Self-learning navigation algorithm for vision-based mobile robots using machine learning algorithms

Size: px
Start display at page:

Download "Self-learning navigation algorithm for vision-based mobile robots using machine learning algorithms"

Transcription

1 Journal of Mechanical Science and Technology 25 (1) (2011) DOI /s y Self-learning navigation algorithm for vision-based mobile robots using machine learning algorithms Jeong-Min Choi 1, Sang-Jin Lee 2 and Mooncheol Won 2,* 1 Hyundai Wia Corp. Machine tool research center, #462-18, Sam-dong, Uiwang, Gyeonggido ,Korea 2 Department of Mechatronics Engineering, Chungnam National University, Daejeon , Korea (Manuscript Received May 22, 2010; Revised September 28, 2010; Accepted November 11, 2010) Abstract Many mobile robot navigation methods use, among others, laser scanners, ultrasonic sensors, vision cameras for detecting obstacles and following paths. However, humans use only visual (e.g. eye) information for navigation. In this paper, we propose a mobile robot control method based on machine learning algorithms which use only camera vision. To efficiently define the state of the robot from raw images, our algorithm uses image-processing and feature selection steps to choose the feature subset for a neural network and uses the output of the neural network learned through supervised learning. The output of the neural network uses the state of a reinforcement learning algorithm to learn obstacle-avoiding and path-following strategies using camera vision image. The algorithm is verified by two experiments, which are line tracking and obstacle avoidance. Keywords: Pattern recognition; Feature selection; Reinforcement learning; Mobile robot; Robot vision; Obstacle avoidance Introduction This paper was recommended for publication in revised form by Associate Editor Yang Shi * Corresponding author. Tel.: , Fax.: address: mcwon@cnu.ac.kr KSME & Springer 2011 In contrast to humans, who use only visual information for navigation, many mobile robots use laser scanners and ultrasonic sensors along with vision cameras to navigate. The goal of our research is to develop a navigation algorithm for mobile robots using only visual information. Also, by using a reinforcement learning algorithm [1], we expect that the mobile robot will learn the right actions for navigating by itself. Research on task learning using visual information has been introduced by the following papers: Gasket et al. use reinforcement learning ( Advantage Learning ) to teach the mobile robot navigation with a match matrix consisting of a stored carpet image and a camera image [2]. Asada et al. use reinforcement learning using position and the size of a goal and a ball detected from the camera image to teach the robot how to shoot at a goal [3]. Nehmzow uses supervised learning with the neural network and the distribution of edge pixels in the extracted edge image for navigation [4]. Regueiro et al. proposed a leaning method using reinforcement learning with the state defined by existence and nonexistence of the edge in a grid image [5]. Shibata et al. use an original image to teach the box-pushing task using reinforcement learning (Actor-critic architecture) [6]. However, these methods need complex image-processing to detect specific patterns [2, 3] or use camera vision along with other sensors such as IR and ultrasonic sensors for learning [5] [6]. In contrast to the above-mentioned studies we propose a navigation method using only camera-vision information processed by a simple image-processing algorithm and the machine learning algorithm. Also, we verify our algorithm by two experiments which are ine tracking and obstacle avoidance. 2. Learning process and experimental settings 2.1 Learning process The suggested learning process used in this paper consists of two parts as shown in Fig. 1. The first part (Fig. 1(a)) refers to the off-line neural network [7] learning as a road environment recognizer. The input features of the neural network are obtained from camera images. The raw training dataset for the neural network is obtained by the manual operation of robot angular velocities under the constant forward velocity 0.3m/s. The training data consist of camera images (inputs of the neural network) and momentary angular velocities (outputs of the neural network). Also, we adopted the feature selection technique to find the best input features from camera images to optimize the performance of

2 248 J.-M. Choi et al. / Journal of Mechanical Science and Technology 25 (1) (2011) (a) The neural network learning process (off-line) Fig. 1. The suggested learning process. (b) Navigation learning process (on-line) Fig. 3. The desired path for the line tracking experiment. Fig. 2. The mobile robot for experiments. the environment recognizer. The second part (Fig. 1(b)) is navigation learning using reinforcement learning by on-line experiments with the mobile robot. We use the Q-learning algorithm [8] which is one of the popular reinforcement learning methods. The mobile robot learns the proper angular velocity and forward velocity for the navigation environment by using the Q-learning algorithm. The state of Q-learning is the output of the environment recognizer for the current image and the current forward velocity of the mobile robot. The action of Q-learning is the forward velocity and angular velocity of the mobile robot. The reward of Q-learning is defined by the forward velocity, angular velocity and change of edge pixels. 2.2 The mobile robot and experiment settings We used the mobile robot depicted in Fig. 2 for our experiments. The mobile robot is 0.5m wide, 0.4m long, and 0.77m high. It has a monocular vision camera that can change its Fig. 4. Experiment environment for obstacle avoidance. pitch angle. The maximum velocity of the mobile robot is about 1.5m/s. Our first experiment consisted of tracking the line shown in Fig. 3. The goal of this experiment was to track the desired path which is composed of a straight and a curved line. The second experiment is obstacle avoidance. One of the test environments is depicted in Fig. 4. We set box obstacles in a corridor with black strips at the border between the walls and the floor. The goal of this experiment is to avoid the obstacle and to follow the corridor. 3. Supervised learning of the neural network for environment recognition The navigation method proposed in this paper uses only

3 J.-M. Choi et al. / Journal of Mechanical Science and Technology 25 (1) (2011) Table 1. Feature candidates for the neural network. Fig. 5. The image processing procedure. No. Feature (edge extraction method distribution direction) 1 Sobel filter Horizontal direction 2 Sobel filter Vertical direction 3 X differentiation filter Horizontal direction 4 X differentiation filter Vertical direction 5 Y differentiation filter Horizontal direction 6 Y differentiation filter Vertical direction 7-45 differentiation filter Horizontal direction 8-45 differentiation filter Vertical direction 9 45 differentiation filter Horizontal direction differentiation filter Vertical direction Fig. 6. The image processing procedure. (a) The horizontal (x direction) distribution data (b) The vertical (y direction) distribution data Fig. 7. The generation of the distribution data. camera-generated visual information in order to recognize the navigation environment. Therefore, this paper uses the pattern recognition method of the neural network as an algorithm to detect lanes or obstacles more effectively. Also finding the optimal input data for the neural network is necessary to improve the performance of pattern recognition. Therefore we used the feature selection algorithm [9] to choose the best input feature set of the neural network. 3.1 Training data acquisition for the neural network The raw training data for the neural network was obtained in the following way: We acquired camera images and momentary angular velocities ranging between -20 /s 20 /s during the manual operation with a computer keyboard and RF Communication every 0.1s. At this time the forward velocity of the mobile robot was fixed at 0.3m/s. The edge is detected by imaged acquired here and the edge pixel distribution is defined as input. Momentary angular velocity is defined as output to compose a learning data set of the neural network. At this time, since the momentary angular velocity is the angular velocity of fixed velocity, 0.3m/s, it cannot be the absolute value applied to every velocity. In this paper, this value is used as an index describes in which situation the robot is placed. Through this process we made 110 data sets for the linetracking experiment and 518 data sets for the obstacleavoidance experiment. 3.2 Image processing to generate input features After data acquisition, we processed the original images in order to get more meaningful and useful input features for the neural network. Fig. 5 shows the image-processing procedure that we adopted. The first step required reducing the image size from 160x140 to 40x30. The second step involved changing the color image into the gray level one. The third step consisted of making the edge image extracted by five kinds of filters (sobel filter, x, y, -45, +45 differentiation filter). Fig. 6 shows the result of this step about the image of the obstacle avoidance experiment. And the last step involved counting edge pixels in separate areas which are 10 for each horizontal and vertical (x and y) direction to get the distribution data. This procedure is shown in Fig. 7. After the image processing procedure, we finally obtained 10 distribution data sets from each image. 10 distribution data sets are shown in Table 1. These data became input feature candidates for the neural network to recognize the road environment. The reason that the distribution of the horizontal and vertical direction of the edge pixel is generated by the final results of the image processing is that information such as distance of lanes or obstacles and obstacle volumes is indirectly collected through this information. Although this paper does not directly

4 250 J.-M. Choi et al. / Journal of Mechanical Science and Technology 25 (1) (2011) Table 2. Feature combinations of each subset number. Subset No. Combined feature (refer to Table. 1) 4_210 7, 8, 9, 10 4_115 3, 4, 9, 10 3_22 1, 5, 6 2_1 1, 2 1_2 2 Fig. 8. The structure of the neural network for road state recognition. Fig. 9. Feature selection results of the obstacle avoidance experiment training data. calculate obstacle volumes or distances, distribution data that draws out those values are used as inputs of a neural network that recognizes the navigation environment for the mobile robot to navigate using only one camera image. 3.3 Feature selection for the optimal feature subset Because 10 input feature candidates do not have the same meaning and importance for the relation between input and output of the neural network, we needed a procedure which chooses the optimal input feature subset. This procedure is called feature selection. We carried out feature selection to optimize the performance of the neural network as the navigation environment recognizer and to reduce the training time of the neural network. Three popular methods for feature selection are the wrapper method [10], the embedded method [11], and the filter method [12]. The feature subset search strategy and the way to measure the performance of the feature subset are important factors for feature selection. In this paper, we use the wrapper method with the forward method as the feature subset search strategy and validation error by a cross validation for the performance measure. The wrapper method was used to train the neural network for all feature subsets, so it has a weak point in that the computation cost is high. Therefore, we carried out feature selection with 385 feature subsets (from 1 to 4 feature combinations) instead of 1024 feature subsets (from 1 to 10 feature combinations). After training the neural network for each 385 feature subset with the cross validation, we choose the feature subset which has minimum validation error as the optimal feature subset. 3.4 Results of the feature selection and the neural network training The neural network for road state recognition using in the feature selection was designed with the multi-layer feed forward neural network which has 2 hidden layers. Fig. 8 shows the structure of the neural network system. The number of inputs is 1040 in accordance with the number of feature combinations. The output layer has 1 neuron which is the correspondence of momentary angular velocity with the input image. Each hidden layer has 10 and 5 neurons individually. The scaled conjugate gradient method [13] was used to train the weights of the neural network. Fig. 9 shows a part of the feature selection results obtained from the obstacle avoidance experiment training data. It is difficult to show all results for 385 feature subsets, so we chose 5 results and show their training and validation errors. The 4_210 feature combination is the optimal feature subset, because it has the smallest validation error. So we can expect that the 4_210 feature combination will result in the best performance when used for the neural network input feature. Table 2 shows combined features of each feature subset number in Fig. 9. By comparing the 4_210 and 4_155 feature subsets, we can notice the difference of importance between 7, 8 features and 3, 4 features. Although they have commonly 9, 10 features, there is the a difference on their validation errors. Also, in the case of the 2_1 feature subset combined with 1, 2 features it has enough edge information for all directions, because it is generated by the sobel filter. However it shows poor performance when compared to the 4_210 feature combination. The reason why the 4_210 feature combination is the optimal feature subset for the obstacle avoidance experiment is that most images contain the diagonal black areas which represent the borders between the walls and the floor. As a result of the feature selection, the optimal feature subset is the feature combination composed of 2, 5, 7, 8 features for the line tracking experiment and 7, 8, 9, 10 features for the obstacle avoidance experiment. Therefore, we used these optimal feature subsets as input features to train weights of each neural network for the two experiments. We carried out the on-line experiments with these trained neural networks as the navigation environment recognizer.

5 J.-M. Choi et al. / Journal of Mechanical Science and Technology 25 (1) (2011) Table 3. State definition for the line tracking experiment. State Output Range Navigation learning algorithm We used reinforcement learning to build the navigation learning algorithm. Reinforcement learning is a learning method in which which the action of the state is learned when maximizing the reward from the environment through trial and error. Unlike supervised learning that needs a pair of input-output training data, reinforcement learning needs only reward to evaluate adequacy of actions. In this paper, we use the Q-learning algorithm which is one of the reinforcement learning algorithms. 4.1 Q-learning algorithm The Q-learning algorithm is a model-free method learning algorithm. It learns the Q-value which means the expected value of a pair of state and action. This algorithm is learning the Q-value table by iteration with the delayed reward of action in state. The update rule of the Q-value is given as U = r + γ max Q( s, a) (1) t+ 1 t+ 1 t+ 1 a Q( s, a ) Q( s, a ) + [ U Q( s, a )] (2) α t t t t t+ 1 t t Eq. (1) reflects the delayed reward and the Q-value of next state and Eq. (2) is the update equation of Q-value. The action in the specific state is determined by Eq. (3) to be chosen as one of possible actions maximizing the Q-value. a arg max Q( s, a') (3) a ' Table 4. Action definition for the line tracking experiment Action Angular velocity ( /s) Also, to learn optimal actions it is needed to select random actions by exploration instead of always selecting learned actions based on Eq. (3). So, we use the ε-greedy strategy [1] which selects random actions by ε probability. Table 5. Visual state definition for the obstacle avoidance experiment. Visual state Output Range Table 6. Velocity state definition for the obstacle avoidance experiment. Velocity state Forward velocity range (m/s) Fig. 10. The reward area for the line tracking experiment States, actions and rewards design for the line tracking experiment We designed states using the output of the neural network trained for the line tracking experiment. States were discretized by 7 discrete output ranges of the neural network. Because the neural network output has the range from -1.0 to 1.0, we discretize states as shown in Table 3. The 0 state means that the mobile robot should turn left rapidly and the 6 state means that the mobile robot should turn right rapidly. The 3 state means that the mobile robot will continue along a straight line. We considered only the angular velocity for the design of actions, and we fixed the forward velocity at 0.3 m/s. Table 4 shows the definition of actions. We picked five angular velocities ranging between -15 /s 15 /s. We introduce a reward area to design rewards and use the number of edge pixels in it. Fig. 10 shows the reward area for the line tracking experiment. If the number of edge pixels in it is large, we could think that the mobile robot drives well along with the line. Also, if it doesn t contain any edge pixels, it means that the mobile robot is off the line. So, we design rewards such that the more edge pixels in it, the bigger rewards are gotten. Eq. (4) is the designed reward for the line tracking experiment. 4.2 States, actions and rewards design The design of states, actions and rewards in reinforcement learning is the major factor affecting learning results. Therefore, it is important to design those that are suitable for the goal of our experiments. 1.0, if N 40, N 40 t t 1 r = 10.0, if N 10, N 30 t t t 1 1.0, else (4) The mobile robot receives aplus reward when more than 40 edge pixels are in the reward area for two control loop times (500ms). On the other hand, the mobile robot receives a minus reward when it runs off the line at current loop time.

6 252 J.-M. Choi et al. / Journal of Mechanical Science and Technology 25 (1) (2011) Table 7. Action definition for the obstacle avoidance experiment. Angular velocity Velocity change -20 /s 0 /s 20 /s Decrease (-0.2m/s) No change (0m/s) Increase (0.2m/s) Fig. 11. The reward area for the obstacle avoidance experiment. Fig. 12. Learning time and the average reward for each episode of the line tracking experiment States, actions and rewards design for the obstacle avoidance experiment States for the obstacle avoidance experiment are designed by the output of the neural network and the forward velocity of the mobile robot. By including the forward velocity as a state, the mobile robot could learn the right actions, unlike the line tracking experiment. Table 5 and table 6 show the state definitions for visual states and forward velocity states respectively. Like the state definition of the line tracking experiment, the 0 and 4 visual state means that the mobile robot should turn left and right respectively. Actions of the mobile robot are designed with combinations of the forward velocity change and the angular velocity. The forward velocity change is keeping current velocity, 0.2m/s increase, and decrease. The designed angular velocity is -20 /s, 0 /s, 20 /s. Therefore, the number of actions is 9 (see Table 7). Because of this kind of action design, the mobile robot can learn proper forward velocities and angular velocities for each state. Table 7 shows the action definitions for the obstacle avoidance experiment. We adopted the reward area as in the line tracking experiment. Fig. 11 shows the reward area for the obstacle avoidance experiment. In this experiment, the edge pixels in the reward area represent obstacles or walls to avoid. Therefore, we designed rewards so that whenfewer edge pixels are in the reward area, bigger rewards are obtained. Also, we considered the forward velocity and the angular velocity for the design of the reward for the mobile robot to encourage it to go as fast and straight as possible. r = ( N N ) + ( v 0.5) 0.02 w (5) t t t t t Eq. (5) is the designed reward for the obstacle avoidance experiment. This equation consists of rewards linked to the number of edge pixels in the reward area and the forward velocity and the angular velocity of mobile robot. So, the mobile robot obtains big rewards when it runs in a straight line at a (a) An image at the start point (c) An image at the middle of a curvy path fast speed and avoids obstacles in the reward area. We tuned the contribution of each term by adjusting the constant in front of the corresponding terms. 5. Experiment results (b) An image at the beginning of a curvy path (d) An image after the turning Fig. 13. Images during the line tracking after learning. Our algorithm is verified by two on-line experiments. All experiments were implemented in good lighting conditions, because performance of the visual camera, which is our unique sensor, is seriously affected by lighting conditions. The pitch angle of the camera is set at 45 for the line tracking experiment and at 40 for the obstacle avoidance experiment in orderto catch images of the further distance. The loop time of the learning algorithm is 500ms. 5.1 Results of the line tracking experiment An episode for line tracking experiment ends when the mobile robot reaches the end point or runs markedly off the desired path. Fig. 12 shows learning time and the average reward for each episode. The average reward is the reward per second. At the beginning of the learning process the mobile robot ob-

7 J.-M. Choi et al. / Journal of Mechanical Science and Technology 25 (1) (2011) Fig. 14. Learning time and the average reward for each episode of the obstacle avoidance experiment. (a) The path of the mobile robot at the beginning of the learning process (b) The path of the mobile robot after learning Fig. 15. Navigation paths of the mobile robot during and after learning. tains small average rewards and achieves little learning time, because it frequently runs off the line. However, by increasing the number of the episodes, learning time and the average rewards increase. After the 15 th episode, both average rewards and learning time are roughly on saturation levels. From this we can infer that the mobile robot learns proper actions which are suitable for learning states and drives well along the desired path. Fig. 13 shows images taken during the line tracking after learning is completed. We know through these images that the mobile robot has learned to keep the black line in the reward area while driving. This result comes from the reward design that bigger rewards are obtained when more edge pixels are in the reward area. Although we conducted the line tracking navigation test in the environment combined only with straight and curved lanes as seen in Fig. 3, the algorithm that is suggested can be applied to other situations. The reason is that the algorithm suggested here uses vertical and horizontal direction distribution information of edge pixels, and that although the line used in this test is not a curved line but a corner line with a rightangled shape, it shows distribution data similar to those of the curved line. 5.2 Results of the obstacle avoidance experiment Fig. 14 shows learning time and the average reward for each episode for the obstacle avoidance experiment. The episode ends when the robot reaches the end point or approaches the obstacle or a wall too closely. (a) The path of the mobile robot in the new environment #1 (b) The path of the mobile robot in the new environment #2 Fig. 16. Navigation paths of the mobile robot in new environments. Like the line tracking experiment, learning time and average rewards are small in the early episode because the mobile did not learn the right actions yet. Also, learning time is long, but the average reward is not big at the 12 th episode, because the mobile robot runs around obstacles at a slow speed. However, after the 17 th episode (after the mobile robot learns enough) the mobile robot gets nearly equal rewards and learning time except for a few episodes (20, 24, 25, 31) during which the mobile robot executes random actions by ε-greedy learning strategy with a change of 10% exploration probability (ε=0.1). However, after the 35 th episode that exploration rate is set at 0 the mobile robot gets equal rewards during the nearly same learning time. So, we could know that the Q-table is learned correctly for this navigation environment. Navigation paths of the mobile robot during learning are shown in Fig. 15. The black vertical lines represent the boundaries between walls and the floor. In the beginning of learning (Fig. 15(a)), we notice that the navigation is finished because the mobile robot approaches the obstacle too closely. After enough learning (Fig. 15(b)), the mobile robot runs well while avoiding obstacles and walls. Fig. 16 shows the navigation path using the learned Q-table in the new environment with two obstacles. Fig. 16(a) shows that the mobile robot turns right to avoid the first obstacle and then turns left, unlike the movement reflected in Fig. 15(b). This is because the second obstacle is detected. Fig. 16(b) shows the navigation path in a more difficult environment than that in Fig. 16(a). The first obstacle is located more closely than in Fig. 16(a), and a bigger second obstacle than Fig. 16(a) is set. In this case, the mobile robot follows a more curvy path than in Fig. 16(a) in order to avoid the second obstacle. Through these experiments, we know that it is possible to navigate in an environment which is different from the learning environment.. The algorithm suggested here can be applied to circular or other shapes of obstacles in the place of a square shaped box as used in this paper. Since the algorithm suggested in this paper uses the distribution of edge pixels, edge pixel exists in similar location of the image, and similar distribution data will be acquired regardless of the obstacle shapes, if the obstacles are placed in the similar location. Therefore, the robot can perform obstacle-avoidance by the algorithm suggested here

8 254 J.-M. Choi et al. / Journal of Mechanical Science and Technology 25 (1) (2011) regardless of the shape of the obstacle. 6. Conclusions We proposed a navigation learning algorithm using only visual information for mobile robots by mimicking human behavior. We used the pattern recognition technique of the neural network to enable the mobile robot to differentiate between the environment and the camera image. Also, we adopted the feature selection procedure in order to optimize the performance of the neural network. By using this procedure, we found the best feature subset for the input of the neural network. Output of the neural network and the forward velocity of the mobile robot were used for the state of Q-learning which is our navigation learning algorithm. We introduced the reward area to define rewards for Q-learning. Actions are forward velocity and angular velocity. We verified the proposed algorithm by the line tracking and obstacle avoidance experiments. We confirmed that the mobile robot navigates well in both experiments after sufficient learning. Our future work will focus on developing an algorithm which can navigate more complex indoor environments such as offices. Nomenclature Q (s, a) : Q-value s : State a : Action r : Reward γ : Discount factor α : Learning rate ε : Exploration rate N : The number of edge pixels in the reward area t at time t v : The forward velocity of the mobile robot at time t t : The angular velocity of the mobile robot at time t w t References [1] Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction, The MIT Press (1998). [2] Chris Gaskett, Luke Fletcher and Alexander Zelinsky, Reinforcement Learning for a Vision Based Mobile Robot, IEEE (2000). [3] Minoru Asada, Shoichi Koda, Sukoya Tawaratsumida and Koh Hosoda, Vision-based reinforcement learning for purposive behavior acquisition, IEEE International Conference on Robotics and Automation (1995). [4] Ulrich Nehmzow, Vision Processing for Robot Learning, J. Industrial Robot, 26 (2) (1999) [5] Carlos V. Regueiro, Jos e E. Domenech, Roberto Iglesias, and Jos e L. Correa, Acquiring contour following behavior in robotics through Q-learning and image-based states, PWASET, 15 (2006). [6] Katsunari Shibata and Masaru Iida, Acquisition of box pushing by direct-vision-based reinforcement learning, SICE Annual Conference (2003). [7] Christopher M. Bishop, Neural Networks for Pattern Recognition, Oxford University Press (1997). [8] C. J. C. H. Watkins. Learning from Delayed Rewards. Cambridge University (1989). [9] Isabelle Guyon and Andre Elisseeff, An Introduction to Variable and Feature Selection, Journal of Machine Learning Research 3 (2003) [10] R. Kohavi and G. John, Wrappers for feature selection, Artificial Intelligence, 97 (1-2) (1997) [11] L. Breiman, J. H. Friedman, R. A. Olshen and C. J. Stone, Classification and Regression Trees, Wadsworth and Brooks (1984). [12] H. Stoppiglia, G. Dreyfus, R. Dubois and Y. Oussar, Ranking a random feature for variable and feature selection, JMLR, 3 (2003) [13] Martin F. Moller, November, A Scaled Conjugate Gradient Algorithm for Fast Supervised Learning, Neural Networks, 6 (1990) Mooncheol Won received a B.Sc. and an M.Sc. degree from Seoul National University, Korea, in the Department of Naval Architecture and Ocean Engineering. He also received a Ph.D. degree in mechanical engineering from the University of California at Berkeley, USA. Currently, he is a professor in the Department of Mechatronics Engineering at Chungnam National University, Korea. His research interests include control of maritime and mechatronics systems, and machine learning applications of robotic systems. Jeong-Min Choi received a B.Sc. degree Chungnam National University, Korea in the Department of Mechatronics Engineering. Currently, he is in the researcher in the department of Research Center of Hyundai Wia Corp., Korea. His research interests include machine learning applications of robotic systems, especially reinforcement learning and neural networks. robotic systems. Sang-Jin Lee received the B.Sc. degree in the department of mechatronics engineering from Chungnam National University, Korea. He is in the master s course in the department of mechatronics engineering of Chungnam National University, Korea. His research interests include machine learning applications of

Continuous Valued Q-learning for Vision-Guided Behavior Acquisition

Continuous Valued Q-learning for Vision-Guided Behavior Acquisition Continuous Valued Q-learning for Vision-Guided Behavior Acquisition Yasutake Takahashi, Masanori Takeda, and Minoru Asada Dept. of Adaptive Machine Systems Graduate School of Engineering Osaka University

More information

Path Planning for a Robot Manipulator based on Probabilistic Roadmap and Reinforcement Learning

Path Planning for a Robot Manipulator based on Probabilistic Roadmap and Reinforcement Learning 674 International Journal Jung-Jun of Control, Park, Automation, Ji-Hun Kim, and and Systems, Jae-Bok vol. Song 5, no. 6, pp. 674-680, December 2007 Path Planning for a Robot Manipulator based on Probabilistic

More information

ADAPTIVE TILE CODING METHODS FOR THE GENERALIZATION OF VALUE FUNCTIONS IN THE RL STATE SPACE A THESIS SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL

ADAPTIVE TILE CODING METHODS FOR THE GENERALIZATION OF VALUE FUNCTIONS IN THE RL STATE SPACE A THESIS SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL ADAPTIVE TILE CODING METHODS FOR THE GENERALIZATION OF VALUE FUNCTIONS IN THE RL STATE SPACE A THESIS SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL OF THE UNIVERSITY OF MINNESOTA BY BHARAT SIGINAM IN

More information

An Image Based Approach to Compute Object Distance

An Image Based Approach to Compute Object Distance An Image Based Approach to Compute Object Distance Ashfaqur Rahman * Department of Computer Science, American International University Bangladesh Dhaka 1213, Bangladesh Abdus Salam, Mahfuzul Islam, and

More information

Reinforcement Learning for Appearance Based Visual Servoing in Robotic Manipulation

Reinforcement Learning for Appearance Based Visual Servoing in Robotic Manipulation Reinforcement Learning for Appearance Based Visual Servoing in Robotic Manipulation UMAR KHAN, LIAQUAT ALI KHAN, S. ZAHID HUSSAIN Department of Mechatronics Engineering AIR University E-9, Islamabad PAKISTAN

More information

Visual Attention Control by Sensor Space Segmentation for a Small Quadruped Robot based on Information Criterion

Visual Attention Control by Sensor Space Segmentation for a Small Quadruped Robot based on Information Criterion Visual Attention Control by Sensor Space Segmentation for a Small Quadruped Robot based on Information Criterion Noriaki Mitsunaga and Minoru Asada Dept. of Adaptive Machine Systems, Osaka University,

More information

A threshold decision of the object image by using the smart tag

A threshold decision of the object image by using the smart tag A threshold decision of the object image by using the smart tag Chang-Jun Im, Jin-Young Kim, Kwan Young Joung, Ho-Gil Lee Sensing & Perception Research Group Korea Institute of Industrial Technology (

More information

6. NEURAL NETWORK BASED PATH PLANNING ALGORITHM 6.1 INTRODUCTION

6. NEURAL NETWORK BASED PATH PLANNING ALGORITHM 6.1 INTRODUCTION 6 NEURAL NETWORK BASED PATH PLANNING ALGORITHM 61 INTRODUCTION In previous chapters path planning algorithms such as trigonometry based path planning algorithm and direction based path planning algorithm

More information

A Simple Interface for Mobile Robot Equipped with Single Camera using Motion Stereo Vision

A Simple Interface for Mobile Robot Equipped with Single Camera using Motion Stereo Vision A Simple Interface for Mobile Robot Equipped with Single Camera using Motion Stereo Vision Stephen Karungaru, Atsushi Ishitani, Takuya Shiraishi, and Minoru Fukumi Abstract Recently, robot technology has

More information

CS4758: Rovio Augmented Vision Mapping Project

CS4758: Rovio Augmented Vision Mapping Project CS4758: Rovio Augmented Vision Mapping Project Sam Fladung, James Mwaura Abstract The goal of this project is to use the Rovio to create a 2D map of its environment using a camera and a fixed laser pointer

More information

3D Grid Size Optimization of Automatic Space Analysis for Plant Facility Using Point Cloud Data

3D Grid Size Optimization of Automatic Space Analysis for Plant Facility Using Point Cloud Data 33 rd International Symposium on Automation and Robotics in Construction (ISARC 2016) 3D Grid Size Optimization of Automatic Space Analysis for Plant Facility Using Point Cloud Data Gyu seong Choi a, S.W.

More information

Gauss-Sigmoid Neural Network

Gauss-Sigmoid Neural Network Gauss-Sigmoid Neural Network Katsunari SHIBATA and Koji ITO Tokyo Institute of Technology, Yokohama, JAPAN shibata@ito.dis.titech.ac.jp Abstract- Recently RBF(Radial Basis Function)-based networks have

More information

Localization algorithm using a virtual label for a mobile robot in indoor and outdoor environments

Localization algorithm using a virtual label for a mobile robot in indoor and outdoor environments Artif Life Robotics (2011) 16:361 365 ISAROB 2011 DOI 10.1007/s10015-011-0951-7 ORIGINAL ARTICLE Ki Ho Yu Min Cheol Lee Jung Hun Heo Youn Geun Moon Localization algorithm using a virtual label for a mobile

More information

3D Corner Detection from Room Environment Using the Handy Video Camera

3D Corner Detection from Room Environment Using the Handy Video Camera 3D Corner Detection from Room Environment Using the Handy Video Camera Ryo HIROSE, Hideo SAITO and Masaaki MOCHIMARU : Graduated School of Science and Technology, Keio University, Japan {ryo, saito}@ozawa.ics.keio.ac.jp

More information

Neural Networks for Obstacle Avoidance

Neural Networks for Obstacle Avoidance Neural Networks for Obstacle Avoidance Joseph Djugash Robotics Institute Carnegie Mellon University Pittsburgh, PA 15213 josephad@andrew.cmu.edu Bradley Hamner Robotics Institute Carnegie Mellon University

More information

Calibration of Inertial Measurement Units Using Pendulum Motion

Calibration of Inertial Measurement Units Using Pendulum Motion Technical Paper Int l J. of Aeronautical & Space Sci. 11(3), 234 239 (2010) DOI:10.5139/IJASS.2010.11.3.234 Calibration of Inertial Measurement Units Using Pendulum Motion Keeyoung Choi* and Se-ah Jang**

More information

Combining Deep Reinforcement Learning and Safety Based Control for Autonomous Driving

Combining Deep Reinforcement Learning and Safety Based Control for Autonomous Driving Combining Deep Reinforcement Learning and Safety Based Control for Autonomous Driving Xi Xiong Jianqiang Wang Fang Zhang Keqiang Li State Key Laboratory of Automotive Safety and Energy, Tsinghua University

More information

CS4758: Moving Person Avoider

CS4758: Moving Person Avoider CS4758: Moving Person Avoider Yi Heng Lee, Sze Kiat Sim Abstract We attempt to have a quadrotor autonomously avoid people while moving through an indoor environment. Our algorithm for detecting people

More information

Stacked Denoising Autoencoders for Face Pose Normalization

Stacked Denoising Autoencoders for Face Pose Normalization Stacked Denoising Autoencoders for Face Pose Normalization Yoonseop Kang 1, Kang-Tae Lee 2,JihyunEun 2, Sung Eun Park 2 and Seungjin Choi 1 1 Department of Computer Science and Engineering Pohang University

More information

Efficient L-Shape Fitting for Vehicle Detection Using Laser Scanners

Efficient L-Shape Fitting for Vehicle Detection Using Laser Scanners Efficient L-Shape Fitting for Vehicle Detection Using Laser Scanners Xiao Zhang, Wenda Xu, Chiyu Dong, John M. Dolan, Electrical and Computer Engineering, Carnegie Mellon University Robotics Institute,

More information

Flexible Calibration of a Portable Structured Light System through Surface Plane

Flexible Calibration of a Portable Structured Light System through Surface Plane Vol. 34, No. 11 ACTA AUTOMATICA SINICA November, 2008 Flexible Calibration of a Portable Structured Light System through Surface Plane GAO Wei 1 WANG Liang 1 HU Zhan-Yi 1 Abstract For a portable structured

More information

Self-Organization of Place Cells and Reward-Based Navigation for a Mobile Robot

Self-Organization of Place Cells and Reward-Based Navigation for a Mobile Robot Self-Organization of Place Cells and Reward-Based Navigation for a Mobile Robot Takashi TAKAHASHI Toshio TANAKA Kenji NISHIDA Takio KURITA Postdoctoral Research Fellow of the Japan Society for the Promotion

More information

View-based Programming with Reinforcement Learning for Robotic Manipulation

View-based Programming with Reinforcement Learning for Robotic Manipulation View-based Programming with Reinforcement Learning for Robotic Manipulation Yusuke MAEDA*, Takumi WATANABE** and Yuki MORIYAMA* *Yokohama National University **Seiko Epson Corp. Background Conventional

More information

HOG-Based Person Following and Autonomous Returning Using Generated Map by Mobile Robot Equipped with Camera and Laser Range Finder

HOG-Based Person Following and Autonomous Returning Using Generated Map by Mobile Robot Equipped with Camera and Laser Range Finder HOG-Based Person Following and Autonomous Returning Using Generated Map by Mobile Robot Equipped with Camera and Laser Range Finder Masashi Awai, Takahito Shimizu and Toru Kaneko Department of Mechanical

More information

A Fast and Accurate Eyelids and Eyelashes Detection Approach for Iris Segmentation

A Fast and Accurate Eyelids and Eyelashes Detection Approach for Iris Segmentation A Fast and Accurate Eyelids and Eyelashes Detection Approach for Iris Segmentation Walid Aydi, Lotfi Kamoun, Nouri Masmoudi Department of Electrical National Engineering School of Sfax Sfax University

More information

Behavior Learning for a Mobile Robot with Omnidirectional Vision Enhanced by an Active Zoom Mechanism

Behavior Learning for a Mobile Robot with Omnidirectional Vision Enhanced by an Active Zoom Mechanism Behavior Learning for a Mobile Robot with Omnidirectional Vision Enhanced by an Active Zoom Mechanism Sho ji Suzuki, Tatsunori Kato, Minoru Asada, and Koh Hosoda Dept. of Adaptive Machine Systems, Graduate

More information

A Study on Object Tracking Signal Generation of Pan, Tilt, and Zoom Data

A Study on Object Tracking Signal Generation of Pan, Tilt, and Zoom Data Vol.8, No.3 (214), pp.133-142 http://dx.doi.org/1.14257/ijseia.214.8.3.13 A Study on Object Tracking Signal Generation of Pan, Tilt, and Zoom Data Jin-Tae Kim Department of Aerospace Software Engineering,

More information

Study on Synchronization for Laser Scanner and Industrial Robot

Study on Synchronization for Laser Scanner and Industrial Robot Study on Synchronization for Laser Scanner and Industrial Robot Heeshin Kang 1 1 Korea Institute of Machinery and Materials, Daejeon, Korea 173 Abstract On this paper, a study of robot based remote laser

More information

A Fast Circular Edge Detector for the Iris Region Segmentation

A Fast Circular Edge Detector for the Iris Region Segmentation A Fast Circular Edge Detector for the Iris Region Segmentation Yeunggyu Park, Hoonju Yun, Myongseop Song, and Jaihie Kim I.V. Lab. Dept. of Electrical and Computer Engineering, Yonsei University, 134Shinchon-dong,

More information

Visual Servoing Utilizing Zoom Mechanism

Visual Servoing Utilizing Zoom Mechanism IEEE Int. Conf. on Robotics and Automation 1995, pp.178 183, Nagoya, May. 12 16, 1995 1 Visual Servoing Utilizing Zoom Mechanism Koh HOSODA, Hitoshi MORIYAMA and Minoru ASADA Dept. of Mechanical Engineering

More information

Adaptive Building of Decision Trees by Reinforcement Learning

Adaptive Building of Decision Trees by Reinforcement Learning Proceedings of the 7th WSEAS International Conference on Applied Informatics and Communications, Athens, Greece, August 24-26, 2007 34 Adaptive Building of Decision Trees by Reinforcement Learning MIRCEA

More information

A Fuzzy Reinforcement Learning for a Ball Interception Problem

A Fuzzy Reinforcement Learning for a Ball Interception Problem A Fuzzy Reinforcement Learning for a Ball Interception Problem Tomoharu Nakashima, Masayo Udo, and Hisao Ishibuchi Department of Industrial Engineering, Osaka Prefecture University Gakuen-cho 1-1, Sakai,

More information

Design of Obstacle Avoidance System for Mobile Robot using Fuzzy Logic Systems

Design of Obstacle Avoidance System for Mobile Robot using Fuzzy Logic Systems ol. 7, No. 3, May, 2013 Design of Obstacle Avoidance System for Mobile Robot using Fuzzy ogic Systems Xi i and Byung-Jae Choi School of Electronic Engineering, Daegu University Jillyang Gyeongsan-city

More information

OBSTACLE DETECTION USING STRUCTURED BACKGROUND

OBSTACLE DETECTION USING STRUCTURED BACKGROUND OBSTACLE DETECTION USING STRUCTURED BACKGROUND Ghaida Al Zeer, Adnan Abou Nabout and Bernd Tibken Chair of Automatic Control, Faculty of Electrical, Information and Media Engineering University of Wuppertal,

More information

Matching Evaluation of 2D Laser Scan Points using Observed Probability in Unstable Measurement Environment

Matching Evaluation of 2D Laser Scan Points using Observed Probability in Unstable Measurement Environment Matching Evaluation of D Laser Scan Points using Observed Probability in Unstable Measurement Environment Taichi Yamada, and Akihisa Ohya Abstract In the real environment such as urban areas sidewalk,

More information

Real time game field limits recognition for robot self-localization using collinearity in Middle-Size RoboCup Soccer

Real time game field limits recognition for robot self-localization using collinearity in Middle-Size RoboCup Soccer Real time game field limits recognition for robot self-localization using collinearity in Middle-Size RoboCup Soccer Fernando Ribeiro (1) Gil Lopes (2) (1) Department of Industrial Electronics, Guimarães,

More information

Learn to Swing Up and Balance a Real Pole Based on Raw Visual Input Data

Learn to Swing Up and Balance a Real Pole Based on Raw Visual Input Data Learn to Swing Up and Balance a Real Pole Based on Raw Visual Input Data Jan Mattner*, Sascha Lange, and Martin Riedmiller Machine Learning Lab Department of Computer Science University of Freiburg 79110,

More information

Toward Interlinking Asian Resources Effectively: Chinese to Korean Frequency-Based Machine Translation System

Toward Interlinking Asian Resources Effectively: Chinese to Korean Frequency-Based Machine Translation System Toward Interlinking Asian Resources Effectively: Chinese to Korean Frequency-Based Machine Translation System Eun Ji Kim and Mun Yong Yi (&) Department of Knowledge Service Engineering, KAIST, Daejeon,

More information

DEVELOPMENT OF POSITION MEASUREMENT SYSTEM FOR CONSTRUCTION PILE USING LASER RANGE FINDER

DEVELOPMENT OF POSITION MEASUREMENT SYSTEM FOR CONSTRUCTION PILE USING LASER RANGE FINDER S17- DEVELOPMENT OF POSITION MEASUREMENT SYSTEM FOR CONSTRUCTION PILE USING LASER RANGE FINDER Fumihiro Inoue 1 *, Takeshi Sasaki, Xiangqi Huang 3, and Hideki Hashimoto 4 1 Technica Research Institute,

More information

Dynamic Obstacle Detection Based on Background Compensation in Robot s Movement Space

Dynamic Obstacle Detection Based on Background Compensation in Robot s Movement Space MATEC Web of Conferences 95 83 (7) DOI:.5/ matecconf/79583 ICMME 6 Dynamic Obstacle Detection Based on Background Compensation in Robot s Movement Space Tao Ni Qidong Li Le Sun and Lingtao Huang School

More information

A Visualization Tool to Improve the Performance of a Classifier Based on Hidden Markov Models

A Visualization Tool to Improve the Performance of a Classifier Based on Hidden Markov Models A Visualization Tool to Improve the Performance of a Classifier Based on Hidden Markov Models Gleidson Pegoretti da Silva, Masaki Nakagawa Department of Computer and Information Sciences Tokyo University

More information

Introduction to Mobile Robotics

Introduction to Mobile Robotics Introduction to Mobile Robotics Gaussian Processes Wolfram Burgard Cyrill Stachniss Giorgio Grisetti Maren Bennewitz Christian Plagemann SS08, University of Freiburg, Department for Computer Science Announcement

More information

Traffic Signs Recognition using HP and HOG Descriptors Combined to MLP and SVM Classifiers

Traffic Signs Recognition using HP and HOG Descriptors Combined to MLP and SVM Classifiers Traffic Signs Recognition using HP and HOG Descriptors Combined to MLP and SVM Classifiers A. Salhi, B. Minaoui, M. Fakir, H. Chakib, H. Grimech Faculty of science and Technology Sultan Moulay Slimane

More information

Fully Automatic Methodology for Human Action Recognition Incorporating Dynamic Information

Fully Automatic Methodology for Human Action Recognition Incorporating Dynamic Information Fully Automatic Methodology for Human Action Recognition Incorporating Dynamic Information Ana González, Marcos Ortega Hortas, and Manuel G. Penedo University of A Coruña, VARPA group, A Coruña 15071,

More information

Modifications of VFH navigation methods for mobile robots

Modifications of VFH navigation methods for mobile robots Available online at www.sciencedirect.com Procedia Engineering 48 (01 ) 10 14 MMaMS 01 Modifications of VFH navigation methods for mobile robots Andre Babinec a * Martin Dean a Františe Ducho a Anton Vito

More information

ECE 285 Class Project Report

ECE 285 Class Project Report ECE 285 Class Project Report Based on Source localization in an ocean waveguide using supervised machine learning Yiwen Gong ( yig122@eng.ucsd.edu), Yu Chai( yuc385@eng.ucsd.edu ), Yifeng Bu( ybu@eng.ucsd.edu

More information

Convolutional Restricted Boltzmann Machine Features for TD Learning in Go

Convolutional Restricted Boltzmann Machine Features for TD Learning in Go ConvolutionalRestrictedBoltzmannMachineFeatures fortdlearningingo ByYanLargmanandPeterPham AdvisedbyHonglakLee 1.Background&Motivation AlthoughrecentadvancesinAIhaveallowed Go playing programs to become

More information

A Method of weld Edge Extraction in the X-ray Linear Diode Arrays. Real-time imaging

A Method of weld Edge Extraction in the X-ray Linear Diode Arrays. Real-time imaging 17th World Conference on Nondestructive Testing, 25-28 Oct 2008, Shanghai, China A Method of weld Edge Extraction in the X-ray Linear Diode Arrays Real-time imaging Guang CHEN, Keqin DING, Lihong LIANG

More information

Predict the box office of US movies

Predict the box office of US movies Predict the box office of US movies Group members: Hanqing Ma, Jin Sun, Zeyu Zhang 1. Introduction Our task is to predict the box office of the upcoming movies using the properties of the movies, such

More information

Study on the Signboard Region Detection in Natural Image

Study on the Signboard Region Detection in Natural Image , pp.179-184 http://dx.doi.org/10.14257/astl.2016.140.34 Study on the Signboard Region Detection in Natural Image Daeyeong Lim 1, Youngbaik Kim 2, Incheol Park 1, Jihoon seung 1, Kilto Chong 1,* 1 1567

More information

Geometrical Feature Extraction Using 2D Range Scanner

Geometrical Feature Extraction Using 2D Range Scanner Geometrical Feature Extraction Using 2D Range Scanner Sen Zhang Lihua Xie Martin Adams Fan Tang BLK S2, School of Electrical and Electronic Engineering Nanyang Technological University, Singapore 639798

More information

Laser Sensor for Obstacle Detection of AGV

Laser Sensor for Obstacle Detection of AGV Laser Sensor for Obstacle Detection of AGV Kyoung-Taik Park*, Young-Tae Shin and Byung-Su Kang * Nano-Mechanism Lab, Department of Intelligent Precision Machine Korea Institute of Machinery & Materials

More information

Classification Algorithm for Road Surface Condition

Classification Algorithm for Road Surface Condition IJCSNS International Journal of Computer Science and Network Security, VOL.4 No., January 04 Classification Algorithm for Road Surface Condition Hun-Jun Yang, Hyeok Jang, Jong-Wook Kang and Dong-Seok Jeong,

More information

Future Computer Vision Algorithms for Traffic Sign Recognition Systems

Future Computer Vision Algorithms for Traffic Sign Recognition Systems Future Computer Vision Algorithms for Traffic Sign Recognition Systems Dr. Stefan Eickeler Future of Traffic Sign Recognition Triangular Signs Complex Signs All Circular Signs Recognition of Circular Traffic

More information

A Symmetry Operator and Its Application to the RoboCup

A Symmetry Operator and Its Application to the RoboCup A Symmetry Operator and Its Application to the RoboCup Kai Huebner Bremen Institute of Safe Systems, TZI, FB3 Universität Bremen, Postfach 330440, 28334 Bremen, Germany khuebner@tzi.de Abstract. At present,

More information

COLLABORATIVE AGENT LEARNING USING HYBRID NEUROCOMPUTING

COLLABORATIVE AGENT LEARNING USING HYBRID NEUROCOMPUTING COLLABORATIVE AGENT LEARNING USING HYBRID NEUROCOMPUTING Saulat Farooque and Lakhmi Jain School of Electrical and Information Engineering, University of South Australia, Adelaide, Australia saulat.farooque@tenix.com,

More information

Augmented Reality of Robust Tracking with Realistic Illumination 1

Augmented Reality of Robust Tracking with Realistic Illumination 1 International Journal of Fuzzy Logic and Intelligent Systems, vol. 10, no. 3, June 2010, pp. 178-183 DOI : 10.5391/IJFIS.2010.10.3.178 Augmented Reality of Robust Tracking with Realistic Illumination 1

More information

Implementation of Enhanced Web Crawler for Deep-Web Interfaces

Implementation of Enhanced Web Crawler for Deep-Web Interfaces Implementation of Enhanced Web Crawler for Deep-Web Interfaces Yugandhara Patil 1, Sonal Patil 2 1Student, Department of Computer Science & Engineering, G.H.Raisoni Institute of Engineering & Management,

More information

Cover Page. Abstract ID Paper Title. Automated extraction of linear features from vehicle-borne laser data

Cover Page. Abstract ID Paper Title. Automated extraction of linear features from vehicle-borne laser data Cover Page Abstract ID 8181 Paper Title Automated extraction of linear features from vehicle-borne laser data Contact Author Email Dinesh Manandhar (author1) dinesh@skl.iis.u-tokyo.ac.jp Phone +81-3-5452-6417

More information

3D Modelling with Structured Light Gamma Calibration

3D Modelling with Structured Light Gamma Calibration 3D Modelling with Structured Light Gamma Calibration Eser SERT 1, Ibrahim Taner OKUMUS 1, Deniz TASKIN 2 1 Computer Engineering Department, Engineering and Architecture Faculty, Kahramanmaras Sutcu Imam

More information

A Nondestructive Bump Inspection in Flip Chip Component using Fuzzy Filtering and Image Processing

A Nondestructive Bump Inspection in Flip Chip Component using Fuzzy Filtering and Image Processing A Nondestructive Bump Inspection in Flip Chip Component using Fuzzy Filtering and Image Processing 103 A Nondestructive Bump Inspection in Flip Chip Component using Fuzzy Filtering and Image Processing

More information

More on Learning. Neural Nets Support Vectors Machines Unsupervised Learning (Clustering) K-Means Expectation-Maximization

More on Learning. Neural Nets Support Vectors Machines Unsupervised Learning (Clustering) K-Means Expectation-Maximization More on Learning Neural Nets Support Vectors Machines Unsupervised Learning (Clustering) K-Means Expectation-Maximization Neural Net Learning Motivated by studies of the brain. A network of artificial

More information

Subpixel Corner Detection Using Spatial Moment 1)

Subpixel Corner Detection Using Spatial Moment 1) Vol.31, No.5 ACTA AUTOMATICA SINICA September, 25 Subpixel Corner Detection Using Spatial Moment 1) WANG She-Yang SONG Shen-Min QIANG Wen-Yi CHEN Xing-Lin (Department of Control Engineering, Harbin Institute

More information

Performance analysis of a MLP weight initialization algorithm

Performance analysis of a MLP weight initialization algorithm Performance analysis of a MLP weight initialization algorithm Mohamed Karouia (1,2), Régis Lengellé (1) and Thierry Denœux (1) (1) Université de Compiègne U.R.A. CNRS 817 Heudiasyc BP 49 - F-2 Compiègne

More information

Forward Feature Selection Using Residual Mutual Information

Forward Feature Selection Using Residual Mutual Information Forward Feature Selection Using Residual Mutual Information Erik Schaffernicht, Christoph Möller, Klaus Debes and Horst-Michael Gross Ilmenau University of Technology - Neuroinformatics and Cognitive Robotics

More information

Analysis of Image and Video Using Color, Texture and Shape Features for Object Identification

Analysis of Image and Video Using Color, Texture and Shape Features for Object Identification IOSR Journal of Computer Engineering (IOSR-JCE) e-issn: 2278-0661,p-ISSN: 2278-8727, Volume 16, Issue 6, Ver. VI (Nov Dec. 2014), PP 29-33 Analysis of Image and Video Using Color, Texture and Shape Features

More information

Notes 9: Optical Flow

Notes 9: Optical Flow Course 049064: Variational Methods in Image Processing Notes 9: Optical Flow Guy Gilboa 1 Basic Model 1.1 Background Optical flow is a fundamental problem in computer vision. The general goal is to find

More information

Applicability Estimation of Mobile Mapping. System for Road Management

Applicability Estimation of Mobile Mapping. System for Road Management Contemporary Engineering Sciences, Vol. 7, 2014, no. 24, 1407-1414 HIKARI Ltd, www.m-hikari.com http://dx.doi.org/10.12988/ces.2014.49173 Applicability Estimation of Mobile Mapping System for Road Management

More information

Topics in AI (CPSC 532L): Multimodal Learning with Vision, Language and Sound. Lecture 12: Deep Reinforcement Learning

Topics in AI (CPSC 532L): Multimodal Learning with Vision, Language and Sound. Lecture 12: Deep Reinforcement Learning Topics in AI (CPSC 532L): Multimodal Learning with Vision, Language and Sound Lecture 12: Deep Reinforcement Learning Types of Learning Supervised training Learning from the teacher Training data includes

More information

Practical Characteristics of Neural Network and Conventional Pattern Classifiers on Artificial and Speech Problems*

Practical Characteristics of Neural Network and Conventional Pattern Classifiers on Artificial and Speech Problems* 168 Lee and Lippmann Practical Characteristics of Neural Network and Conventional Pattern Classifiers on Artificial and Speech Problems* Yuchun Lee Digital Equipment Corp. 40 Old Bolton Road, OGOl-2Ull

More information

Autonomous Sensor Center Position Calibration with Linear Laser-Vision Sensor

Autonomous Sensor Center Position Calibration with Linear Laser-Vision Sensor International Journal of the Korean Society of Precision Engineering Vol. 4, No. 1, January 2003. Autonomous Sensor Center Position Calibration with Linear Laser-Vision Sensor Jeong-Woo Jeong 1, Hee-Jun

More information

Dept. of Adaptive Machine Systems, Graduate School of Engineering Osaka University, Suita, Osaka , Japan

Dept. of Adaptive Machine Systems, Graduate School of Engineering Osaka University, Suita, Osaka , Japan An Application of Vision-Based Learning for a Real Robot in RoboCup - A Goal Keeping Behavior for a Robot with an Omnidirectional Vision and an Embedded Servoing - Sho ji Suzuki 1, Tatsunori Kato 1, Hiroshi

More information

The Fly & Anti-Fly Missile

The Fly & Anti-Fly Missile The Fly & Anti-Fly Missile Rick Tilley Florida State University (USA) rt05c@my.fsu.edu Abstract Linear Regression with Gradient Descent are used in many machine learning applications. The algorithms are

More information

Jo-Car2 Autonomous Mode. Path Planning (Cost Matrix Algorithm)

Jo-Car2 Autonomous Mode. Path Planning (Cost Matrix Algorithm) Chapter 8.2 Jo-Car2 Autonomous Mode Path Planning (Cost Matrix Algorithm) Introduction: In order to achieve its mission and reach the GPS goal safely; without crashing into obstacles or leaving the lane,

More information

Wide area tracking method for augmented reality supporting nuclear power plant maintenance work

Wide area tracking method for augmented reality supporting nuclear power plant maintenance work Journal of Marine Science and Application, Vol.6, No.1, January 2006, PP***-*** Wide area tracking method for augmented reality supporting nuclear power plant maintenance work ISHII Hirotake 1, YAN Weida

More information

10-701/15-781, Fall 2006, Final

10-701/15-781, Fall 2006, Final -7/-78, Fall 6, Final Dec, :pm-8:pm There are 9 questions in this exam ( pages including this cover sheet). If you need more room to work out your answer to a question, use the back of the page and clearly

More information

Research on Evaluation Method of Product Style Semantics Based on Neural Network

Research on Evaluation Method of Product Style Semantics Based on Neural Network Research Journal of Applied Sciences, Engineering and Technology 6(23): 4330-4335, 2013 ISSN: 2040-7459; e-issn: 2040-7467 Maxwell Scientific Organization, 2013 Submitted: September 28, 2012 Accepted:

More information

Clustering with Reinforcement Learning

Clustering with Reinforcement Learning Clustering with Reinforcement Learning Wesam Barbakh and Colin Fyfe, The University of Paisley, Scotland. email:wesam.barbakh,colin.fyfe@paisley.ac.uk Abstract We show how a previously derived method of

More information

Wearable Master Device Using Optical Fiber Curvature Sensors for the Disabled

Wearable Master Device Using Optical Fiber Curvature Sensors for the Disabled Proceedings of the 2001 IEEE International Conference on Robotics & Automation Seoul, Korea May 21-26, 2001 Wearable Master Device Using Optical Fiber Curvature Sensors for the Disabled Kyoobin Lee*, Dong-Soo

More information

Notes on Multilayer, Feedforward Neural Networks

Notes on Multilayer, Feedforward Neural Networks Notes on Multilayer, Feedforward Neural Networks CS425/528: Machine Learning Fall 2012 Prepared by: Lynne E. Parker [Material in these notes was gleaned from various sources, including E. Alpaydin s book

More information

Robustness of Selective Desensitization Perceptron Against Irrelevant and Partially Relevant Features in Pattern Classification

Robustness of Selective Desensitization Perceptron Against Irrelevant and Partially Relevant Features in Pattern Classification Robustness of Selective Desensitization Perceptron Against Irrelevant and Partially Relevant Features in Pattern Classification Tomohiro Tanno, Kazumasa Horie, Jun Izawa, and Masahiko Morita University

More information

Improvement of SURF Feature Image Registration Algorithm Based on Cluster Analysis

Improvement of SURF Feature Image Registration Algorithm Based on Cluster Analysis Sensors & Transducers 2014 by IFSA Publishing, S. L. http://www.sensorsportal.com Improvement of SURF Feature Image Registration Algorithm Based on Cluster Analysis 1 Xulin LONG, 1,* Qiang CHEN, 2 Xiaoya

More information

State Estimation for Continuous-Time Systems with Perspective Outputs from Discrete Noisy Time-Delayed Measurements

State Estimation for Continuous-Time Systems with Perspective Outputs from Discrete Noisy Time-Delayed Measurements State Estimation for Continuous-Time Systems with Perspective Outputs from Discrete Noisy Time-Delayed Measurements António Pedro Aguiar aguiar@ece.ucsb.edu João Pedro Hespanha hespanha@ece.ucsb.edu Dept.

More information

UNIVERSITY OF NORTH CAROLINA AT CHARLOTTE

UNIVERSITY OF NORTH CAROLINA AT CHARLOTTE UNIVERSITY OF NORTH CAROLINA AT CHARLOTTE Department of Electrical and Computer Engineering ECGR 4161/5196 Introduction to Robotics Experiment No. 5 A* Path Planning Overview: The purpose of this experiment

More information

Review on Methods of Selecting Number of Hidden Nodes in Artificial Neural Network

Review on Methods of Selecting Number of Hidden Nodes in Artificial Neural Network Available Online at www.ijcsmc.com International Journal of Computer Science and Mobile Computing A Monthly Journal of Computer Science and Information Technology IJCSMC, Vol. 3, Issue. 11, November 2014,

More information

S-SHAPED ONE TRAIL PARALLEL PARKING OF A CAR-LIKE MOBILE ROBOT

S-SHAPED ONE TRAIL PARALLEL PARKING OF A CAR-LIKE MOBILE ROBOT S-SHAPED ONE TRAIL PARALLEL PARKING OF A CAR-LIKE MOBILE ROBOT 1 SOE YU MAUNG MAUNG, 2 NU NU WIN, 3 MYINT HTAY 1,2,3 Mechatronic Engineering Department, Mandalay Technological University, The Republic

More information

Binary Decision Tree Using K-Means and Genetic Algorithm for Recognizing Defect Patterns of Cold Mill Strip

Binary Decision Tree Using K-Means and Genetic Algorithm for Recognizing Defect Patterns of Cold Mill Strip Binary Decision Tree Using K-Means and Genetic Algorithm for Recognizing Defect Patterns of Cold Mill Strip Kyoung Min Kim,4, Joong Jo Park, Myung Hyun Song 3, In Cheol Kim, and Ching Y. Suen Centre for

More information

EE795: Computer Vision and Intelligent Systems

EE795: Computer Vision and Intelligent Systems EE795: Computer Vision and Intelligent Systems Spring 2012 TTh 17:30-18:45 WRI C225 Lecture 02 130124 http://www.ee.unlv.edu/~b1morris/ecg795/ 2 Outline Basics Image Formation Image Processing 3 Intelligent

More information

A Connectionist Learning Control Architecture for Navigation

A Connectionist Learning Control Architecture for Navigation A Connectionist Learning Control Architecture for Navigation Jonathan R. Bachrach Department of Computer and Information Science University of Massachusetts Amherst, MA 01003 Abstract A novel learning

More information

Vision-based Frontal Vehicle Detection and Tracking

Vision-based Frontal Vehicle Detection and Tracking Vision-based Frontal and Tracking King Hann LIM, Kah Phooi SENG, Li-Minn ANG and Siew Wen CHIN School of Electrical and Electronic Engineering The University of Nottingham Malaysia campus, Jalan Broga,

More information

Using the Kolmogorov-Smirnov Test for Image Segmentation

Using the Kolmogorov-Smirnov Test for Image Segmentation Using the Kolmogorov-Smirnov Test for Image Segmentation Yong Jae Lee CS395T Computational Statistics Final Project Report May 6th, 2009 I. INTRODUCTION Image segmentation is a fundamental task in computer

More information

CS 4758: Automated Semantic Mapping of Environment

CS 4758: Automated Semantic Mapping of Environment CS 4758: Automated Semantic Mapping of Environment Dongsu Lee, ECE, M.Eng., dl624@cornell.edu Aperahama Parangi, CS, 2013, alp75@cornell.edu Abstract The purpose of this project is to program an Erratic

More information

Implementation of a Face Recognition System for Interactive TV Control System

Implementation of a Face Recognition System for Interactive TV Control System Implementation of a Face Recognition System for Interactive TV Control System Sang-Heon Lee 1, Myoung-Kyu Sohn 1, Dong-Ju Kim 1, Byungmin Kim 1, Hyunduk Kim 1, and Chul-Ho Won 2 1 Dept. IT convergence,

More information

Image Feature Generation by Visio-Motor Map Learning towards Selective Attention

Image Feature Generation by Visio-Motor Map Learning towards Selective Attention Image Feature Generation by Visio-Motor Map Learning towards Selective Attention Takashi Minato and Minoru Asada Dept of Adaptive Machine Systems Graduate School of Engineering Osaka University Suita Osaka

More information

Reinforcement Learning-Based Path Planning for Autonomous Robots

Reinforcement Learning-Based Path Planning for Autonomous Robots Reinforcement Learning-Based Path Planning for Autonomous Robots Dennis Barrios Aranibar 1, Pablo Javier Alsina 1 1 Laboratório de Sistemas Inteligentes Departamento de Engenharia de Computação e Automação

More information

CS 4758 Robot Navigation Through Exit Sign Detection

CS 4758 Robot Navigation Through Exit Sign Detection CS 4758 Robot Navigation Through Exit Sign Detection Aaron Sarna Michael Oleske Andrew Hoelscher Abstract We designed a set of algorithms that utilize the existing corridor navigation code initially created

More information

BabyTigers-98: Osaka Legged Robot Team

BabyTigers-98: Osaka Legged Robot Team BabyTigers-98: saka Legged Robot Team Noriaki Mitsunaga and Minoru Asada and Chizuko Mishima Dept. of Adaptive Machine Systems, saka niversity, Suita, saka, 565-0871, Japan Abstract. The saka Legged Robot

More information

Applying the Episodic Natural Actor-Critic Architecture to Motor Primitive Learning

Applying the Episodic Natural Actor-Critic Architecture to Motor Primitive Learning Applying the Episodic Natural Actor-Critic Architecture to Motor Primitive Learning Jan Peters 1, Stefan Schaal 1 University of Southern California, Los Angeles CA 90089, USA Abstract. In this paper, we

More information

Triangular Mesh Segmentation Based On Surface Normal

Triangular Mesh Segmentation Based On Surface Normal ACCV2002: The 5th Asian Conference on Computer Vision, 23--25 January 2002, Melbourne, Australia. Triangular Mesh Segmentation Based On Surface Normal Dong Hwan Kim School of Electrical Eng. Seoul Nat

More information

A STRUCTURAL OPTIMIZATION METHODOLOGY USING THE INDEPENDENCE AXIOM

A STRUCTURAL OPTIMIZATION METHODOLOGY USING THE INDEPENDENCE AXIOM Proceedings of ICAD Cambridge, MA June -3, ICAD A STRUCTURAL OPTIMIZATION METHODOLOGY USING THE INDEPENDENCE AXIOM Kwang Won Lee leekw3@yahoo.com Research Center Daewoo Motor Company 99 Cheongchon-Dong

More information