AUTOMATIC RECOGNITION OF OBJECT DETECTION USING MATLAB

Transcription

1 AUTOMATIC RECOGNITION OF OBJECT DETECTION USING MATLAB A.Anitha 1,J.Gayatri 2,Ashwini.k 3 Department of Electronics and Communication and Instrumentation,VTU University. R.Y.M.Engineering College, Bellary Abstract- Monitoring military, conflicts, illegal immigrants etc. areas rely currently on technology and man power, however automatic monitoring has been advancing in order to avoid potential human errors that can be caused by different reasons. This introduces an automatic recognition of object, which uses image processing to detect and extract moving objects within a restricted area, and a neural network to recognize the extracted object. Experimental results provides a simple, efficient and fast solution to the problem of detecting, extracting and recognizing moving objects within one system. Keywords: surveillance camera, pixel, threshold, segment. 1. INTRODUCTION Automatic recognition systems for still and moving objects can be invalid in security applications, such as monitoring border areas, buffer zones and restricted areas. A simple recognition system would comprise a camera fixed high above the monitored zone, where images of the zone are captured and consequently processed. Processing the captured images can be in three phases, namely, detection of a moving object, extraction of the object and finally recognition of the object. Optical flow and background subtraction have been used for detecting moving objects in image sequences. Other works such as moving objects segmentation using optical flow estimation presented a method for the segmentation of moving objects, where a powerful variation method using active contours for computing the optical flow is used. However, the high computational time to extract the optical flow and the lack of discrimination of the foreground from the background, make this method unsuitable for real time processing. On the other hand, background subtraction detects moving objects by subtracting estimated background models from images. This method is sensitive to illumination changes and small movement in the background, e.g. leaves of trees. Moving object detection is achieved by comparing two subsequent still images from a surveillance camera, captured every two seconds. The difference between the pixel values is calculated and the output is obtained using a deterministic rule. Further processing is also carried out in order to provide a clearer image of the detected object and its surrounding by improving the result image contrast using background compensation. 2. SYSTEM OVERVIEW MOVING OBJECT DETECTION \ An image-processing scheme to detect moving objects in real time is a key technology for the automatic surveillance. 749

2 The first image, which is called reference image, represents the reference pixel values for comparison purpose, and the second image, which is called the input image, contains the moving object. The two images are compared and the differences in pixel values are determined. If the input image pixel values are not equal to the background color (B) of the output image O(x,y), which contains the moving object, will be then set to black (gray level 0 ) or white (gray level 255 ) as follows: B= (3) extraction method is summarized in the following subsections. Fig 2.1: Moving object detection phase in a third image, which is called output image, with a black or white background. The reference and input images are compared by taking their difference where the output of this comparison DRI(x,y) is determined by the following rule: D R (x,y) (1) Where R(x, y) and I(x,y) are the reference and input images, respectively at image coordinates (x,y). The obtained difference matrix is then threshold in order to determine the background color of the third image; output image O(x,y), which contains the moving object. To threshold the difference matrix, we first calculate its average pixel value using: D average = (2) Where n x and n y are, respectively, the numbers of difference matrix pixels in the x and y directions. The Fig. 2.2: Object extraction phase 2.1 HORIZONTAL SCANNING Starting at image coordinates (x=0, y=0) until coordinate (x=n x, y=n y ), the total pixel value of the column is defined as: Total Y X = ` -(5) This total value is compared to a threshold value θx which is defined as: x = n y.b (6) Where n y is the maximum number of vertical pixels (n y = 256 in this work), and B is the gray level value of the image background as defined in equation (3). For white (B=255), whereas for black (B=0).The comparison deterministic rules are defined as follows: For B=255, if Total Y x ( -c) then S x = x if Total Y x <( -C) then E x = x For B =0, if Total Y x <( C) then S x = x 750

3 if Total Y x C) then E x = x Where S x and Ex are respectively the starting and ending x coordinates of the object. C is a correction value that is added to eliminate unwanted pixels which could be due to noise. 2.2VERTICAL SCANNING Starting also at image coordinates (x=0, y=0) until coordinate (x=n x, y=n y ), the total pixel value of the row is defined as: This total value is compared to a threshold value θ y defined as: Total X y = - (7) This total value is compared to a threshold value θ y defined as: n x.b - (8) 2.3 OBJECT RECOGNITION The detected and extracted moving object is recognized by a trained supervised neural network that is based on the back propagation learning algorithm. This algorithm is chosen due to its implementation simplicity and efficiency in pattern classification. Object recognition is a difficult problem due to the large feature space and the complexity of feature dependencies. First, there exist positional complexities resulting from the 3D position and orientation of the object as well as the 3D position and orientation of the camera. Further, changes in lighting, background, and occlusion can create dramatically different images for the same object. In addition, to these complexities, we try to perform object recognition over classes of objects. in the same class. Object recognition is done by neural network concept shown in the figure Where n x is the maximum number of horizontal pixels (n x =256 in this work), and B is the grey level value of the image background as defined in equation (3). The comparison deterministic rules are defined as follows: For B= 255, if Total X y ( then S y = y If Total X y ( ) then E y = y Where S y and E y are respectively the starting and ending y coordinates of the object, and C is the correction value. The size of the extracted object image is (Ex-Sx).(Ey-Sy). The application of this algorithm continues until the row at the last y-coordinate pixel is accounted for. If there is another object within the image, its y-coordinate positions are determined similarly to the first object. Fig. 2.3: Object recognition phase in On the other hand, the generated training data may not be an accurate representation of reality and may create an artificial bias. Furthermore, when dealing with millions of images simultaneously, special precautions must be taken to respect the strict hardware constraints. Here, first discusses the image generation process. Next, it explores two nearest neighbor related algorithms. 751

4 The first uses cover trees, and the second implements modified Nister trees. Finally, the paper ends with future work and conclusions. When computing a classifier for object recognition one faces two main philosophies: generative and discriminative models. Formally, the two categories can be described as follows: 2.4 OBJECT RECOGNITION IN REAL WORLD An object recognition system finds objects in the real world from an image of the world, using object models which are known a priori. This task is surprisingly difficult. In this chapter it will be discussed different steps in object recognition and introduce some techniques that have been used for object recognition in many applications. The object recognition problem can be defined as a labeling problem based on models of known objects. Formally, given an image containing one or more objects of interest (and background) and a set of labels corresponding to a set of models known to the system, the system should assign correct labels to regions, or a set of regions, in the image. The object recognition problem is closely tied to the segmentation problem: without at least a partial recognition of objects, segmentation cannot be done, and without segmentation, object recognition is not possible. to the presence of multiple occluding objects in images. The object recognition task is affected by several factors. We classify the object recognition problem into the following classes. 2.5 SEGMENTED The images have been segmented to separate objects from the background. Object recognition and segmentation problems are closely linked in most cases. In some applications, it is possible to segment out an object easily. In cases when the objects have not been segmented, the recognition problem is closely linked with the segmentation problem. as the spine or axis of the cylinder, a two-dimensional cross-sectional figure, and a sweeping rule that defines how the cross section is to be swept along the space curve. The cross section can vary smoothly along the axis. This representation is shown in Figure 3.5, the axis of the cylinder is shown as a dash line, the coordinate axes are drawn with respect to the cylinder s central axis, and the cross sections at each point are orthogonal to the cylinder s central axis and missing cost can take many different forms. Applications will determine the exact form of these functions. 2.6 IMAGE DATABASE Meaningful training of a neural network is vital if we are to generalize it successfully. Two aspects in training are noted here: firstly the number of images used and, secondly, the sufficiency of the extracted patterns from these images. This section will briefly describe the objects considered in this work; bearing in mind the assumed scenario for this application which is monitoring a secured area with a surveillance camera being positioned higher than the ground level. The assumed secured area is a border controlled area or a buffer zone between two countries, thus the choice of objects in this application. The considered objects in this work are classified into three groups: humans, animals, and vehicles. Images of objects from each group will be used for training and later on generalizing (or testing) the neural network. The objects in each group were as follows: human (Female and Male), animal (Goat), vehicle (Car, Jeep, Motorbike, and Loader). 752

5 excluding two directions which are {3 and 7}, since the object movement will be towards or away from the camera Fig 2.4 Female Male Goat Car 2.7 NEURAL NETWORK ARBITRATION The neural network consists of an input layer with 100 neurons receiving the input feature vector, one hidden layer with 200 neurons which was determined after many experiments involving the adjustment of the number of hidden ROLE OF ROBOT Robot is placed in the area captured region by the camera. Any object coming in this region is detected by camera and the distance and angle is covered with respect to the robot, also it shows the forward and backward direction. Detection is done by MATLAB software process which is shown in figure Detection is done by subtracting the original image to the reference image. Extraction is done on the basis of object position change. Detected object is captured and software sends the command to Robot to follow the object. It will follow until it reached the object. When robot reached the object it sends the command by distance null. For recognition process camera takes the snap of the objects and gives them id for separation. Whatever the suspicious object is, can be commanded by giving the object id in the software by manually. Robot will respond only for that suspicious object and follow the object where it goes in the captured region. Fig. 2.5 Object recognition phase 2.8 ROBOT APPLICATION We are using Robot to show one of the application parts of this topic.here, Robot is the moving object which is commanded by the matlab software. Robot lives only in the captured camera region.when any object or moving object comes in the captured camera region,it is detected by the camera and command is given by the software to Robot follow the object, show the object distance, reach near the object. Robot is having hardware part which is mainly divided into two part Transmitter part and receiver part. 2.9 OBJECT RECOGNITION RESULTS The implementation results of the trained neural network were as follows: using the training image set (36 images) at both tolerance levels yielded 100% recognition as would be expected. The recognition system implementation using the testing image set (56 images that were not previously exposed to the neural network) yielded different correct object recognition according to the tolerance level. At Low tolerance level 39 objects were correctly recognized, thus achieving 70% correct identification rate, whereas, at High tolerance level 51 objects were correctly recognized, thus achieving 91.1% correct identification rate. Combining the results using 753

6 testing images (56) and training images (36), yields an overall correct identification rate of 81.5% with Low tolerance, and 94.6% with High tolerance. Table II shows the extracted objects recognition results. 3. SIMULATION RESULTS AND ANALYSIS 3.1 SIMULATION ANALYSIS The simulation analysis is the following analysis. Fig 3.3 object segmented Fig 3.1 Robot and object detection 3.2 RESULT ANALYSIS Robot is masked and object detected Distance of the object with respect to the Robot is cm. Angle of the object with respect to the Robot is rad. Position left from the Robot. Finally, the robot has reached. Fig 3.4 object ID creation 3.3 RECOGNITION OF THE OBJECT Object is placed in the secured region, database is created and tested. For recognition purpose ID is created. Fig 3.5 object recognition Fig 3.2: Four object is placed. Object can even recognized when they are partially obstructed from view. Deterministic rule provide a clear image for the detected object Multi detection, extraction and multi recognition objects are possible within one system 754

7 4 CONCLUSION modified Nester trees on localized image patches, accuracy generally improved with training set size. The proposed system receives still images captured every two seconds from a surveillance camera, which monitors a restricted zone, such as international border crossings, buffer zone in conflict areas, or any monitored area where movement across that area is to be detected. The implementation of three phases: Firstly, moving object detection which is achieved using image pixel value difference, a deterministic rule to determine the\ moving object, and background compensation. Secondly, the extraction of the detected object, which is achieved by using a set of deterministic rules to find pixel variations within the image of the detected object(s) and eliminating phantom objects that may have been obtained in the first phase. This second phase has also further processing of the extracted object image, such as squaring, framing and scaling the image to a predefined size in preparation for the next phase. Finally, the third phase is the recognition of the extracted object using a supervised neural network based on the simple but efficient back propagation learning algorithm. The proposed system provides solutions to the problem of monitoring secured areas; namely, the detection of movement across the area, the extraction of the moving object, and the recognition of the object. The object recognition will benefit from tens of millions or hundreds of millions of generated images. The only obstacle in answering that question lies in transforming the problem and representation for efficient use given memory and computation constraints. REFERENCES [1]. A. Petrosino, "Moving Object Detection for Real- Time Applications," in Proc. IEEE 14 th Int. Conf. Image Analysis and Processing, Modena, Italy, Sep. 2007, pp [2] G. Zhang, J. Jia, W. Xiong, T. T. Wong, P. A. Heng, and H. Bao, Moving Object [3 ]L. Maddalena, and A [4] H. Fujiyoshi, and T. Kanade, Layered detection for multiple overlapping objects,. [5] C. Zhan, X. Duan, S. Xu, Z. Song, and M. Luo,. [6] G. Chen, H. Zhou, and J. Yan, A Novel Method for Moving Object Detection in Foggy Day, 5. FUTURE WORK The object detection and classification can be significantly improved by using large datasets of generated images. By using cover trees on the pixel values of images and 755

8 Author s profile ISSN: X A. Anitha The author A.ANITHA is of native From BellaryDistrict of Karnataka,India.She Born at Date-ofbirthis This author completedm.tech in DigitalElectronics form BITM Engg College.She is working as Assistant Professorof ECE department inrymec college of Engineering and Technology,Bellary,for the past Two years.her area of interest includes DigitalElectronics. J. Gayatri The author J.Gayatri is of native from Bellary District of Karnataka, India. She Born at Date-ofbirth is This author completed M.Tech indigital Electronics form BITM, Bellary. She is working as an Assistant Professor of IT department in RYM Engineering college of Bellary for the pastthree years. Her area of interest includes Image processing. K.Ashwini. The author Ashwini.k is of native from Bellary District of Karnataka, India. She Born at Date-of birth is This author completed M.Tech indigital Electronics form BITM, Bellary. She isworking as an Assistant Professor of IT departmentin RYM Engineering college of Bellary. Her area of interest includes Imageprocessing. 756