Micro-Doppler Based Human-Robot Classification Using Ensemble and Deep Learning Approaches

Transcription

1 Miro-Doppler Based Human-Robot Classifiation Using Ensemble and Deep Learning Approahes Sherif Abdulatif, Qian Wei, Fady Aziz, Bernhard Kleiner, Urs Shneider Department of Biomehatroni Systems, Fraunhofer Institute for Manufaturing Engineering and Automation IPA {sherif.abdulatif, qian.wei, fady.aziz, bernhard.kleiner, These authors ontributed to this work equally. arxiv: v [s.cv] 6 Feb 18 Abstrat Radar sensors an be used for analyzing the indued frequeny shifts due to miro-motions in both range and veloity dimensions identified as miro-doppler (µ-d) and miro-range (µ-r), respetively. Different moving targets will have unique µ-d and µ-r signatures that an be used for target lassifiation. Suh lassifiation an be used in numerous fields, suh as gait reognition, safety and surveillane. In this paper, a 5 GHz FMCW Single-Input Single-Output (SISO) radar is used in industrial safety for real-time human-robot identifiation. Due to the real-time onstraint, joint Range- Doppler (R-D) maps are diretly analyzed for our lassifiation problem. Furthermore, a omparison between the onventional lassial learning approahes with handrafted extrated features, ensemble lassifiers and deep learning approahes is presented. For ensemble lassifiers, restrutured range and veloity profiles are passed diretly to ensemble trees, suh as gradient boosting and random forest without feature extration. Finally, a Deep Convolutional Neural Network (DCNN) is used and raw R-D images are diretly fed into the onstruted network. DCNN shows a superior performane of 99% auray in identifying humans from robots on a single R-D map. I. INTRODUCTION Identifying a moving target and lassifying its motions are grasping great interest in many fields. Numerous safety and surveillane appliations now require systems that an be used for human motion identifiation, whih are known as biometri systems. Due to reent development in the field of industry, robots are now widely used for assembly line automation. Therefore, the need of a safe Human Robot Interation (HRI) environment is highly needed to avoid potential threats due to robots powerful movements [1]. In order to provide a safe HRI in suh unstable environments, a reliable, real-time and aurate human detetion system is required. Camera-based systems are used for human detetion based on feet and head reognition of the objets in the sene []. In [], infrared ameras are used to enhane night vision human detetion for surveillane appliations. However, all of the vision-based sensors suffer limitations in different lighting and weather onditions. In industrial safety appliations, the standardized sensor for safe human detetion is LIDAR (Light Detetion and Ranging), where refletions from human legs at knee level are proessed to identify presene of a human in the sanned area [4]. LIDAR sensors, though have many limitations in deteting refletions from dark surfaes and problems with outdoor harsh environmental onditions. Unlike other vision based sensors, radar systems an still detet targets behind obstales or hard surfaes and an work in harsh outdoor irumstanes [5]. Due to developments in radar tehnology, espeially in the miro-doppler (µ-d) field, radar an be used nowadays to extrat target s bulk motion (Range-Veloity) in addition to miro-motions as limbs swings or even very fine vibrations due to vital signs [6]. Reently, µ-d effet in radar has been extensively addressed in different appliations related to biometri systems, suh as gait reognition and limbs deomposition [7]. The author in [8] foused on distinguishing human target from different objets, suh as animals based on their orresponding µ-d signatures. Others foused on using radar for safe pedestrian reognition in automotive industry to apply an emergeny brake to the ar when a pedestrian is deteted at a lose distane [9]. Most of the previously mentioned papers applied human detetion based only on veloity analysis, where CW radars were used to extrat µ-d signatures of the whole human gait yle i.e, more than 1 seond duration. Signatures are used with either onventional mahine learning or Deep Convolutional Neural Networks (DCNN) approahes on the olleted dataset to detet possible human presene. However, these tehniques did not address the real-time lassifiation aspet. Moreover, the size of the training dataset proposed for designing a lassifier is in the order of hundreds. Suh amount is not enough for designing robust lassifiers, espeially in DCNN, as it is proven that for the training of a DCNN, more training data an inrease the model performane [1]. In this paper, the human-robot lassifiation problem is addressed on olleted FMCW radar Range-Doppler (R-D) maps that an be omputed in muh shorter time intervals (tenth of a seond) [11]. Aordingly, real-time onstraint an be fulfilled and the amount of data used to train deep models inreased from hundreds to thousands of datasets. Finally, a omparison between different learning approahes is presented. The paper is organized as follows: Setion II introdues the dataset preparation and R-D map omputations. Setion III shows the use of onventional learning approahes on hand rafted features. A novel data-driven approah based on ensemble tree lassifiers is presented in Setion IV. In Setion V, a DCNN is designed and applied on R-D maps as images. Finally, all approahes are ompared with suggested future work in Setion VI. II. DATASET PREPARATION As mentioned above, R-D maps are used to build the dataset for all presented learning approahes. A proedure for R-D mapping was proposed in [1]. As illustrated in Fig. 1, the proedure applies Fast Fourier Transform (FFT) to eah measurement with a speified FFT length (N F F T ) on

2 both dimensions of the data matrix. Eah olumn stands for one hirp with N s samples, while one measurement onsists of N p hirps. Aordingly, eah R-D map an still ontain abundant information about the µ-d and the µ-r of the moving objet, as long as the range and veloity resolutions of the radar are suffiiently high. A 51 FFT is used to ahieve aeptable resolution on both dimensions. After FFT, the range information is estimated within eah hirp, while the Doppler data is estimated aross all hirps in one sample. In this way, one R-D map an be aquired from eah data matrix. Eah element in the data matrix speifies the bak-sattered power at these partiular range and veloity. These maps an be visualized as a (51 51) heat map, where Doppler is represented on the horizontal axis, ranges on the vertial axis and the olor RGB values from to 55 represent the baksattered power. D-FFT Ns A. Radar Parametrization Np Fig. 1: R-D mapping proedure. # Measurements The FMCW radar used for data olletion operates at a arrier frequeny f = 5 GHz at a maximum operating bandwidth of B = GHz with a hirp sweeping time of T p =.5 ms. Based on the given bandwidth and Eq. 1, a reasonable range resolution for our appliation is derived as R res = 7.5 m, where is the speed of light. Sine the proposed human robot detetion is required on a room level, the maximum measurement range is limited to R max = 5 m. Aordingly, the number of samples per hirp ( see Eq. ) an be alulated as N s 67 samples. R res = (1) B N s = BR max In [7], infrared motion aptured data was olleted on different walking human subjets and it was proved that the feet has the maximum swinging veloity omponent that an reah up to 4.5 m/s. Based on Eq., a maximum veloity v max = 6 m/s an be ahieved with the given arrier frequeny and hirp duration. This shows that our attained maximum veloity an safely over human walking test subjets. For the requirement of human robot lassifiation, a veloity resolution of v res =.1 m/s is proposed. Aording to Eq. 4, the number of hirps per measurement to the next power of is N p = 18 hirps. Using the given hirp duration (T p ), one R-D map of N s N p size will theoretially take a measurement duration of 64 ms. After taking proessing delays into onsideration, a final measurement duration of one R-D map is still lower than.1 s. This duration is far better for a real-time lassifiation onstraint. Based on the proposed parameters, both range and () veloity resolution are high enough to indue R-D maps that ontain suffiient information about the µ-d and the µ-r of a omplex human target. All presented radar design equations are derived in [1]. v max = () 4f T p B. Dataset Build-up N p = f T p v res (4) To ollet the data, measurements have been arried out on 1 test walking human subjets inluding different heights and genders. There exist unountable types of mahines under the name robot. Sine industry is addressed for the proposed task, two industrial robots are inluded as subjets for data olletion and testing. The first robot is a moving mobile robot assistant (Care-O-bot) developed by the robotis department in Fraunhofer IPA [1]. The seond robot is a 6-axis roboti arm developed by Stäubli [14]. The subjets were moving in random aspet angles in the radar detetable area during eah experiment. However, the ase of exat lateral motion was not inluded to avoid the extreme radial veloity fading effets mentioned in [9]. During the data olletion, eah data sample is labeled with the urrent target lass. This labeling is required for supervised learning to allow models to know the standard solution, and thus learn the right model parameters. The olleted data together with the labels is divided into two subsets. The first subset will be used for onstrution of the model. This subset is then divided into two parts, one part alled training set is used to learn the model parameters, while the other part named validation set is used to tune the model hyper-parameters, suh as the number of hidden layers and neurons of eah layer in a neural network. The seond subset (test set) is employed to assess the generalization performane of the final model by omparing the model preditions with the true lass labels. The hold-out method is used to reate the test set. For a hold-out split, the omplete dataset D is divided into two disjoint subsets. One is used as the training set S, the other is employed as test set T. Mathematially, it an be expressed as: D = S T, S T =. The models are trained on S and tested on T to gain a glimpse about the generalization performane. Sine ontiguous samples from one measurement an be very similar due to the relatively high sampling frequeny, a test on randomly hosen samples from all measurements annot reflet the true generalization performane of the model. Consequently, separate experiments of eah objet type is hosen to build the test set, suh that it takes up to % of the entire data of this type. By this means, the whole test set ontains around % of all olleted data. Finally, a human-robot dataset with 774 training samples and 189 test samples is obtained, in both of whih the amount of human samples and robot samples are omparable. C. R-D Interpretations The human walking motion is desribed as suessive periodi yles in whih two phases an fully desribe a gait yle [6]. The first phase (swinging) in whih there is only one swinging foot and the other is touhing the ground. Within this phase, the human appears on the R-D map as a

3 broad distribution in both R-D axes as shown in Fig. a. This distribution represents a variety of veloities due to the bulk moving body parts (torso and head), in addition to the swinging effet of different body limbs (arms, legs and feet). In the seond phase (stane), no limbs swinging and only bulk motion is observed. This phase is the dominant phase oupying 6% of the gait yle. By omparing robot motion and human stane phase shown in Fig. b and Fig., respetively. It is lear that the stane phase represents the main hallenge in our humanrobot differentiation task, sine both R-D maps are very similar in the narrow horizontal distribution. Range[m] Veloity[m/s] (a) Human in swinging phase. Range [m] Range [m] 1 - Veloity [m/s] () Robot in motion. - Veloity [m/s] (b) Human in stane phase. Fig. : Comparison between R-D maps of human and robot. III. CONVENTIONAL LEARNING ON HANDCRAFTED FEATURES After obtaining the dataset as desribed in the previous setion, we tested several onventional mahine learning methods with hand-rafted features extrated from the R-D maps. Before the feature extration, the multi-otsu method is used on R-D maps as an unsupervised image thresholding to extrat the R-D data orresponding to the target from the bakground [15]. By leveraging it, the original ontinuous RGB values in eah R-D map are quantized into 1 disrete levels. The lowest 5 levels from the 1 are negleted sine they an be onsidered as noise. Fig. illustrates the effet of applying multi-otsu method to the R-D map of a human swinging phase shown in Fig. a. Range [m] Veloity [m/s] Fig. : The R-D map of human after multi-otsu thresholding Subsequently, features are extrated from eah R-D map to represent the Human/Robot differenes disussed previously. The features are also hosen, suh that they do not reflet an exat veloity or range to avoid overfitting. Thus, the extrated features only onsiders the distribution over R-D maps. The distribution an be represented as two features for the range and veloity profiles, whih are omputed as the differenes between the maximum and the minimum deteted values in both R-D dimensions. Moreover, features as the variane in veloities σv and in ranges σr an be seen as the polynomial features of the standard deviation in both veloities σ v and ranges σ R with a degree of [16]. Furthermore, the ovariane between range and Doppler values is also onsidered as a feature. Finally, we have 7 features used with different onventional mahine learning tehniques. After the feature extration, several lassial mahine learning methods have been implemented, trained and evaluated with the feature data. The employed methods are (a) Deision Tree, (b) Logisti Regression, () Support Vetor Mahine (SVM) and (d) K-Nearest Neighbors (K-NN). However, depending on features from one single R-D map, the performane of all methods is less than satisfatory. This ondued to the use of feature vetor sequene aumulated from several suessive R-D maps (a so-alled sample buffer). Eah feature vetor sequene is onatenated by feature vetors extrated from all suessive samples in a buffer. Aordingly, the number of features used will inrease from 7 (in the ase of a buffer of size 1) to 7 (in the ase of a buffer of size 1). As shown in Fig. 4, the lassifiation auraies of all the methods inrease as the buffer size inreases. The best test auray is reahed using SVM as 95.% at buffer size 1. However, suh large buffer size indues a lateny of more than 1 seond, whih leads to problems in safety ritial real-time appliations. Test Auray [%] Deision Tree Logisti Regression Support Vetor Mahine K-Nearest Neighbors Buffer Size Fig. 4: Auray urves of onventional learning methods. IV. ENSEMBLE TREES WITH RESTRUCTURED R-D DATA There are two main drawbaks of using onventional mahine learning methods with hand-rafted features. On one hand, manual onstrution of features from raw data is timeonsuming and requires suffiient domain knowledge. On the other hand, as desribed in the previous setion, suffiient performane an only be obtained at the ost of a large buffer size, whih diretly orrelates with the inferene lateny. A. Ensemble Learning Generally, the preditive power aquired by ombining a bunh of models is better than only using one single model.

4 An Instane of this idea is ensemble learning, a family of mahine learning methods whih perform the learning task by onstruting a group of individual learners and ombine their outputs together as the final output. In mahine learning, one must always be faed with the bias-variane trade-off. Bias and variane are two types of error of a preditive model. A simple model is prone to underfitting of the training data; thus, having high bias, but low variane. Conversely, omplex models tend to overfit the training set and thus having low bias, but large variane. From this dilemma, two opposite proedures an be proposed for dereasing predition error: reduing the variane of omplex models and reduing the bias of simple models. Krogh and Vedelsby proved in [17] that the error of an entire ensemble Ê an be determined by: Ê = E D (5) where E represents the average error of all individual learners, while D evaluates the degree of diversity of individual learners. This indiates that, in order to redue the preditive error of an ensemble model, the individual learners should be diverse. Two ommon types of ensemble learning are bagging and boosting. To realize the diversity of individual learners, both bagging and boosting leverage varying training sets, on whih individual learners are trained. The differene lies in how the varying training sets are obtained. Bagging is the abbreviation of bootstrap aggregating, whih dereases the predition error by reduing the variane of omplex individual learners. In bagging, varying training sets are built by randomly sampling from the whole dataset with replaement (bootstrap sample). After building a predefined K training sets, K individual learners will be trained on these K training sets. This means that, the individual learners an be generated in parallel; hene, there is no strong dependeny between them. The hypothesis of the entire ensemble an be aquired by unweighted averaging of the hypotheses of all K individual learners. Therefore, the estimated bias remains unhanged, while the estimated variane dereases by a fator of K 1 [16]. Boosting improves the predition performane by reduing the bias of weak individual learners. It onstruts diverse training sets by iteratively assigning weights to data samples. The weight, with whih eah data sample is attahed depends on how well this data sample an be predited by the urrent ensemble. By doing so, the training data distribution is modified, and thus resulting in more attention to the data portion whih is not well predited so far. Weak learners are simple models whih an learn the training data with an auray not muh higher than 5% (with a high bias and a low variane). Shapire proved in [18] that a group of weak learners ould be ombined into a strong ensemble ahieving arbitrarily high training auray. Boosting employs this idea and onstruts a set of weak learners sequentially. Eah individual weak learner is indued with the urrent weighted training set obtained in the manner desribed previously. After generating the predefined number of individual K learners, the ensemble hypothesis is obtained by weighted vote of preditions made by all weak learners. By ombining weak learners that fous more on urrently mispredited samples, both bias and variane of an ensemble will gradually derease. B. Restruture of R-D Map Before feeding the lassifiers, the obtained R-D map should be first restrutured. Firstly, R-D map is of a two-dimensional struture 51 51, whih must be flattened to a onedimensional feature vetor before feeding into the ensemble lassifiers. To onvert R-D maps into feature vetors, elements of eah R-D map are averaged along both dimensions. This results into a two 51-dimensional vetors obtained; one representing the Doppler profile (row vetor) and the other one representing the range profiles (olumn vetor). Seondly, to guarantee a rational lassifiation based on target motion dynamis, instead of the absolute values, suh as veloity or range. The information regarding absolute measurement values should be removed from the data, suh that only the R-D distributions would be used. We propose a method to eliminate suh information ontaining onrete target motion parameters as follows: in both Doppler and range profile vetors, the elements orresponding to high power areas (distribution) are shifted to the middle of eah vetor. Sine the positions of these elements in both Doppler profile and range profile orrelate to the absolute veloity and range of the target, respetively. By shifting the large-valued elements in both vetors to the middle, lues to the veloity and range of the target are eliminated. This algorithm works by normalizing the power values of both veloity and range profiles to the sum of all of their elements, respetively, to get a weights vetor for eah. Then, a weighted average is applied to both veloity and range indies based on their orresponding weights, to get an average value lose to the high power area in eah profile. Aordingly, the profiles are shifted from the omputed indies to the middle of the vetors. To redue the omplexity, 18 elements are removed from both ends of the shifted vetors to get 56 elements per profile. After the dimension redution, one 51-dimensional feature vetor is built by onatenating both restrutured Doppler and range profiles. C. Performane To study the feasibility of applying ensemble learning to restrutured R-D maps, random forest and gradient boosting are tested. They an be regarded as outstanding representatives of bagging and boosting, respetively. As shown in Table I, the random forest ahieves a worse performane ompared to the gradient boosting. This an be explained by two reasons: (a) Aording to [19], boosted trees perform better than the random forest for a low dimensionality problem with a data dimension up to 4. (b) Gradient boosting an ahieve better results than random forest for a binary lassifiation problem []. Thus, gradient boosting is onsidered as a better hoie for further investigations and omparisons. TABLE I: Auray of random forest and gradient boosting. Training Auray Test Auray Random Forest 99.88% 9.4% Gradient Boosting 1.% 97.85%

5 V. CNN WITH R-D MAPS The CNNs proved to be extremely effetive in image lassifiation, and sine R-D maps are essentially images as well. Therefore, the appliation of CNNs is reasonable in this senario. As previously illustrated in Fig., a typial R- D map of human has a broad horizontal distribution whih represents a variety of veloities of different body parts. Through omparison, one an see that the R-D map of a robot will only have a narrow horizontal distribution due to its rigid body motion. For human eyes, the differene between both patterns is already distinguishable. Aordingly, a CNN is able to differentiate them as well. For the proposed CNN model, an input image size of ( ) is used. This provides a trade-off between performane and proessing time. The graysale olor mode is employed due to the following reason: in R-D maps, the olor represents the bak-sattered power whih orrelates to the target position relative to the radar and an be affeted by metalli parts on targets (e.g., wearable metal artiles suh as wathes or rings on human targets). In our approah, all of this information should not be onsidered. Compared to the graysale, the RGB olor mode is more sensitive to the noise resulting from unwanted objets and lutters. Furthermore, by using one-hannel graysale images (shown in Fig. 5) as input, the omputational omplexity of both training and predition is redued. (a) Human (b) Robot Fig. 5: Graysale R-D maps of human and robot fed to CNN. A. Network Arhiteture and Training The network arhiteture used in our approah is inspired by Leun s LeNet-5 [1]. As illustrated in Fig. 6, it ontains a stak of 6 onvolutional layers with Retified Linear Unit (ReLU) ativation funtion. Eah onvolutional layer onsists of 16 onvolutional kernels with a size of and is followed by a max-pooling layer. Furthermore, there is a fully-onneted layer onsisting of 16 neurons after the onvolutional layer stak. The output layer at the end has one neuron with sigmoid ativation whih is fully onneted to the 16 neurons of the previous layer. The hoie of the optimizer has an immediate effet on the result of the training, as well as, the required time. For the training of our proposed CNN, the modern adaptive optimizer Adam proposed in [] is employed. The Adam optimization algorithm is urrently one of the most popular algorithms for training various types of Deep Neural Networks (DNN), suh as CNNs in omputer vision appliations and Reurrent Neural Networks (RNNs) for natural language proessing. It enhanes the lassial stohasti gradient desent algorithm by enabling the omputation of individual adaptive learning rates for different parameters. In this manner, the Adam optimizer delivers good optimization results, while maintaining a fast onvergene speeds. Dropout is a simple and yet effetive regularization tehnique proposed in [] to prevent a deep learning model from omplex o-adaptations on training data, namely overfitting. It randomly ignores neurons to a predefined ratio (i.e., dropout rate) during training. During the forward propagation of eah training step, the ignored neurons an not ontribute to the ativation of their onneted neurons in latter layer temporally. Moreover, the weights of the ignored neurons is not updated during the bak propagation. In our ase, dropout has been applied to the last fully-onneted layer in all implementations and a dropout rate of.5 is hosen. The hoie of this hyperparameter is based on trial-and-error. After the best validation auray, ahieved in the th training epoh, both validation auray and loss, start to osillate around the same level. This indiates that the model already fit the data to its best. B. Performane The proposed CNN ahieves the best performane and outperforms the boosting-based approah. Both training and test auraies ahieved auraies of 98.4% and 99.65%, respetively. Table II, illustrates the onfusion matrix of the trained CNN model performed on the test set. Aordingly, the mislassifiation in both lasses is about.5% of the ases, whih is suitable for the required task. TABLE II: Human-Robot onfusion matrix of CNN. True/Predited Human Robot Human 99.4%.56% Robot.6% 99.44% VI. CONCLUSION This work presents the use of FMCW radars to distinguish humans from robots in an industrial environment based on their veloity and range distributions. Sine the required task involves aspets of human safety, a real-time detetion tehnique is proposed based on R-D maps that onsume muh shorter time intervals than µ D signatures. Aordingly, the use of R-D maps inreases the number of datasets used for designing lassifiers, thus a more robust lassifiation is expeted. Moreover, the range dimension an be used as an additional feature in lassifiation and an help in determining how far a target is from the radar. Based on the target s lassified type and range from the radar, speial ations an be taken in suh situations. About 1, equally distributed R-D maps are olleted from different human and industrial robot subjets. To generalize the lassifiation on different surrounding lutters and noise, the measurements are all taken in different test areas. Furthermore, during eah experiment test subjets are moved in front of the radar with different aspet angles to address different motion patterns and angles of inident. The olleted R-D maps are then applied to different mahine learning and deep learning approahes.

6 Fig. 6: Proposed CNN arhiteture on R-D maps. Conventional learning tehniques are first applied for the Human/Robot lassifiation problem based on handrafted features. The features are extrated to inlude the distribution in both R-D dimensions without speifi ranges or veloities. Seven features are extrated from eah R-D map and applied to lassial learning approahes as SVM and K-NN. Unfortunately, the ahieved performane at one R-D map is always lower than 9% whih is not suffiient for our task. To ahieve an aeptable performane, the seven features are extrated over suessive R-D maps and onatenated as a sequene of feature vetors. In that ase, SVM attained the best test auray of 95% on 1 suessive buffers. Although the ahieved auray is aeptable, 1 buffers requires a lateny of 1 s on our urrent radar parametrization. This refleted that onventional tehniques affets the required real-time aspet. This motivated the demonstration of a novel ensemble tree learning lassifiers on restrutured R-D data. In this part, a - dimensional mean is omputed on eah R-D map to get range and Doppler profiles. The extrated profiles are then shifted and restrutured to inlude the patterns without exat veloity or range values. The restrutured R-D data is onatenated in one vetor as 51 samples and fed into random forest and gradient boosting lassifiers. The gradient boosting attains a performane of 97% on one R-D map, outperforming the random forest and onventional tehniques. Finally, a 6 layers CNN was trained on graysale R-D maps and the trained network was shown to outperform all other tehniques with a test auray of 99% on one R-D map (lateny of.1 s). This work addresses only the lassifiation of a single moving target in the radar area. Extending the idea with Multiple-Input Multiple-Output (MIMO) radars an inlude estimating the angle of arrival; thus, the exat position of a moving target an be identified. Based on suh positioning, the detetion and lassifiation of multiple subjets in the test area an be addressed. The presented approahes inluded the lassifiation of moving human and robots only. However, stati human detetion must also be onsidered by means of vital signs or deteted radar ross setions. REFERENCES [1] M. Vasi and A. Billard, Safety issues in human-robot interations, in IEEE International Conferene on Robotis and Automation (ICRA), May 1, pp [] J. K. Suhr and H. G. Jung, Rearview amera-based bakover warning system exploiting a ombination of pose-speifi pedestrian reognitions, IEEE Transations on Intelligent Transportation Syst., Jun. 17. [] T. Takeda, K. Kuramoto, S. Kobashi, and Y. Hata, A fuzzy human detetion for seurity system using infrared laser amera, in IEEE International Symposium on Multiple-Valued Logi (ISMVL), May 1, pp [4] T. Taipalus and J. Ahtiainen, Human detetion and traking with kneehigh mobile D LIDAR, in IEEE International Conferene on Robotis and Biomimetis (ROBIO), De. 11, pp [5] Q. Lu, C. Liu, Z. Zeng, J. Li, and X. Zhang, Detetion of human s motions through a wall using UWB radar, in IEEE International Conferene on Ground Penetrating Radar (GPR), Jun. 16, pp [6] V. C. Chen, The miro-doppler effet in radar. Arteh House, 11. [7] S. Abdulatif, F. Aziz, B. Kleiner, and U. Shneider, Real-time apable miro-doppler signature deomposition of walking human limbs, in IEEE Radar Conferene (RadarConf), May 17, pp [8] Y. Kim and T. Moon, Human detetion and ativity lassifiation based on miro-doppler signatures using deep onvolutional neural networks, IEEE Geosiene and Remote Sensing Letters, vol. 1, no. 1, pp. 8 1, Jan. 16. [9] A. Bartsh, F. Fitzek, and R. Rasshofer, Pedestrian reognition using automotive radar sensors, Advanes in Radio Siene: ARS, vol. 1, Jan. 1. [1] I. Goodfellow, Y. Bengio, A. Courville, and Y. Bengio, Deep learning. MIT press Cambridge, 16, vol. 1. [11] B. Jokanovi and M. Amin, Fall detetion using deep learning in range-doppler radars, IEEE Transations on Aerospae and Eletroni Systems, vol. 54, no. 1, pp , Feb. 18. [1] B. J. Lipa and D. E. Barrik, FMCW signal proessing, FMCW signal proessing report for Mirage Systems, 199. [1] U. Reiser, T. Jaobs, G. Arbeiter, C. Parlitz, and K. Dautenhahn, Careo-bot R vision of a robot butler, in Your virtual butler. Springer, Jan. 1, pp [14] S. Robots, 17. [Online]. Available: [15] P.-S. Liao, T.-S. Chen, and P.-C. Chung, A fast algorithm for multilevel thresholding, Journal of Information Siene and Engineering, vol. 17, no. 5, pp , Sept. 1. [16] Z. Zhihua, Mahine Learning. Tsinghua University Press, 16. [17] A. Krogh and J. Vedelsby, Neural network ensembles, ross validation, and ative learning, in Advanes in neural information proessing systems, 1995, pp [18] R. E. Shapire, The strength of weak learnability, Mahine learning, vol. 5, no., pp , Jun [19] R. Caruana, N. Karampatziakis, and A. Yessenalina, An empirial evaluation of supervised learning in high dimensions, in Proeedings of the international onferene on Mahine learning. ACM, Jul. 8, pp [] M. Fernández-Delgado, E. Cernadas, S. Barro, and D. Amorim, Do we need hundreds of lassifiers to solve real world lassifiation problems, Journal of Mahine Learning Researh, vol. 15, no. 1, pp , Jan. 14. [1] Y. Leun, L. Bottou, Y. Bengio, and P. Haffner, Gradient-based learning applied to doument reognition, in Proeedings of the IEEE, Nov. 1998, pp [] D. P. Kingma and J. Ba, Adam: A method for stohasti optimization, De 14. [Online]. Available: [] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, Dropout: a simple way to prevent neural networks from overfitting. Journal of mahine learning researh, vol. 15, no. 1, pp , Jan. 14.