Dynamic Human Body Tracking And Recognition For Mobile Robot

With the development of social economy,human hands have been gradually liberated from manual labor,and a lot of manual labor had been carried out by robots.From the initial remote-controlled mobile robots that carry various cargoes,to the current automatic tracking robots that can serve specific personnel,the detection and tracking technology of the target,human body,is its main technological advancement point.In recent years,digital image processing technology has become more and more mature,and tracking and recognizing human body based on visual image processing has become a research hotspot among many scholars and enterprises.There are two methods to do for dynamic human body tracking and recognition for mobile robots the indoor environment: classical methods and deep learning based methods.In the classical method,the location of the human body is obtained mainly based on the feature of the image's color,contour,texture and so on.As for the deep learning based method,convolutional neural networks are used to extract the feature of the image,followed by a classifier to do classification.The details of the research contents are as follows:Firstly,an improved method is proposed for dynamic human body tracking based on the classical method,in which the gray projection method is used to compensate the dynamic background of the mobile robot into a static background.It is combined with the Camshift algorithm,the particle filter algorithm and the background cross-frame algorithm which are suitable for different tracking situations.The method sets evaluation conditions for each tracking algorithm to determine whether the algorithm is effective or not.The tracking process is switched among different algorithms,which can avoid the sho the shortcomings of single tracking algorithm and makes the tracking task more flexible.However,the classical method is only suitable for simple tracking scenarios.It lacks the ability to recognize human identity,so it only applies to single person tracking.Secondly,the simplified Faster-RCNN is combined with the improved Kalman filter algorithm to track the dynamic human body in the system of mobile robot.In this algorithm,Faster RCNN is simplified by removing the original full connection layer,so the human candidate positions are generated only useing the basic convolution neural networks(CNNs)and the regional proposal network(RPN).These positions are taken as the observation values of the improved Kalman filtering algorithm,which is combined with the estimated value of Kalman filter algorithm to get the optimal position.Compared with the classical tracking method,this method has advantages of higher accuracy,high real-time performance and high stability.Thirdly,face alignment and face recognition are performed using multi-task convolutional neural network(MTCNN)and Facenet.The method uses a multi-task convolutional neural network to identify the tracked human body area mentioned above.It can find the position of the face,and find five key points of the face to align it.Then the aligned face images are fed into the Facenet to determine the identity of the human body.Finally,the control method of mobile robot is studied.Considering characteristics of dynamic human tracking and recognition in the environment of mobile robots,this paper acquires tracks images by using a self-developed four-wheel differential robot equipped with a binocular camera.Then the distance and angle between the mobile robot and the moving human body are taken as the input of the robot,the fuzzy control method is used to control the motion of the robot to achieve the tracking task the moving human body.