Dynamic Human Detection And Tracking Algorithm Based On Depth Information

With the rapid development of computer vision and artificial intelligence,intelligent information products are becoming more and more widely used.Whereas the aggravating trend of aging population brings impact to the social system’s sustainable development and the higher labor cost contributes to the price increase,Societies ’ve made an urgent request for service robots.Robot visual technology is a momentous research direction in the robot field,and the detection and tracking of pedestrian as an essential part of human-computer interaction,have become an important branch of service robot vision technology.In the absence of human intervention,the computer can perceive the external environment through a powerful data processing system.By solving,analyzing and dealing with the video image sequence,the service robots can comprehend external environment.And on the basis of trying to understand the behavior of the target,Pedestrian detection and tracking technology can provide a great convenience for the upper level of service robot,so that people can free from the heavy,repetitive monotonous task freed.Visual perception is a complex system,as the visual information contains a great amount of data,from which to extract useful feature information need to implement complex algorithms and time-consuming computing.Therefore,accurate and real-time detection and tracking of pedestrians is the hotspot and important research direction of human-computer interaction in service robots,and it is also a major challenge.In this paper,a pedestrian detection and tracking algorithm for service robots is in depth analyzed for the service robot visual perception system.Under the light interference,occlusion and dynamic background,the algorithm can still achieve real-time effective tracking.The main research work of this paper includes:(1)Target Human Body Extraction Based on Depth ImageIn this paper,the Kinect sensor integrated in the turtlebot robot can obtain the depth image data.Through the analysis of the image acquisition equipment,and understand the relevant software as well as technical principles,the depth of the image information can be processed after the extraction of the target area,solving the initialization problem of tracking process of the target pedestrian location.(2)Improvement of Tracking-Learning-Detection(TLD)Framework and Comparison of AlgorithmsThe target tracking algorithm based on the TLD framework can update the positive and negative samples in real time through training,so that the algorithm can obtain new target appearance features,even if the frame is missing or the camera moves fast,the algorithm can still achieve long-term tracking,performing good accuracy and robustness.But this algorithm can’t automatically lock the target,the block is also likely to cause serious target loss or misdirect.In response to this problem,,this paper improves the TLD algorithm by combining the advantages of particle filter based on the deep image object extraction to achieve more accurate tracking effect.Experiments show that when the pedestrian appears again,the algorithm can quickly achieve the target re-examination.Finally,by discussing and comparing the accuracy and timeliness of each algorithm,the robustness of proposed algorithm is further proved.