Research On Human Motion Object Detection And Tracking Algorithm Based On Depth Image

With the rapid development of the smart city in recent years,intelligent transportation and self-driving technology have attracted more and more attention.As the core of self-driving and intelligent monitoring,pedestrian target detection and tracking has a high value of application and research.The effect of traditional visible light target detection and tracking is easily affected by target’s moving,light changes,complex background interference and occlusion.The sequence of images collected by the Kinect depth camera makes it possible to detect and track pedestrian targets accurately.Therefore,it is of considerable significance to study pedestrian target detection and tracking technology based on the depth image.In this paper,pedestrian target detection and tracking are analyzed from the aspects of neighborhood background modeling detection,feature learning detection,and average shift tracking.The main work and innovations of this paper are as follows:(1)For the depth images,I studied and analyzed the problems of neighborhood modeling and detection and optimized them,effectively utilizing the pixel depth feature to suppress ghosting quickly.Since the intersection of pedestrian target and the ground area will cause itslocal detection failure,this paper proposes upper and lower hierarchical modeling.Besides,due to specific errors in imaging,I added the layered distance parameter in the detection mechanism to extract the complete target better.To improve time efficiency,in this paper,I propose a multi-background point update method,which selectively updates the background points and avoids invalid updates.Experimental results show that this optimization mechanism takes less time and can effectively remove ghosts and global denoising.(2)For feature learning detection,I focused on the Hog+SVM way.Because the algorithm is time-consuming based on multi-scale window detection,an optimization method of depth estimation and depth matching detection is proposed.In this method,the original algorithm is only used to detect the initial or specific frames,while the local estimation and matching detection are carried out for the subsequent frames to limit the number of windows and extract the windows with high matching degree of deoth model.When the number of frames reaches a certain threshold,the original algorithm will be enabled to update the detection;Experiments show that this method can effectively improve the detection efficiency compared with other methods.(3)Aiming at the research of the pedestrian target tracking algorithm,I firstly introduce relevant classical algorithms such as Kalman Meanshift.Then,the Meanshift tracking algorithm is mainly studied,and the problems of target jump and size fixation in the Meanshift algorithm are deeply analyzed.Using the depth feature,I proposed a multivalued connectivity technique based on the depth image to segment it to obtain the depth connected object.Then I used the method of connectivity first matching to reduce the number of trace iterations.Meantime,according to the target’s characteristics under occlusion and the model’s matching degree,the occlusion calculation mechanism is established.Finally,the target recovery mechanism is added to update the target model to improve the tracking robustness.Experiments show that the algorithm has better accuracy and robustness.