Visual Target Tracking Algorithm Based On Multi-feature And Random Set Filter

Visual target tracking technology has always been a hotspot and focus in the field of computer vision,especially for unknown and time-varying visual multi-target tracking.And it has important research significance in the fields of military security,national defense construction and civil.In recent years,deep learning has been widely used in visual target tracking,and visual target tracking technology based on depth feature of neural network and correlation feature detection technology makes a breakthrough progress.This paper focuses on the application of multi-Bernoulli(MB)filter in visual multi-target tracking,and make a thorough and systematic research base on depth learning technology.The content and main achievements are as follows:In an unknown and time-varying scenario,visual multi-target tracking has the problem of complex object modeling and anti-interference between close targets and background clutter.Therefore,this paper proposes a visual multi-target tracking based on convolution feature and Multi-Bernoulli filter.In feature extraction,the proposed algorithm gets the convolution kernel by K-means method,and convolutes with the target image to get the target features.In order to reduce the interference of the background,the algorithm fuses the target feature and the background feature which is obtained by sampling the background,and gets a high robustness and accurately feature.In tracking filtering,the algorithm establishes the target likelihood model by the convolution feature of the target,and accomplishes the unknown and time-varying visual multi-target tracking based on MB filter.At the same time,this paper proposes an adaptive template updating strategy to avoid the template damage which is caused by the target adjacent and target occlusion.The strategy updates the target template adaptively by comparing the speed of target state change and the degree of target occlusion,which can improve the tracking accuracy.Finally,the proposed algorithm is validated by infrared and color video data in standard database.The results show that,compared with the traditional MB filter,the proposed algorithm can adapt to the target state change,effectively solve the problem of target adjacency and occlusion in complex scenario,and has higher tracking accuracy.Aiming at the uncertainty of target's birth time and state in multi-target visual tracking,which leads to errors in state estimation,or even target missing,this paper introduces a detection algorithm based on deep learning,and proposes a visual multi-target tracking algorithm based on SSD detection and MB filter.The proposed algorithm detects the targets by SSD,and approximates the probability distribution of the new targets in the form of random finite sets.Then,the algorithm accomplishes the visual multi-target tracking based on MB filter.The algorithm can effectively improve the tracking accuracy,because it can not only solve the problems of missing detection and false detection in detection algorithm,but also avoid the problem of target estimation offset after continuous recursion of MB filter.Finally by testing the standard visual data,the experimental results show that the proposed algorithm can effectively deal with the complex target regeneration problem,and improve the tracking accuracy of visual multi-target because of fusing the detection result and the filter result.In order to solve the problem that traditional MB algorithm cannot get the complete track of each target and multiple targets are still indistinguishable,this paper proposes a visual multi-target tracking algorithm based on Generalized Label Multi-Bernoulli(CLMB).The proposed algorithm brings the concept of label into the theory of random finite sets,labels the multi-Bernoulli component,and uses ?-GLMB filtering algorithm to achieve multi-target tracking.By using the conjugate relationship between label and target state,the proposed algorithm can get the complete track of each target.The experimental results show that the proposed algorithm can obtain accurate track of each target and improve the tracking accuracy in a certain extent.