Research On Visual Multi-Target Tracking Algorithm Based On Detection And Multi-Bernoulli Filter

Visual multi-target tracking is an important branch of target tracking field.As a classic computer vision problem,multi-target tracking has a wide range of applications in the fields of intelligent video surveillance,intelligent traffic control and driverless cars.Visual multi-target tracking applications in complex environments not only include issues like light variation,target deformation and occlusion,but also include complex issues like unknown number of targets,uncertain newborn targets,image blurring and clutter interference.These issues have always been problems in visual multi-target tracking.This paper mainly focuses on the application of Multi-Bernoulli filter(MBF)in visual multi-target tracking,combined with target detection to carry out in-depth research,and achieved the following research results:1.To solve the problems of the MBF in the multi-target tracking filed,such as the newborn targets cannot be detected,the tracking accuracy decreases or the targets are underestimated when the targets are occluded,and the underestimated targets cannot be associated with their previous tracks when they are redetected.The YOLOv3 method is introduced to predetect the targets under the framework of MBF,and then the predetected targets are presented by the convolution features and their similarity matrix are calculated between the adjacent frames.The detection strategy for newborn targets and the re-identifiable strategy for underestimated targets are proposed to identify the targets,which can keep the targets associated with their previous tracks.Moreover,the adaptive template update method is proposed by merging the high confidence detection and the current estimated results,meanwhile the occlusion targets are considered.Finally,the proposed algorithm is tested on some challenging video sequences from the public standard dataset.The results show that the proposed algorithm has a good performance on detecting the newborn targets and the underestimated targets with high tracking accuracy.2.Aiming at the practical application of the target detector,due to the problems of target occlusion,complex background,and video blur,the lack of some targets' detection and inaccurate of detection cause the decrease in tracking accuracy and a large number of trajectory fragments and ID switches.In this paper,the visualized multi-Bernoulli tracking method is integrated under the target tracking framework associated with Intersection Over Union(IOU).When the target trajectory does not find a matching high confidence detection,the multi-Bernoulli filter is used to tracking the target to solve the problem of trajectory fragmentation.At the same time,according to the IOU threshold judgment and similarity comparison.Associate the target trajectory with the re-detected high-quality detection to prevent the ID switches.Experimental results show that the proposed algorithm can effectively reduce trajectory fragments and ID switches,and improve the accuracy of multi-target tracking.3.Aiming at visual multi-target tracking,the multi-Bernoulli filter uses sampled particles to approximate the posterior probability density distribution.If the number of particles is too small,it is difficult to include real target state;if the number of particles is too large,it will increase the calculation.In this paper,the correlation filter is introduced as a weak filter to sample particles under the framework of multi-Bernoulli filter.First,the VGG19 convolution feature with strong generalization ability is extracted to train multiple correlation filters.Secondly,multiple target states are obtained using the trained correlation filter.Finally,the obtained target state set is expanded as a sampled particle set.The proposed dual-filter target tracking algorithm adopts a weak filter to sample particles,which contains the real state of the target effectively,thereby improving the tracking accuracy of multi-Bernoulli filter.