Detection Based Data Association Method For Multi-Target Tracking In Complex Scene

With the advancement of social civilization and science technology,the demand of people for intelligent video equipment is growing.Thus,computer vision tech-nology attracts more and more attention.As an important component of computer vision technology,multi-target tracking has become a popular research issue in the field of artificial intelligence.At present,although many works have made several technical breakthroughs,the multi-target tracking in complex scene still faces many challenges:(1)Due to the influence of background noise,the reliability of moving target detection in video sequences is low.When these target detections are used as candidate regions of tracked targets,it is prone to generate target lost and drift.(2)The field of view may appear various complex scenes,such as frequent occlusions,similar targets and target deformations.These complex scenes increase the proba-bility of generating the tracking errors,e.g.,identity switching.(3)Because of the special requirement that the targets' number is variable,multi-target tracker needs to automatically initialize the new target according to the detection results.However,detection errors seriously affect the target initialization accuracy.To solve the above three challenges,the paper studies the detection based data association method for multi-target tracking in complex scene.The main contents and innovations are sum-marized as follows.1.To address the problem that moving target detection of traditional optical flow method is susceptible to noise interference,a moving target detection frame-work based on adaptive weight coefficient is proposed to avoid the optical flow field error caused by manually setting the weight coefficient.We propose an adaptive weight coefficient optimization strategy based on two-layer fuzzy C-Means(FCM)clustering algorithm.The adaptive optimization strategy overcomes the shortcom-ings that Horn-Schunck optical flow method cannot self-learn the convergence thresh-old,and reduces the interference degree of background noise to the target fore-ground.We design a moving target detection algorithm in video sequences based on the adaptive weight coefficient.Compared with other manually setting weight coefficients,the moving target detection result under the adaptive weight coefficient can achieve optimal or sub-optimal performance.2.Based on the target detection,the paper studies the detection based data as-sociation method for multi-target tracking.To solve the problem that the appearance of target detection region in complex scene is unreliable,we propose a multi-target tracking method using spatial-temporal feature fusion with acceleration feature.The proposed method can improve the multi-target tracking accuracy in complex scenes such as background changes and similar targets.A novel feature in the temporal di-mension is presented and called the acceleration feature descriptor,which provides additional motion cues for constructing a strong discriminative target feature model.Acceleration features are combined with multiple features in the spatial dimension to establish a target spatiotemporal feature model.The tracker employs the Hungar-ian algorithm to solve the cost matrix of data association,and completes the task of multi-target tracking.The proposed tracking method reduces the false alarm rate of multi-target tracking caused by the unreliable appearance of target,thereby multi-target tracking accuracy is improved.3.As for the situation that the targets are occluded each other in the multi-target tracking process,an anti-occlusion multi-target tracking method based on progres-sive spatial-temporal feature model is proposed to solve the problem of trajectory fragment and identity switch caused by the occlusion.According to the association states,overlap and depth ordering of the target,the proposed method establishes an occlusion reasoning model and identifies the occluded target in online.Progressive spatial-temporal feature model is proposed to re-detect the occluded target.By em-ploying the histogram of optical flow feature when spatial features are unreliable,the progressive spatial-temporal feature model reduces the influence of the occluded tar-get's appearance change.Therefore,this method can improve multi-target tracking accuracy,precision and trajectory completeness.4.To solve the problem that the target appearance features generate partial or significant variations frequently,a hierarchical data association multi-target tracking method based on the main-parts and spatial-temporal feature model is proposed.We construct main-parts feature model to represent the target with local or no appearance variations.By calculating the appearance feature affinity of the target in main-parts region,the tracker improves the accuracy of the cost matrix in the data association process.The discriminating ability of target feature model with partial occlusion and deformation becomes stronger.Spatial-temporal feature model is built by fus-ing the appearance and the optical flow features for the target's overall rectangular image.The model aims to distinguish different targets with significant appearance changes and to avoid tracking errors when the target is globally occluded.Tracklet confidence is exploited to implement hierarchical data association,which is accu-rately completed by using main-parts and spatial-temporal feature models for the targets with local and significant appearance variations.This strategy reduces the probability of tracking errors such as identity switch.5.To solve the problem that new targets initialization in complex scene is often inaccurate,we propose a deep spatial-temporal feature fusion multi-target tracking method with target initialization filter.The filter reduces the false positive new target caused by redundant detection results,while deep spatial-temporal feature model improves the discriminative ability of the new target feature model.We present a target initialization filter based on Gaussian mapping.It filters out the redundant new targets according to the overlap ratio of Gaussian map between targets to improve the accuracy of identifying new targets.By fusing deep convolution,size,position and optical flow histogram features,a deep spatial-temporal feature model is established for the new target,and is inherited or updated frame-by-frame.Our proposed method effectively improves multi-target tracking accuracy and robustness.