Research On Multiple Object Tracking Integrating Regional Regression,Motion And Appearance

Multi-object tracking(MOT)plays a crucial role in scene understanding tasks for video analysis.However,the MOT task remains a challenging problem due to poor detection quality,uncertain number of targets,background clutter,similar appearance,variation of occlusions and motion pattern.This dissertation focuses on developing effective methods to handle these problems.The contributions are shown as follows:(1)As for the poor quality of observations provided,a multi-object tracking framework combining with regional regression algorithm is proposed.It aims to improve the precision of multi-object tracking and reduce the number of false negative rate.This work begins with an analysis,which illustrates the great influence of video pedestrian detection's quality to the multi-object tracking.Without regard to the region proposal network in detection frameworks,observation boxes provided and the tracked boxes are jointly fed into the region regression network.Thus,the accuracy of multi-object tracking is greatly increased.Besides,considering of the tracking information,a novel non-maximum suppression(NMS)is proposed.Moreover,combined with the proposed tracker,an integrated detection framework is designed,which can improve the stability of the detector.(2)Considering the compatible problems of both motion and appearance models under the scenario with similar appearance,variation of occlusions and motion pattern,a general architecture named as MIF is presented by seamlessly blending the motion integration,three-dimensional(3D)integral image and adaptive appearance feature fusion.Due to the impact of camera motion to the distribution of pedestrian motion model,the intension of camera motion is designed to adaptively adjust the parameters of pedestrian motion model.Besides,it can also be used to recognize current frame's motion state and compute the intension of camera motion.Combined with the intention metric,the integrated motion model is designed.Meanwhile,a 3D integral image based spatial blocking method is presented to efficiently cut useless connections between trajectories and candidates with spatial constraints.Thus,the time complexity of cost computation at data association stage is reduced to linear complexity.In consideration of the rich historical feature information,and the feature misalignment problems between tracked boxes and observation boxes,an occlusion aware appearance model with visibility prediction branch is proposed.Furthermore,an adaptive appearance features fusion mechanism based on pose and occlusions is also designed to handle the misalignment problems.Finally,the MOT Challenge datasets are used to evaluate the performance of the algorithm proposed above.Experiments demonstrate the advantages of our proposed methods.Our proposed tracker achieves the state-of-the-art accuracy on all of the online MOT Challenge with 48.1,60.4 and 60.1MOTA(The Multi-Object Tracking Accuracy)on MOT15?17.Meanwhile,the unified detection framework based on our proposed tracker also achieves considerable performance on MOT17 Det with 4.8% MODA(The MultiObject Detection Accuracy)more than the original detector.Overall,experiments demonstrate the advantages of our proposed algorithms.