Research On Target Tracking Using Optimal Matching Under Complex Scenarios

Visual target tracking technology has always been one of the hot topics in the field of image and visual information computing,and it has essential theoretical research significance and practical application value in both military and civilian domains.Due to the rapid development of science and technology,and the continuous improvement of the performance of video capture equipment,target tracking technology has made considerable progress.However,it is still a challenging problem to maintain a high tracking accuracy in complex scenarios.Complex scenarios mainly include:deformation,occlusion,fast motion,illumination variation,similar object interference,background clutters,etc.Aiming at these issues,this thesis focuses on the research of target tracking using optimal matching under complex scenarios,the methods of anti-interference matching tracking,sparse optimization tracking and deep learning model tracking are regarded as research objects,which are validated by unmanned aerial vehicle(UAV)cooperative tracking simulation platform.The specific contributions are as follows:(1)To overcome the effects of similar object interference,illumination variation,background clutters,a novel target tracking method of anti-interference matching under foreground constraint is proposed.Firstly,the method combines super-pixel segmentation and mean shift clustering to establish robust discriminant appearance model.To avoid the interference of similar objects in the foreground region,a decision-making algorithm for anti-interference matching is utilized to improve the matching accuracy.Meanwhile,to provide a more accurate target representation,an online model updating method is constructed,which can append appropriate feature compensations to the feature sets when a severe occlusion occurs.(2)To avoid the influence of target deformation and local occlusion,a novel tracking method based on sparse optimization of local sensing is proposed.In the framework of particle filter,to make full use of the local feature information of the target,the selected target region is segmented uniformly and non-overlapping,and added to the set of target template.The sparse matching is verified by calculating the similarity between local feature blocks and target templates in the candidate samples amid the sparse tracking process.In the light of the degree of occlusion,a occlusion decision-making method is used to detect occlusion,which ensures the set of target templates more complete.(3)To solve the effects of fast motion,rotation,background clutter on tracking,this research presents a target tracking method based on location-classification-matching model.The method combines the color feature of the target with the feature representation of convolution neural network.In the location model,the double-layer convolution network is used to calculate the target feature score in the search regions,and the feature point of maximum score is taken as the center to sample,then the candidate target regions of the current frame are obtained.In the classification model,three-layer convolution network is used to inter-class screen for candidate regions,and the classification score of each region is calculated to gain sub-optimal tracking results.In the matching model,conventional color features are used to perform intra-class optimization matching for sub-optimal target regions to determine the final tracking target.Moreover,this thesis updates the network in the location and the classification,respectively,and updates the matching model online and in real time.(4)To deal with the issue of low accuracy of UAV tracking under complex scenarios,this thesis establish a cooperative tracking simulation platform for multiple UAVs.Aiming at the interference of complex scenarios in the tracking process,the tracking methods proposed in this paper are embedded in UAVs,and a cooperative strategy is formulated according to the target recognition rate of each UAV to ensure the tracking accuracy.An optimization model is established to guarantee that UAVs are in the best observation positions,which takes the potential energy of UAV motion as the objective function and the cooperative strategy as the constraint condition.The position of each UAV is dynamically adjusted by solving the optimization model to ensure the accuracy of cooperative tracking.