Research On Multi-source Target Fusion Tracking Method Based On Sparse Representation

The target tracking is one of the main research directions in computer vision field,which has been widely used in video surveillance,military guidance,unmanned driving,human-computer interaction and so on,and has attracted much attention from domestic and foreign researchers.As an important branch of target tracking,multi-source target tracking is implemented by combining image data from multiple sensors.Because the data from different sensors are redundant and complementary,it is possible to achieve better tracking performance than a single sensor.The infrared-visible fusion tracking is one of the most popular technologies.How to efficiently and accurately track the target and analyze the changes of the motion state is a challenging problem in the current multi-source tracking.Aiming at the problem,this paper proposes two fusion tracking algorithms of infrared-visible target based on joint sparse representation:1.The fusion tracking algorithm of infrared-visible target based on sparse representation and L1-APG.Firstly,the target models of infrared and visible are built respectively by using sparse representation method,and the optimization problem is constructed by minimizing their joint reconstruction error.Secondly,the optimization problem is solved by employing L1-APG algorithm.Finally,the computational complexity of the algorithm is further reduced by using the minimum error boundary constraint to reduce resample of particle,thus gain real-time fusion tracking.2.The fusion tracking algorithm of infrared-visible target based on sparse representation and occlusion detection.The sparse representation of the target is used to describe the target appearance model,and a method of simultaneous tracking and recognizing is introduced on the basis of sparse representation.To solve the problem that the occluded tracking results are improperly added to the reference template set during the target template update,we establish the occlusion detection model that can calculate the size of the occluded area.The collaborative learning method is used to update the reference model and reduce the impact of occlusion on the tracking results.The test results of infrared and visible image sequences demonstrate that the two proposed methods in this paper can perform well in dealing with target intersection,target rotation,illumination change,and target occlusion.