Research On Video Object Tracking Algorithm By Sparse Representation

Object tracking has been widely applied in many fields such as video surveillance,intelligent transportation, human-computer interaction, industrial control, medical science,military, and so on. It has been an active topic in computer vision field. In the trackingsystem, tracking algorithm is the core technology of the tracking system. And it affects therobustness, accuracy and timeliness of the tracking performance. Nowadays, with thedevelopment of compression sensing theory, sparse representation, as an important part ofthis area, receives widespread attention. L1-tracker in the frame of particle filter is adoptedby many state-of-art object tracking algorithms which can handle these challengingproblems, such as occlusion, deformation of the object and complex background, lightingchange, and so on. But its major defect is poor real-time performance. In this paper, in orderto improve the real-time performance of L1-tracker, the following two aspects are adopted.Firstly, one of major cause which affects the efficiency of L1-tracker is the giganticdimensions of the small positive and negative template, which cause the sparse calculationinto linear growth. In this paper, complete dictionary is reconstructed by the concept of theHaar-like feature and feature block. The single pixel is replaced by the pixel block to makeup positive and negative trivial templates. So the dimensions of over-complete dictionaryare reduced in sparse representation and the calculation of the sparse is significantlyreduced. On the premise of robustness, the number of the target template is reducedadoptively in order to decrease the calculation of sparse representation.Secondly, updating frequency of the templates is further controlled according to sparsecoefficient.In this paper the proposed algorithm are tested on public video clips and capturedvideos in the laboratory. The experimental results show that our algorithm has betterperforms than the latest state-of-art five tracking algorithms in real time application andhandling the problems such as short time occlusion, the pose of target and illuminationchanges of tracking.