Research On Robust Visual Object Tracking Algorithm Based On Correlation Filter

Visual object tracking is one of the most important components in computer vision field. To deal with complex tracking scenarios, most tracking algorithms choose tracking-by-detection methods. The trackers based on discriminant correlation filters(DCF) are prominent, but most just consider the holistic representation, ignore the target's scale changes, and only use a single tracker model to update the appearance of target constantly. Therefore, they are weak at dealing with occlusion and severe deformation.To solve the problem of KCF performance degradation when scale variations and occlusion happen, an algorithm of salient patch-based visual tracking with multi-cues integration(SPMCI) is proposed. By analyzing the local and holistic structures of the target apparent, SPMCI uses local patch as the basis model of target apparent. A lot of patches will be produced by tracking algorithm based on average patch technology, and not only introduce the unnecessary background, but also increase the computing complexity. So saliency map of the target is exploited to control the distribution and number of patch firstly. Then the scale pyramid of patch is constructed for scale estimation. To further improve the performance of the algorithm, the appearance cue, distribution cue and trajectory of patches are integrated for translation estimation. The experiments reveal that SPMCI algorithm has improved the precision of results and the ability to adapt different scenes of tracking.How to deal with the apparent model pollution caused by the long time occlusion and large deformation is a difficult problem. Based on SPMCI, a tracker using apparent change detection and multi-trackers relay(MTR) is proposed. Color histogram matching method is exploitd as a basis for judging apparent changes, and the factor-PSR(Peak to Sidelobe Ratio) is adopted to distinguish between deformation and full occlusion. According to the results of detection, a new tracker model is constantly initialized or the apparent model is updated. Finally, experimental results show that MTR improves the tracking performance in long time occlusion and large deformation scene.