Robust visual tracking via identifying multi-scale patches

Abstract

The complex changes of target and its surroundings introduce several tracking challenges, such as occlusion, deformation and so on. Many challenges coexist in a video which makes tracking still under successfully solved. The present trackers deal with coexisting challenges in a common model for all components of target. However, different components often undergo different challenges at the same time, while some with deformation and others with occlusion. The common model cannot adapt to these challenges simultaneously. An effective method is to separately deal with the challenges. This paper proposes a new robust tracker via separately tracking and identifying the multi-scale patches of target to cope with the coexisting challenges. It is achieved by three respects. Firstly, we define a new basic tracker by introducing the gaussian mixture model into Kernelized Correlation Filters (KCF). For the KCF is very sensitive to the similar surroundings, we construct a regular term and a loss function via the gaussian mixture model to optimize the classifier formed by KCF. Secondly, we define a new appearance representation model of target by multi-scale patches. To deal with the different variations of patches, we separately construct and update their appearance representations. Thirdly, with the tracked result of each patch computed by our basic tracker, we use the structure information and the Hough Vote to decide the target. Then, our method improves the accuracy by rejecting the failed tracked patches. Many experiments have been achieved on the Tracking Benchmark, and the quantitative and qualitative evaluations show that the proposed tracker performs better than most of the present trackers.