Part-Based Tracking with Appearance Learning and Structural Constrains

Abstract

Adaptive tracking-by-detection methods are widely used in computer vision for tracking objects. Despite these methods achieve promising results, deformable targets and partial occlusions continue to represent key problem in visual tracking. In this paper, we propose a part-based visual tracking method. First, we take advantage of the existing online learning appearance model to learning the appearance of each part. Second, we propose a novel part initialization method and an affine invariant structural constrain between these parts. Third, a tracking model based on the appearance of each part and the spatial relationship between the parts is proposed. We make use of an optimization algorithm to find the best parts during tracking, update the appearance model and the structural constraints between parts simultaneously. In this paper we show our method has many advantages over the pure appearance learning based tracking model. Our method can effective solve the partial occlusion problem, and relieve the drift problems. What’s more, our method achieves great result while tracking the target of which geometric appearance changes drastically over time.