Compressive Tracking based on Superpixel Segmentation

Abstract

The compressive sensing trackers, which utilize a very sparse measurement matrix to capture the targets' appearance model, perform well when the tracked targets are well defined. However, such trackers often run into drifting problems due to the fact that the tracking result is a bounding box which also includes background information, especially in the case of occlusion and low contrast situations. In this paper, we propose an online compressive tracking algorithm based on superpixel segmentation (SPCT). The proposed algorithm employs a weighted multi-scale random measurement matrix along with an efficient superpixel segmentation to preserve the image structure of the targets during tracking. The superpixel segmentation is used to distinguish the target from its surrounding background, to obtain the weighted features within the bounding box. Furthermore, a feedback strategy is also proposed to update the classifier model to reduce the drifting risk. Extensive experimental results have demonstrated that our proposed algorithm outperforms several state-of-the-art tracking algorithms as well as the compressive trackers.