Improved object tracking algorithm based on tracking-leaning-detection framework

Abstract

In object tracking, a novel tracking framework which is called “Tracking-Leaning-Detection” was proposed by Zdenka Kalal. This framework decomposes the object tracking task into tracking, learning and detection. In every frame that follows, the tracker and the detector work simultaneously to obtain the location of the object independently, and the learning acts as an information exchanging center between tracker and detector. To make up defects of the framework's robustness, we reconstruct the detector with local binary pattern feature. Firstly, local binary pattern descriptor of every scanning-window is calculated to generate local binary pattern feature vector. Secondly, the new Local Binary Pattern feature vector is generated by histogram statistics of the local binary pattern feature vector, and the positive and negative samples (image patches) are transformed in the same way. Thirdly, the new local binary pattern statistics feature vector of the scanning-window is matched with the positive and negative samples set based on normalized cross correlation. Finally, the detection results and the tracking results are fused and the detector is updated online. The experimental results on the public data set show that the proposed algorithm has better tracking performance and robustness.