Compressive tracking based on random channel haar-like feature

Abstract

Compressive tracking based on random channel Haar-like feature (RCCT) is proposed in this paper to improve tracking accuracy. Firstly, the color video frame is converted into grayscale image for tracking in real-time compressive tracking (CT), which may lose some information. Therefore, Haar-like features with random position and size are generated from three channels (RGB with random), represent the target better. What's more, it costs much time to detect new target round the position of the target in the current frame in the CT algorithm, and causes the target to drift when the speed of the target increases suddenly. Searching the new target in the vicinity of prediction target is proposed to reduce search time and to avoid missing the target. We have done experiments with large number of public data sets. Experimental results show that the RCCT algorithm reduces the average error of the target center compared with CT algorithm and other improved algorithms, and it performs favorably at the circumstances of the change of illumination and target position.