Visual Object Tracking Method for Mobile Robots Based on DSST

Abstract

Aiming at achieving fast and accurate objecting tracking performance for mobile robots, discriminative space scale tracking (DSST) correlation filter combined with Kalman filter is proposed. A confidence criterion according to the oscillation severity and average peak-to correlation energy (APCE) of the response graph is set to make a judgement whether the target is occluded or has pixel drift. Kalman filter will start tracking when the target is judged being occluded. The position state obtained from Kalman filter will be used to update the tracking model for the next frame. Experiments are taken on the OTB100 datasets and shows an improvement on the accuracy and success rate compared with other algorithms. Finally, the method is tested on our robot system and experiments show that our method satisfies the real-time requirements.