A Robust Object Tracking and Visual Servo Method for Mobile Robot

The general Siamese network based object tracking methods tend to generate the final score map from high-level features and treat features from each position equally, which may lead to the problems of large search region and low efficiency. In order to solve these, this paper proposes a fully-connected Siamese network tracking method based on the calculation of histogram of gradient feature similarity and on feedback of the fading-memory Kalman filter. This strategy enables real-time correction and compensation, which means it could re-track the target although it is occluded or temporarily lost. The target’s bounding box obtained by object tracking method is used to produce the control command and achieve the image-based visual servo. Comparative experiments with other methods are conducted on several public datasets to prove its effectiveness. In addition, we design a mobile robot tracking system to test the algorithmic performance in real-world scenarios. Experimental results show that the robot is able to track the target accurately, and continue to track the target despite occlusion or temporary disappearance.