Monocular Robot Tracking Scheme Based on Fully-Convolutional Siamese Networks

Tracking a specified target is a fundamental task of service robots. This paper studies and designs a monocular mobile robot tracking system based on Fully-Convolutional Siamese Networks. This tracking mobile robot system uses the Siamese convolution network as the tracker to lock the target. The bearing conversion algorithm and speed conversion algorithm are combined to implement following the target stably and accurately. The bearing conversion algorithm keeps the tracking target in the middle of the image and calculates the moving angle according to the distance from the center. The speed conversion algorithm sets different speeds according to the ratio of bounding box area. Implementation of the monocular mobile robot tracking system is grounded on the Robotic Operating System (ROS) platform. The experiment shows the effectiveness of the implemented the monocular robot tracking system which can achieve real-time following target at variable speeds according to the forward speed of the tracking target with safety distance even though having occlusion.