Research of Autonomous Robot Object Tracking System based on Novel YOLOX-SC Algorithm

In recent years, extensive research has been conducted on robot visual tracking problems in non-stationary scenarios. In order to improve the real-time tracking performance of the system, this paper proposes a robot visual tracking control system based on an improved YOLOX algorithm. The Swin Transformer self-attention and CBAM (Convolutional Block Attention Module) attention mechanism modules will be applied to the network architecture of the YOLOX algorithm to enhance the detection performance of object features. The experimental platform uses an industrial camera to obtain target pose data, and after deploying the model with lightweight optimization, the system detects and tracks moving objects in real time from video frames. Simultaneously, the gimbal is servo-controlled based on the image position deviation of the moving target. Experimental data shows that the quantized YOLOX-SC network model achieves a detection accuracy of 87.5% and a recognition speed of 120 FPS. This ensures real-time control of the system while improving the accuracy of object detection. Comparative results also indicate that deploying the improved detection algorithm on the experimental platform significantly enhances visual tracking performance compared to the original detection network.