Dynamic target recognition and obstacle avoidance in robot vision system based on YOLOv4 algorithm

This paper aims to solve the problem of reducing false detection and missed detection of dynamic targets under complex backgrounds or forms and the difficulty of robots dodging obstacles in complex. Using FPN, PANet, CSPDarknet53, and CSP modules, the method improves YOLOv4-based robot vision to achieve robust dynamic target recognition in complex, occluded scenes, improve multi-scale feature fusion, and reduce computation. By combining with the GAN (Generative Adversarial Network) adversarial network, dynamic targets can be detected more accurately and effectively. The results show that the detection accuracy (mAP) is 90%, the LOSS value of the model tends to be 0, and the stability of the model is strong. The recognition prediction results for each category are as high as more than 90%, which can effectively solve false detection and missed detection. For the research on obstacle avoidance, this paper uses DWA (Dynamic Window Approach) technology for fusion and conducts multiple experiments under different conditions. The obstacle success rate reaches 95% and 90% respectively in static and dynamic environments. By improving the YOLOv4 model algorithm, it can handle dynamic target recognition and obstacle avoidance problems well, providing strong evidence for identification plates