The work aims to propose an AGV obstacle avoidance strategy adopting vision sensor based on improved target detection algorithm aiming at the problems of low identification of environmental information and difficulty in accurate obstacle avoidance in current AGV obstacle avoidance methods. The traditional SSD target detection algorithm was improved by Mobilenet model, and the trained SSD-Mobilenet model was transferred and learned by the AGV working environment data. The obstacle avoidance principle was realized by combining vision, ultrasound and other modules, and the experimental platform with Raspberry Pie 3B+ as the control core was built for relevant experimental research. Experiments showed that the detection accuracy of this method was 94%, and it could accurately identify the types of obstacles. The obstacle avoidance time of the target detection method was 15.8% to 27.3% less than that of the traditional method. This method can effectively improve the accuracy and efficiency of AGV obstacle avoidance, and can be widely used in AGV obstacle avoidance control.