In the intersection scenario, the running state of social vehicles, the control state of traffic lights and the accurate identification of track components have become the technical bottlenecks restricting the promotion and application of track inspection robots. Aiming at the requirements of track health inspection, firstly, the vision inspection system of the track inspection robot and the technical scheme of the navigation system was presented based on "Beidou + 5G". Secondly, the vision detection system model was built based on YOLOv8 algorithm, and the web crawler technology was innovatively used to capture sample data about traffic lights and car taillights from open source video resources to train the vision detection model. Then, transfer learning method and early stop method were used to optimize the detection accuracy of the trained model. The research results show that after adopting YOLOv8 algorithm and optimizing the model with transfer learning method and early stop method, the inspection robot can effectively detect the track components, vehicles and traffic lights at the switch junction, and effectively improve the inspection efficiency and accuracy.