Application of Improved PilotNet Intelligent Traffic Algorithm Based on Baidu PaddlePaddle

In view of the complexity of current traffic scenes, the rapid development of intelligent transportation applications and the improvement of domestic platforms, this paper combines the current mainstream deep learning algorithms to recognize the road line with the PilotNet model improved by BatchNorm and Dropout technologies under the Baidu PaddlePaddle framework, and to detect the traffic signs with the YOLOv3 Tiny model. Due to the limited hardware system resources and the real-time nature of automatic driving, finally, the two types of models are deployed to the Edge Board computing box. In the application of intelligent transportation, single-thread technology cannot respond to the real-time nature of automatic driving, and thread blocking has a great impact. After comparison, dual-thread technology is adopted. Data processing is carried out by capturing the moving image of the car and the control value, and the experiment is completed. In the experiments, the smart car can recognize the traffic signs at a speed of around 0.01s per frame and finish model loading and making decision after the recognition of traffic scenes in 34.86s, which shows that the trained neural network models all have lightweight characteristics and can meet the system application with limited resources.