Research on Multi-class Road Obstacle Recognition and Decision Based on YOLOP Combined YOLOV5 Algorithm
Ganrong Dong, Tiancong Han, Chenxiao Feng · 2023
The paper proposes a visual real-time sensing system. It carries YOLOv5 and YOLOP models and can be carried on the vehicle embedded device. The system simultaneously performs vehicle, pedestrian, traffic light, traffic sign detection, driveable area segmentation and lane detection. The paper not only considers the high accuracy, but also considers the computational cost of running on the vehicle when selecting the model to get the sensing information. The paper chooses YOLOP, which enables FPS to reach 23 (after TRT acceleration) on embedded devices. YOLOv5 also has extremely fast execution speed and accuracy on embedded devices. The combination of the two well meets the speed and accuracy required by the automatic driving decision control. And upload all sensing information to the embedded device. At the same time, a priority decision system for automatic driving is also proposed: pedestrian and vehicle distance information (divided into high, medium and low risk)$>$information of signal lights and traffic signs$>$distance information of lane line$>$information of driving area. Finally, this paper combines perception information with decision control and carries out visual output. Through experiments, it can be seen that the decision has achieved good results and the automatic driving function can be realized in most Chinese real street scenes.