Monocular visual obstacle avoidance method for autonomous vehicles based on YOLOv5 in multi lane scenes

Abstract

This study explores a more effective obstacle avoidance method for autonomous driving based on the monocular vision system of YOLOv5. The study utilizes the YOLOv5 model to detect obstacles and road signs in the environment in real-time, including vehicles, pedestrians, traffic signals, etc., identifying objects of different sizes and angles, providing accurate obstacle avoidance decision inputs. Then, we use a path planning algorithm based on deep reinforcement learning to combine the detected obstacle information with the current state of the vehicle, dynamically generating safe and efficient driving paths. In order to further improve the obstacle avoidance effect, a monocular visual autonomous driving obstacle avoidance aggregation network was introduced in the study, and the MMA (Multi lane monocular visual autonomous driving) obstacle avoidance method was established. The future movement trend of obstacles was predicted by setting an improved monocular obstacle avoidance loss function to sense historical data and the environment. The experimental results show that MMA has significantly enhanced obstacle avoidance and driving efficiency. The accuracy of the MMA obstacle avoidance method fluctuates between 78.76 % and 88.26 %, showing the best overall performance. The accuracy also shows an upward trend with the increase of the data sample size.