Monocular vision approach for Soft Actor-Critic based car-following strategy in adaptive cruise control

The ACC (adaptive cruise control) system is an important component of autonomous vehicle, and measuring inter-vehicle distance and velocity are basic environmental perception functions of it. Since camera-captured images provide richer environmental information at lower cost compared to sensors such as lidar, monocular vision-based solutions are gaining significant attention. However, traditional independent perception methods often overlook the importance of image context information in velocity estimation, and the detection errors of different monocular cameras at different distances may affect the accuracy of car-following. To address these issues, we propose amore effective monocular vision-based structure named TheNeoNet. TheNeoNet utilizes Graph Neural Network (GNN) to segment different pixel regions of the entire optical flow output feature map, thereby learning the relative relationships between different regions and endowing the motion feature extractor with stronger context awareness, significantly improving the accuracy of velocity estimation. After the vision- based perception, Soft Actor-Critic (SAC) method is employed to further optimize the car-following strategy. This method ensures that vehicles travel within a range of high monocular camera detection accuracy while balancing safety and comfort by minimizing predefined performance function indicators. The evaluation results show that compared to the state-of-the-art, the MSE of relative velocity estimation of TheNeoNet decreases by 3.83% and 83.50%, RMSE of distance estimation decreases by 14.49% and 21.39% on TuSimple dataset and KITTI dataset, respectively. The safety and comfort of proposed overall car-following strategy has increased compared with others, and it has good adaptability in various driving scenarios.