Research on Automatic Driving System Based on End-to-End and Real-Time Path Planning

Engineering notes

Key topics: autonomous driving, end-to-end, path planning, occupancy network, occupancy, perception, prediction, planning. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

With the development of autonomous driving technology, visual perception systems face challenges in target detection and localization tasks in complex traffic scenes, deep learning provides new ways to solve this problem, but single-phase and two-phase target detection algorithms need to balance speed and accuracy. In this paper, we study the related work of machine vision in the field of autonomous driving, based on which we propose path planning methods based on multiple neural networks, including CNN, RNN, Transformer and Occupancy Network, etc., and design an integrated framework to make the perception and path planning modules work closely together. The experimental results show that in the end-to-end model performance comparison, the CNN-LSTM-Transformer model converges fast and has the lowest loss and error in the validation set, and has the highest path prediction accuracy and robustness; the evaluation of the real-time path planning algorithm reveals that it has a high real-time performance with an average time delay of only 0.0003 seconds, an average mean square error of 0.0081, an average absolute error of 0.0723, and an average path length of 0.0081, and an average path length of 0.0723. The average mean square error is 0.0081, the average absolute error is 0.0723, and the average path length is 12.3465, so the path quality is good. Future research can further optimize the algorithm to improve the model adaptability and detection efficiency, in order to promote the development of automatic driving technology.

Links and sources

• Official / arXiv page

Comments