Design of a Deep Learning-Based Intelligent Driving Decision System for Vehicle Engineering and Adaptability Verification in Complex Road Conditions

Engineering notes

Experimental results based on the nuScenes dataset and real-vehicle tests demonstrate that the proposed system achieves a decision response time below 80 ms and an average decision accuracy of 92.7% under complex road conditions, outperforming conventional rule-based and single deep learning approaches by 10.6%–28.4%.

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

To address the limitations of conventional intelligent driving decision systems, including insufficient real-time performance, low decision accuracy, and weak adaptability under complex road conditions, this study proposes a deep learning-based intelligent driving decision framework with enhanced multi-source electromagnetic sensing and environmental perception capabilities. Considering the increasing importance of millimeter-wave radar and heterogeneous sensor fusion in intelligent transportation systems, a multi-source data fusion module integrating cameras, LiDAR, millimeter-wave radar, and vehicle status information is first established, where adaptive Kalman filtering is employed to improve data reliability and suppress noise interference. Subsequently, a hybrid decision architecture combining an improved Deep Q-Network (DQN) and a Transformer encoder is developed. Dual Q-learning and prioritized experience replay optimize sequential driving decisions, while the Transformer captures global contextual features of complex road environments, enabling effective integration of local behavioral optimization and holistic situational awareness. Finally, a comprehensive validation framework covering extreme weather, special traffic scenarios, and sudden obstacles is constructed to evaluate system robustness. Experimental results based on the nuScenes dataset and real-vehicle tests demonstrate that the proposed system achieves a decision response time below 80 ms and an average decision accuracy of 92.7% under complex road conditions, outperforming conventional rule-based and single deep learning approaches by 10.6%–28.4%. The proposed framework provides reliable decision support for intelligent vehicles and offers valuable references for electromagnetic sensing, multi-source information fusion, and wireless perception systems in advanced autonomous driving applications.

Links and sources

Comments