Object detection for autonomous driving environmental perception and its applications

Abstract

This study tackles issues in autonomous driving by addressing problems with existing detection methods. Traditional single-modal approaches struggle with adaptability, while multimodal shallow fusion algorithms have weak feature coupling. Decision outputs also lack robustness. This study propose a single-stage object detection model that uses an enhanced YOLOv12 algorithm and point cloud collaboration. This study first optimizes the lightweight feature extraction in the YOLOv12 backbone network and preprocesses the point clouds. Then, it establishes a sensor decision-level adaptive fusion multidimensional perception algorithm. A dynamic scene confidence evaluation factor and a Bayesian probability optimization model are introduced into the framework. The results show that the improved YOLOv12 embedded with a Coordinate Attention mechanism increases the proportion of effective features by more than 20% on average in various complex traffic scenarios. The improvement is most significant in low-light and small-object scenarios. Meanwhile, the proportion of background noise decreases considerably, and feature response intensity increases significantly. The proposed dynamic decision-level fusion algorithm improves both classification accuracy and localization accuracy. The perception accuracy increases by 5.83% compared to the fixed-weight fusion method. This study provides a technical basis and engineering guidance for the large-scale deployment of the next-generation autonomous driving system and the optimization of the vehicle’s environmental perception framework.