When driving at night, poor lighting conditions usually make it difficult for the object detection algorithms to accurately obtain the information on preceding vehicles. Moreover, in the object ranging algorithms, the disparity loss caused by the interference of nighttime environment may render the distance estimation ineffective. To address these issues, the improved algorithms for object detection and vehicle ranging in nighttime driving are proposed. Firstly, the C3k2CASA module is introduced into the YOLOv11n baseline to enable the network to focus on more valuable image features, and the CIFM module is inserted to the Neck to enhance both global and local feature representation. Secondly, the v10Detect module is introduced to suppress redundant predictions in the post-processing stage, and the loss function of model is optimized with PIoU2 to guide the bounding box prediction to regress along a more efficient path. Thirdly, the enhanced SGBM algorithm is developed for stereo matching, and constraints are applied to the matching points to obtain valid disparity stably, and then the distance information to preceding vehicle can be calculated based on the principle of binocular vision ranging. The analysis results show that, in the object detection analysis, the improved algorithm has good performance in the low-light scenarios compared to the original algorithm, and it can effectively balance the model lightweight and the high prediction accuracy in nighttime driving scenarios. In the vehicle ranging experiments, the improved algorithm not only performs identically to the original algorithm in single-vehicle ranging, but also demonstrates significantly enhanced robustness in multi-vehicle scenarios, and it can effectively prevent the problem where the original algorithm may fail due to being affected by the interference from multiple objects during nighttime driving.
Huang et al. (Thu,) studied this question.