面向复杂光照环境的改进CLOCs跨模态目标融合算法

Improved CLOCs Cross-Modal Target Fusion Algorithm for Complex Lighting Environments

摘要: 在自动驾驶领域，复杂光照条件(如弱光、过曝、眩光及空间亮度不均等)会导致视觉特征退化，从而显著降低目标检测的准确性与可靠性，尤其对小尺度及部分遮挡目标影响更为突出。针对上述问题，本文提出了一种基于改进CLOCs的视觉-LiDAR多传感器融合目标检测方法。在统一融合框架下，联合利用二维图像的语义信息与三维点云的结构信息，通过对CLOCs算法进行改进，实现跨模态特征的有效互补，从而提升复杂场景下的检测性能。同时，针对光照变化带来的特征不稳定问题，引入光照鲁棒性优化策略，以增强模型在低光、炫光及过曝等环境中的适应能力。实验结果表明，所提方法在mAP@0.5:0.95和mAP@0.5指标上分别达到55.6%和75.3%，相较于多种主流方法均取得了一定提升。进一步分析表明，该方法在不同尺度目标上均具有良好的检测性能，尤其在中小目标检测任务中表现更为优越。本文方法在检测精度、环境适应性及鲁棒性方面均取得了较好的效果，具有一定的实际应用价值。

Abstract: In the field of autonomous driving, challenging lighting conditions (such as low light, overexposure, glare, and spatially non-uniform brightness) can lead to degradation of visual features, thereby significantly reducing the accuracy and reliability of object detection, especially for small-scale and partially occluded targets. To address these issues, this paper proposes a vision-LiDAR multi-sensor fusion object detection method based on an improved CLOCs framework. Within a unified fusion framework, the semantic information from 2D images and the structural information from 3D point clouds are jointly utilized. By improving the CLOCs algorithm, effective cross-modal feature complementarity is achieved, thereby enhancing detection performance in complex scenarios. In addition, to mitigate feature instability caused by illumination variations, an illumination-robust optimization strategy is introduced to improve the model’s adaptability under low-light, glare, and overexposed conditions. Experimental results demonstrate that the proposed method achieves 55.6% and 75.3% on the mAP@0.5:0.95 and mAP@0.5 metrics, respectively, outperforming several mainstream methods to varying degrees. Further analysis shows that the method maintains strong detection performance across objects of different scales, with particularly notable advantages in detecting small and medium-sized targets. The proposed method exhibits strong performance in terms of detection accuracy, environmental adaptability, and robustness, indicating its potential for practical applications.