面向夜间场景的激光雷达与相机融合的三维目标检测算法

A 3D Object Aetection Algorithm Based on LiDAR and Camera Fusion for Night Scenes

摘要: 基于激光雷达与相机融合的目标检测算法成功提高了自动驾驶系统的感知性能。然而，以往的多模态算法都是面向白天场景进行检测，在夜间场景下会因为夜间环境亮度较低以及由灯光照射引起的过曝问题导致检测性能的下降；为此，本文研究了一种改进的夜间多模态检测算法。首先引入了基于统计的过曝区域像素掩码模块，通过分析像素饱和度特征来抑制过曝区域的干扰；其次引入了基于权重归一化的通道加权融合模块，采用逐通道动态权重分配机制来优化不同模态特征的融合。实验采用NuScenes数据集全集及其夜间子集验证模型性能，实验结果表明改进模型平均类别精度达74.5%，对比基线方法实现了性能上的提升，算法可靠性得到验证，该研究成果展现出较好的工程价值。

Abstract: The target detection algorithm based on the fusion of LiDAR and camera successfully improved the perception performance of the autonomous driving system. However, previous multimodal algorithms are all for detection in daytime scenes. In nighttime scenes, the detection performance will decrease due to the low brightness of the nighttime environment and the overexposure problem caused by lighting. Therefore, this paper studies an improved nighttime multimodal detection algorithm. First, a statistical overexposed area pixel mask module is introduced to suppress the interference of the overexposed area by analyzing the pixel saturation characteristics; secondly, a channel weighted fusion module based on weight normalization is introduced, and a channel-by-channel dynamic weight allocation mechanism is used to optimize the fusion of different modal features. The experiment uses the full set of NuScenes dataset and its nighttime subset to verify the model performance. The experimental results show that the average category accuracy of the improved model reaches 74.5%, which is a performance improvement compared with the baseline method. The reliability of the algorithm is verified, and the research results show good engineering value.