针对自动驾驶环境下的车辆检测问题存在小目标多、目标遮挡严重等情况，提出一种面向自动驾驶的多模态信息融合的目标识别方法。该方法主要包括以下几个方面的改进：1.基于空间注意力机制和混合空洞卷积改进的ResNet50网络：使用选择核卷积替换conv2_x和conv3_x部分的3×3标准卷积，使网络可以根据特征尺寸动态调整感受野的大小。在conv4_x部分使用锯齿状混合空洞卷积[1,2,1,2,1,2]，使网络能够捕获多尺度上下文信息，提高网络特征提取能力。2.改用GIoU损失函数：将YOLOv3中的定位损失函数进行替换，改用GIoU损失函数，GIoU在实际应用时有较好的可操作性。3.基于两种数据融合的人车目标分类识别算法：提出了基于两种数据融合的人车目标分类识别算法，可以有效地提高目标检测的准确率。实验结果表明，该方法与OFTNet 、VoxelNet 和FasterRCNN网络相比，在mAP指标白天提升幅度最高可达0.05，晚上可达0.09，收敛效果也更好。

The vehicle detection problem in the autonomous driving environment has many small targets and serious target occlusion, etc. In this paper, a multimodal information fusion dynamic target recognition method for autonomous driving is proposed.. The method mainly includes the following improvements: 1. Improved ResNet50 network based on spatial attention mechanism and hybrid null convolution: the 3×3 standard convolution in the conv2_x and conv3_x parts is replaced using selective kernel convolution, which allows the network to dynamically adjust the size of the perceptual field according to the feature size. The sawtooth hybrid null convolution [1,2,1,2,1,2] is used in the conv4_x part to enable the network to capture multi-scale contextual information and improve the network feature extraction capability.2. Switch to GIoU loss function: the localization loss function in YOLOv3 is replaced with the GIoU loss function, which has better operability in practical applications.3. Based on Two data fusion algorithm for human-vehicle target classification and recognition: A human-vehicle target classification and recognition algorithm based on two kinds of data fusion is proposed, which can effectively improve the accuracy of target detection. Experimental results show that compared with OFTNet, VoxelNet and FASTERRCNN, the mAP index can be improved by 0.05 in the daytime and 0.09 in the evening, and the convergence effect is better.

国家自然科学基金项目（面上项目）