单目3D目标检测中注意图特征提取网络

Attention Map Feature Extraction Network in Monocular 3D Object Detection

针对单目3D目标检测中,目标深度估计存在深度特征提取不准确的问题,提出一种获取历史全局信息的注意图特征提取网络(AFENet)。通过历史语义卷积获取历史特征中的全局信息,增强上下文特征表达,捕获长时记忆关系,提取更准确的深度特征。实验结果表明,AFENet在KITTI数据集上的检测结果与单目3D目标检测网络相比,在图像检测难度中等和困难情况下平均检测精度提高0.8%和0.4%,可为无人驾驶及智能机器人等应用领域提供技术参考。

To solve the problem of inaccurate depth feature estimation in monocular 3D object detection,an attention map feature extraction network(AFENet)was proposed to obtain historical global information.In this algorithm,the global information of historical features is acquired by historical semantic convolution,and the context information of historical global information is enhanced,the long-term memory relationship is captured,and more accurate depth features are extracted.Experimental results show that compared with monocular 3D object detection network(M3DSSD),the average detection accuracy(3D mAP)of the proposed algorithm on KITTI datasets is 0.8%and 0.4%higher than those of the monocular 3D object detection network(M3DSSD)in medium and difficult image detection cases.