联合多视角互投影融合的三维目标检测方法

针对当前智能车辆目标检测时缺乏多传感器目标区域特征融合问题，提出了一种基于多模态信息融合的三维目标检测方法. 利用图像视图、激光雷达点云鸟瞰图作为输入，通过改进AVOD深度学习网络算法，对目标检测进行优化；加入多视角联合损失函数，防止网络图像分支退化. 提出图像与激光雷达点云双视角互投影融合方法，强化数据空间关联，进行特征融合. 实验结果表明，改进后的AVOD-MPF网络在保留AVOD网络对车辆目标检测优势的同时，提高了对小尺度目标的检测精度，实现了特征级和决策级融合的三维目标检测. 

3D Target Detection Method Combined with Multi-View Mutual Projection Fusion

Aiming at the lack of feature fusion of multi-sensor target regions in the current target detection of intelligent vehicles, a three-dimensional target detection method was proposed based on multi-modal information fusion. Firstly, taking the image view and aerial view of lidar point cloud as input, the target detection was optimized by an improved AVOD deep learning network algorithm. And then, a multi-angle joint loss function was inducted to prevent the branch network image degradation. Finally, a dual-view image and the lidar point cloud projected mutual fusion method was presented to enhance data spatial correlation and to carry out feature fusion. The experimental results show that the improved AVOD-MPF network can improve the detection accuracy of small-scale targets while retaining the advantages of the AVOD network for vehicle target detection, and achieve 3D target detection with feature-level and decision-level fusion.