基于双目点云的三维物体检测准确性分析

双目相机虽然能通过算法生成密集的深度数据，但其在精度上与激光雷达生成的深度数据有着较大的差距，特别是在纹理不明显的区域。针对这种现象，尝试使用激光雷达的精确点云来排除由双目相机产生的pseudo-LiDAR数据中与之差别较大的部分点，然后将优化后的pseudo-LiDAR用以进行三维物体检测。实验结果显示，将pseudo-LiDAR数据中的不准确点（坏点）排除，有助于提高检测准确率，最多可提高21.02%。因此，如何不依赖激光雷达数据来排除pseudo-LiDAR点云中的坏点是进一步提高双目相机系统检测准确率的关键。

Accuracy analysis of 3D object detection based on stereo point cloud

3D object detection is a crucial task in autonomous driving. Recently, the accuracy of LiDAR-based 3D object detection algorithms have improved dramatically. However, if we replace the LiDAR data with the depth map that generates from stereo cameras, the detection performance drops a lot. In a recent research, it is found that by transforming the stereo-based depth map to point cloud representation, referred to as pseudo-LiDAR, the performance of 3D object detection can be improved significantly. However, there is still a big gap of accuracy between LiDAR-based and stereo-based algorithms. One of the main reasons is that there are error points in the transformed pseudo-LiDAR data. To study this, we use the LiDAR points cloud to correct the pseudo-LiDAR data, i.e., to detect and exclude the error points. Then we use the optimized pseudo-LiDAR to perform 3D object detection. Experimental results show that the detection accuracy can be improved by as much as 21.02%.