Ground Truth

The ground-truth data for depth and vehicle motion status in all sequences are obtained through LiDAR and GNSS/INS units, respectively. Specifically, the motion status data for the DCV are collected using the UG005 X1 unit, while the CGI-610 unit is used to collect motion status data for the GVT only in the CCRm scenario. In other scenarios, since the collision target does not move along the camera’s optical axis, the position of the collision target can be determined by measuring the distance between the DCV and the target using LiDAR. Please refer to the paper for details.

We use the LiDAR odometry FASTLIO2 to obtain ground-truth poses and velocitycompensated point clouds. For each pose, we extract the relative poses of several preceding and succeeding point cloud frames to generate local point clouds. Next, the Hidden Point Removal (HPR) operator is applied to eliminate occlusion effects in the local point clouds. Finally, the local point clouds are projected into the RGB camera coordinate frames of the 8-mm and 16-mm lens camera pairs to obtain ground-truth depth.