3 D target detection is the key technology of autopilot system and robot system.At present

most of the research on this technology is based on the RGB-D image obtained by camera.However

in practical application scenarios

the input radar point cloud is more suitable with the 3D object detection system application environment.RGB image and radar point cloud are used as input

and the IP-Basic algorithm is used to eliminate the noise and fill the sparse radar point cloud into a dense depth map.According to the principle of dimensionality reduction

the instance is segmented on the RGB image with Mask R-CNN

and the corresponding region of the target on the depth map is extracted according to the segmentation instance.Combined with the natural invariance of the point cloud

the depth map corresponding to the target is converted into the point cloud which will be inputted into the convolutional neural network for feature learning.Experimental result on public datasets Kitti show that the proposed method achieves good detection precision on the public 3D target detection datasets.