The Pseudo-LiDAR algorithm is one of the general algorithms for 3D object detection with stereo cameras. Because of its low-cost stereo vision scheme and excellent adaptability, the Pseudo-LiDAR algorithm is widely used in the fields of autonomous driving, robots and so on. However, the algorithm will produce too many distortion points in the prediction of the edge of the object, resulting in a large estimation error of the transition region between the object and the background. The depth estimation network adopts a four-branch structure, and the large amount of calculation affects the running speed. To solve the above problems, the Spatial Pyramid Pooling (SPP) four-branch structure within the Pseudo-LiDAR algorithm has been optimized to a single-branch structure to improve the running speed. Innovatively construct a Gaussian confidence filtering module to filter out data with large errors in the transition region; PointVoxel-RCNN (PV-RCNN) is introduced as a 3D object detection network in the Pseudo-LiDAR algorithm to improve the detection accuracy. The experimental results on the KITTI dataset show that compared with the original algorithm, the average accuracy of 3D detection of car objects under simple, medium and difficult levels is improved by 22.36%, 26.48% and 28.55%, respectively