The 3D object detection algorithm is constrained by the large amount of point cloud data, and can not achieve the balance of real-time speed and accuracy. This paper presents an improved 3D object detection algorithm—Pillar RCNN. The algorithm firstly divides the target point cloud space into voxels, presents a 3D object detection backbone network based on sparse convolution which gradually converts voxels into column voxels, quantifies 3D information into dense 2D information, and then processes the dense 2D information through the 2D backbone network. At the same time, the voxel features of different scales in the 3D backbone network and the 2D backbone network detection results are cascaded through the multiscale voxel feature aggregation module, and the result is further refined by the loss function. The algorithm is tested on KITTI public datasets and has a recognition speed of 2.48 ms on RTX 2080Ti hardware platform. Compared with the PointPillars benchmark algorithm, the performance indicators of the three categories of car, pedestrian and bicycle are improved. The detection results of the mode of bicycle and the hard of car are improved by 13.34% and 8.85%, and the detectation accuracies of other categories are improved also, achieving a balance between speed and accuracy.