Abstract: In the automatic driving scene, LiDAR camera is usually used to obtain the point cloud data with high accuracy and perceptible distance. Therefore, achieving object detection by effectively using point cloud data is the key technology to complete the automatic driving task. Point cloud has the problems of sparsity, disorder, and large amount. Traditional deep learning object detection method is difficult to effectively extract features map and meet the accuracy requirements. This paper proposed a 3D object detection method based on the fusion of voxel convolution network and multi-layer perception model. The voxel convolution network was used to extract the global features of point cloud, combined with the local features and distance relationship of point cloud extracted by multi-layer perception. It can improve the accuracy and speed of 3D object classification and position prediction. In this paper, the KITTI dataset was used to compare the proposed method with the classical method. According to the experimental results, the accuracy of the proposed method is significantly improved than the previous methods.