Abstract: Aiming at the inaccurate recognition of 3D detection due to external environmental factors such as illumination and sparse occlusion of laser point clouds, a 3D vehicle target detection method based on improved EPNet is proposed. A detection method based on the combination of point and graph methods is proposed in this paper. The graph method updates points by learning the edge information between points, and combines the two point cloud feature extraction methods to better learn the spatial information between points and reduce the loss of the original structural information of the object. And the way of image and point cloud fusion is improved by using the features of the upper-level fusion module, and a multi-layer feature fusion is proposed, which reduces the influence of illumination and improves the accuracy of vehicle detection. Experiments are carried out on the public KITTI dataset. The results show that the method proposed in this paper has higher detection accuracy than the original network, and the detection accuracy of each level of vehicle has been improved, especially in the Moderate level. The detection accuracy has increased by 1.83%, proving the superiority of the method.