Abstract: Object detection in LiDAR point clouds is crucial for the field of robotics, especially in autonomous driving. With the rise of deep learning and the rapid advancement of image-based object detection algorithms, many recent works have started to transfer best practices from the image-based object detection field. However, due to computational constraints, network architecture designs are still relatively coarse and detection performance is in need of improvement. This paper proposes an optimized network for one-stage anchor-free LiDAR object detection, including voxelization, backbone network design, and feature pyramid introduction. Through these optimizations, we obtain a network structure that performs excellently in detection on cross platform, and can run in real-time on autonomous driving vehicles. Experiments validate that our network architecture outperforms other state-of-the-art methods on the publicly available KITTI dataset.