FuseNet: 3D Object Detection Network with Fused Information for Lidar Point Clouds

3D object detection from lidar point cloud has an important role in the environment sensing system of autonomous driving vehicles. In this paper, we propose two modules for object detection works by more detailed voxel initial information extraction and full fusion of context information. Additionally, we extract density information as the initial feature of the voxels and fully confuse the coordinate and density information with a point-based method to reduce the loss of original data caused by voxelization. Second, we extract the voxel features with a backbone neural network based on 3D sparse convolution. We propose a Cross-connected Region Proposal Network to integrate multiscale and multidepth regional features and to obtain high-quality 3D proposal regions. In addition, we extend the target generation strategy in the anchor-based 3D object detection algorithm, which stabilizes the network performance for multiple objections. Our modules can be flexibly applied to state-of-the-art models and effectively improves the network performance, which proves the effectiveness of the modules that we proposed.

Automated summary

This paper proposes two modules to improve the performance of 3D object detection. The first module reduces data loss by extracting more detailed initial voxel information and fully fusing context information. The second module extracts voxel features using a backbone neural network based on 3D sparse convolution and generates high-quality 3D proposal regions by a cross-connected region proposal network. Additionally, this paper extends the target generation strategy in the anchor-based algorithm, stabilizing the network performance for multiple objects.