Multilevel Feature Aggregation Network with Efficient Voxel-to-keypoint Encoding for 3D Object Detection on Point Clouds

3D object detection from LiDAR point clouds has been one of the key components of computer vision systems and plays a significant role in autonomous driving applications. Existing voxel-based methods convert 3D convolutional features into bird’s-eye view feature representations without considering the dynamic characteristics of 3D proposals during the voxel convolution process, with undesired background points due to weak feature representation capability. To remedy this issue, this paper proposes a new voxel-based two-stage 3D object detection method, a multilevel feature aggregation network (MFANet). Specifically, it takes full advantage of the flexible receptive fields of ball queries and voxel queries. The multilevel features of the 3D proposal are extracted from the 3D sparse convolutional features, the bird’s-eye view features, and an efficient voxel-to-keypoint encoding module is developed by exploiting long-range relationships between points and learning discriminative features for 3D proposal refinement through progressive keypoint sampling and multiscale self-attentions. By aggregating multilevel features via convolutional attention, enriched feature expression of 3D proposals can be obtained. As a result, this method enables the 3D proposals to capture high-level semantic information and fine-grained details of points, which can more accurately estimate object confidence and location. Extensive experiments on the widely used KITTI dataset show that the proposed MFANet significantly improves detection accuracy compared to the state-of-the-art methods.