RTL3D: real‐time LIDAR‐based 3D object detection with sparse CNN

Abstract

LIDAR (light detection and ranging) based real‐time 3D perception is crucial for applications such as autonomous driving. However, most of the convolutional neural network (CNN) based methods are time‐consuming and computation‐intensive. These drawbacks are mainly attributed to the highly variable density of LIDAR point cloud and the complexity of their pipelines. To find a balance between speed and accuracy for 3D object detection from LIDAR, authors propose RTL3D, a computationally efficient Real‐time LIDAR‐based 3D detector. In RTL3D, an effective voxel‐wise feature representation is utilised to organise unstructured point cloud. By employing a sparse feature learning network (SFLN) on voxelised 3D data, RTL3D exploits the sparsity of point cloud and down‐samples 3D data into 2D. Basing on the generated 2D feature map, an optimised dense detection network (DDN) is applied to regress the oriented bounding box without relying on any predefined anchor boxes. The authors also introduce an incremental data augmentation approach which greatly improves the performance of RTL3D. Empirical experiments on public KITTI benchmark demonstrate that RTL3D achieves a competitive performance with state‐of‐the‐art works on 3D detection task. Owning to the simplicity of its single‐stage and anchor‐free design, RTL3D has a real‐time inference speed of 40 FPS.