3D Object Detection Based on Spatially Adaptive Sparse Convolution

Autonomous driving has gradually gained attention in recent years due to its potential to reduce driver burden and improve safety, in which 3D object detection is the basis of the autonomous driving perception system, aiming at predicting the location, size, and class of 3D objects in the vicinity of an autonomous vehicle, which is an important guide for subsequent path planning, motion prediction, and so on. In particular, 3D object detection has difficulties in input point cloud data processing and multimodal fusion. Unlike the image distribution where pixels are regularly distributed in the image plane, the point cloud is a sparse and irregular 3D representation, so it is difficult to apply the traditional convolution directly on the point cloud for detection. In this paper, we carry out a method for 3D object detection for point clouds, using deformable convolution and sparse convolution as a tool to explore the feature extraction method of sparse convolution for irregular point clouds and propose a spatially adaptive sparse convolution-based method.