Deep 3D Object Detection Networks Using LiDAR Data: A Review

As the foundation of intelligent systems, machine vision perceives the surrounding environment and provides a basis for decision-making. Object detection is the core task in machine vision. 3D object detection can provide object steric size and location information. Compared with the 2D object detection widely studied in image coordinates, it can provide more applications of detection systems. Accurate LiDAR data has a stronger spatial capture capability and is insensitive to natural light, which makes LiDAR a potential sensor for 3D detection. Recently, deep neural network has been developed to learn powerful object features from sensor data. However, the sparsity of LiDAR point cloud data poses challenges to the network processing. Plenty of emerged efforts have been made to address this difficulty, but a comprehensive review literature is still lacking. The purpose of this article is to review the challenges and methodologies of 3D object detection networks using LiDAR data. On this account, we first give an outline of 3D detection task and LiDAR sensing techniques. Then we unfold the review of deep 3D detection networks with three kinds of LiDAR point cloud representations and their challenges. We next summarize evaluation metrics and performance of algorithms on three authoritative 3D detection benchmarks. Finally, we provide valuable insights of challenges and open issues.