Comparison of major LiDAR data-driven feature extraction methods for autonomous vehicles

Abstract

Object detection is one of the areas of computer vision that has matured very rapidly. Nowadays, developments in this research area have been playing special attention to the detection of objects in point clouds due to the emerging of high-resolution LiDAR sensors. However, data from a Light Detection and Ranging (LiDAR) sensor is not characterised by having consistency in relative pixel densities and introduces a third dimension, raising a set of drawbacks. The following paper presents a study on the requirements of 3D object detection for autonomous vehicles; presents an overview of the 3D object detection pipeline that generalises the operation principle of models based on point clouds; and categorises the recent works on methods to extract features and summarise their performance.