3D Object Detection for Self-Driving Vehicles Enhanced by Object Velocity

Abstract

A large number of vision sensors has been proposed for enabling self-driving vehicles to perceive their surroundings. Among them, Light Detection And Ranging (LiDAR) presents the unique advantage of acquiring a high resolution 3D representation of the vehicle surroundings, in the form of point clouds, which enables accurate 3D object detection. The success of the first (and current) generation LiDARs has motivated the development of a second generation of this sensor, now based on coherent detection. Second generation LiDARs thus enable not only estimating radial distance, but also radial velocity for each point of the point cloud. The objective of this work is to investigate which benefits can be obtained by considering such an additional information – radial velocity – in 3D object detection. Results show that considering object velocity is particularly helpful in objects represented by a small number of points.

Keywords

• Autonomous driving
• 3D object detection
• coherent LiDAR
• point cloud
• radial velocity
• deep learning