Real-Time Deep ConvNet-Based Vehicle Detection Using 3D-LIDAR Reflection Intensity Data

Abstract

This paper addresses the problem of vehicle detection using a little explored LIDAR’s modality: the reflection intensity. LIDAR reflection measures the ratio of the received beam sent to a surface, which depends upon the distance, material, and the angle between surface normal and the ray. Considering a 3D-LIDAR mounted on board a robotic vehicle, which is calibrated with respect to a monocular camera, a Dense Reflection Map (DRM) is generated from the projected sparse LIDAR’s reflectance intensity, and inputted to a Deep Convolutional Neural Network (ConvNet) object detection framework for the vehicle detection. The performance on the KITTI is superior to some of the approaches that use LIDAR’s range-value, and hence it demonstrates the usability of LIDAR’s reflection for vehicle detection.