Obstacles Occluded Regions Prediction and Speed Control for Autonomous Mobile Robot using Generative Adversarial Networks

Abstract

Many autonomous vehicles use LiDAR (Light Detection and Ranging) as external sensor to recognize surrounding environment. LiDAR sensor enables high resolution measurement by the emission of the laser scan lines. While in case of it is blocked by the obstacles, the sensor can’t measure the backside regions of it, and it is the measuring loss which is called occlusion. It makes the robot motion in unknown environment unsafe so pre-measured map has been used to let these robots recognize environment and to control the motion of it in previous studies. This study proposed an occlusion prediction method for speed control without the pre-measured map in unknown environment. This method is based on GAN (Generative Adversarial Networks) which is a type of deep learning model to predict the occluded regions from LiDAR measurement as local map for the robot. As a result, we showed applicability of this model for the robot speed control using the 3D robot simulator without the pre-measured map in unknown environment.