LiDAR data segmentation in off-road environment using Convolutional Neural Networks (CNN) 2020-01-0696
Recent developments in the area of autonomous vehicle navigation have emphasized algorithm development for the characterization of LiDAR 3D point-cloud data. The LiDAR sensor data provide a detailed understanding of the environment surrounding the vehicle for safe navigation. However, the LiDAR point cloud datasets need point-level labels which require significant amount of annotation effort. We present a framework which generates simulated labeled point cloud data. The simulated lidar data was generated by a physics-based platform, the Mississippi State University Autonomous Vehicle Simulator (MAVS). In this work, we have developed and tested algorithms for autonomous ground vehicle off-road navigation. The MAVS framework generates 3D point cloud for off-road environment that includes trails and trees.
The important first step in off-road autonomous navigation is the accurate segmentation of 3D point cloud data to identify the potential obstacles in the vehicle path. We have used simulated lidar data to segment and detect obstacles using Convolutional Neural Networks (CNN). Our analysis is based on SqueezeSeg, a CNN-based model for point cloud segmentation. The CNN has been trained with the labelled dataset of off-road imagery generated from MAVS and evaluated on the simulated dataset. The segmentation of the lidar data is done by point-wise classification and the results show excellent accuracy in identifying different objects and obstacles in the vehicle path.
Lalitha Dabbiru, Chris Goodin, Nicklaus Scherrer, Daniel Carruth