End-to-End Synthetic LiDAR Point Cloud Data Generation and Deep Learning Validation

LiDAR sensors are common in automated driving due to their high accuracy. However, LiDAR processing algorithm development suffers from lack of diverse training data, partly due to sensors’ high cost and rapid development cycles. Public datasets (e.g. KITTI) offer poor coverage of edge cases, whereas these samples are essential for safer self-driving. We address the unmet need for abundant, high-quality LiDAR data with the development of a synthetic LiDAR point cloud generation tool and validate this tool’s performance using the KITTI-trained PIXOR object detection model. The tool uses a single camera raycasting process and filtering techniques to generate segmented and annotated class specific datasets. This approach will support low-cost bulk generation of accurate data for training advanced selfdriving algorithms, with configurability to simulate existing and upcoming LiDAR configurations possessing varied channels, range, vertical and horizontal fields of view, and angular resolution. In comparison with virtual LiDAR solutions like CARLA [1], this tool requires no game development knowledge and is faster to set up: sensor customization can be done in a front-end panel enabling users to focus more on data generation. The simulator is developed using the Unity Game Engine in conjunction with free and open-source assets, and a build will be shared with the AV community before an open-source release.