Computer Graphics-Driven 3D LiDAR Model for Autonomous Driving Systems

In recent years, research and development applied to the topic of autonomous vehicles has experienced a substantial growth due to its significant impact on economic, cultural, and environmental sectors. Models based on supervised learning, using data from LiDAR sensors, have been one of the most widely used approaches to solve the task of object detection and tracking on roads. However, this solution presents some implementation difficulties due to the challenge of generating large amounts of labeled data needed to train machine and deep learning models in an appropriate way. To tackle this problem, in this paper a LiDAR sensor model was developed using Unreal Engine 4 on CARLA, an open-source simulator. The proposed model goes beyond the state-of-the-art, since it takes into account several aspects of actual LiDAR imaging, including multiple reflections of the light beam, beam divergence, the incidence angle on the impacted obstacle, and the obstacle’s material to calculate the beam’s received intensity and its time delay. Additionally, it has the capability to generate data in three ways: labeled point clouds, temporal signal, and the distance of impacts from each channel of the LiDAR sensor. The performance of the proposed model was evaluated using two approaches: First, by training neural network models with the addition of synthetic data; and second, by comparing the data generated by the developed model against that from a real LiDAR. In the first approach, training of various different object detection models led to detection improvements of up to 10%. In the second approach, a real Velodyne VLP-16 LiDAR was used in a known 3D-modeled environment, resulting in a point distance error of 2.63 cm at its maximum, which is below the 3 cm error specified by the sensor manufacturer.