Data-Driven Performance Evaluation of Machine Learning for Velocity Estimation Based on Scan Artifacts from LiDAR Sensors

Light detection and ranging (LiDAR) sensors are increasingly applied to automated driving vehicles. Microelectromechanical systems are an established technology for making LiDAR sensors cost-effective and mechanically robust for automotive applications. These sensors scan their environment using a pulsed laser to record a point cloud. The scanning process leads in the point cloud to a distortion of objects with a relative velocity to the sensor. The consecutive generation and processing of points offers the opportunity to enrich the measured object data from the LiDAR sensors with velocity information by extracting information with the help of machine learning, without the need for object tracking. Turning it into a so-called 4D-LiDAR. This allows object detection, object tracking, and sensor data fusion based on LiDAR sensor data to be optimized. Moreover, this affects all overlying levels of autonomous driving functions or advanced driver assistance systems. However, since such sensor-specific effects are rarely available in public datasets and the velocities of target objects are not included as ground truth in these datasets, it makes sense to enrich the limited real-world data with synthetic data. Therefore, this article discusses how such datasets can be created and combined to efficiently estimate velocities on real-world data using the novel method named VeloPoints.