Semi-Automatic Ground Truth Trajectory Estimation and Smoothing using Roadside Cameras

High precision trajectory data is crucial for the validation and verification of algorithms for automated driving. For instance, ego-vehicle-localization algorithms, traffic prediction, traffic scene assessment and other components profit from highly accurate trajectory data for evaluation and bench-marking. In this work we present a ground truth data generation pipeline that is able to produce trajectories of traffic participants in a semi-automated process from static roadside cameras. Our approach consists of an assisted manual labeling step, homography projection, followed by feature computation, state estimation and trajectory smoothing using Rauch-Tung-Striebel Smoothers (RTS). We evaluate our approach in a field experiment and compare the produced trajectories to a commercial high precision (GNSS/INS) system, where we reach a mean position error of 1.16m and a mean speed error of 0.73m/s over seven driving sequences. For relative object distances below 40m to the camera origin, position errors are below 0.41m and the respective speed error is below 0.30m/s.