Accurate Trajectory Extraction of Dynamic Targets for Driving Behaviour Analysis

The performance of autonomous driving algorithm relies heavily on the quality and quantity of diverse motion datasets. However, existing interactive motion datasets are typically constructed from complicated sensors or mobile intelligent vehicles, and post-process with manual annotation, which are costly, inefficient and not easy to expand. And these datasets represent only one specific driving scene. In our work, we propose a nearly automatic and complete framework to accurately extract a large number of trajectories and behavioural data from fixed traffic monitoring cameras, which leverages advanced visual algorithms. Our framework contains an automatic and evolutionary optimization for camera calibration, which can be executed iteratively based on a common prior knowledge set. We also introduce a new method for ground position and orientation calculation. A series of postprocessing processes is added to ensure accuracy and diversity. Finally, the data content is expanded with motion- related features, coarse target information and behaviour related features. The results demonstrate that a large number of motion trajectories can easily be obtained through the proposed framework and that the average precision of the trajectory can reach 36 cm, which is precise enough for actual behavioural analysis.