Capturing the true bounding boxes: vehicle kinematic data extraction using unmanned aerial vehicles

Abstract

This paper presents a methodology by which kinematic variables of road vehicles can be extracted from unmanned aerial vehicle (UAV) footage. The oriented bounding boxes of the vehicles are identified based on the aerial view of the intersection, and the kinematic variables, such as position, longitudinal velocity, lateral velocity, yaw angle and yaw rate, are determined. The bounding boxes are converted to the perspective of a roadside camera using homography, to generate labeled data sets for training the machine learning-based perception systems of smart intersections. Compared to ordinary GPS data-based technology, the proposed method provides smoother data and more information about the dynamics of the vehicles. In the meantime, it does not require any additional instrumentation on the vehicles. The extracted kinematic variables can be used for motion prediction of road traffic participants and for control of connected automated vehicles (CAVs) in intelligent transportation systems.

Keywords:

• data sets for machine learning algorithms
• kinematic variables
• unmanned aerial vehicles
• vehicle tracking
• video processing

Acknowledgements

Dénes Takács was supported by the János Bolyai Research Scholarship of the Hungarian Academy of Sciences, and he would also like to thank the Rosztoczy Foundation for their generous support. The authors would like to thank Anil Alan, Xunbi Ji, Sanghoon Oh, Minghao Shen and Hao Wang for their help in the experiments.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This research was supported by the University of Michigan’s Center for Connected and Automated Transportation through the US DOT grant [69A3552348305] and by the National Research, Development and Innovation Office of Hungary under grant no. [NKFI-146201].