Obtaining Long Trajectory Data of Disordered Traffic Using a Swarm of Unmanned Aerial Vehicles

Abstract

For the development of algorithms and models for driver behaviour at microscopic levels, trajectory data is needed.  Access to trajectory datasets under disordered traffic conditions (wide mix of vehicle types with the absence of lane discipline) is very limited. Particularly, existing datasets only cover shorter sections limiting the understanding of driving behaviour. This paper presents the design and preliminary results of a first-of-its-kind experiment to create a long trajectory dataset for an urban arterial road (Chennai city, India) under disordered traffic conditions using a swarm of six Unmanned Aerial Vehicles (UAVs). The steps followed to obtain the detailed trajectory data from UAV traffic videos are 1) Georegistration, 2) Image Stabilization and Stitching, 3) Data Annotation, 4) Vehicle Detection and Classification, and 5) Vehicle Tracking. Finally, to remove the noises and disturbances from the trajectory dataset, a Symmetric Exponential Moving Average (sEMA) filter technique was applied to smooth the positions of the vehicles and the first and second derivative of the positions, i.e., speeds and accelerations were obtained using the central difference method. The trajectory dataset was used to analyze the driver's lateral position preferences. The results highlight the unique behavior of different categories of vehicles under disordered traffic conditions. This dataset is an order of magnitude larger than existing datasets allowing, for the first time, to calibrate and validate car-following and lane-changing models as well as fully two-dimensional models in disordered traffic.

Keywords: Trajectory Data, Unmanned Aerial Vehicles, Deep learning, Vehicle Detection and Tracking, Lateral Distribution of Vehicles, Disordered Traffic