Automatic vehicle trajectory extraction by aerial remote sensing

Research in road users' behaviour typically depends on detailed observational data availability, particularly if the interest is in driving behaviour modelling. Among this type of data, vehicle trajectories are an important source of information for traffic flow theory, driving behaviour modelling, innovation in traffic management and safety and environmental studies. Recent developments in sensing technologies and image processing algorithms reduced the resources (time and costs) required for detailed traffic data collection, promoting the feasibility of site-based and vehicle-based naturalistic driving observation. For testing the core models of a traffic microsimulation application for safety assessment, vehicle trajectories were collected by remote sensing on a typical Portuguese suburban motorway. Multiple short flights over a stretch of an urban motorway allowed for the collection of several partial vehicle trajectories. In this paper the technical details of each step of the methodology used is presented: image collection, image processing, vehicle identification and vehicle tracking. To collect the images, a high-resolution camera was mounted on an aircraft's gyroscopic platform. The camera was connected to a DGPS for extraction of the camera position and allowed the collection of high resolution images at a low frame rate of 2s. After generic image orthorrectification using the flight details and the terrain model, computer vision techniques were used for line rectification: the scale-invariant feature transform algorithm was used for detection and description of image features, and the random sample consensus algorithm for feature matching. Vehicle detection was carried out by median-based background subtraction. After the computation of the detected foreground and the shadow detection using a spectral ratio technique, region segmentation was used to identify candidates for vehicle positions. Finally, vehicles were tracked using a k-shortest disjoints paths algorithm. This approach allows for the optimization of an entire set of trajectories against all possible position candidates using motion-based optimization. Besides the importance of a new trajectory dataset that allows the development of new behavioural models and the validation of existing ones, this paper also describes the application of state-of-the-art algorithms and methods that significantly minimize the resources needed for such data collection. (C) 2013 The Authors. Published by Elsevier Ltd. Selection and/or peer-review under responsibility of Scientific Committee