Applying Graph Theory to Automatic Vehicle Tracking by Remote Sensing
Carlos Lima Azevedo, João Lourenço Cardoso, Moshe E Ben-Akiva · Transportation Research Board 93rd Annual MeetingTransportation Research Board · 2014
Abstract 2 3 The estimation of driving behavior models relies on the access to detailed traffic information such as 4 vehicle trajectories. Recent developments in vision-based technologies have allowed an increased 5 collection of vehicle trajectories around the world, with an emphasis on aerial or high observation point 6 imagery methods. Several computer algorithms have been proposed using images of different traffic 7 scenarios with the specific aim of detecting and tracking road users. 8 Very recently, multiple-object tracking based on constrained flow optimization has been shown to 9 produce very satisfactory results. Generally, this method uses individual image features collected for each 10 candidate vehicle position as main criteria in the optimization process. Although these methods are very 11 effective in controlled scenarios, adverse conditions such as dynamic view points and wider observation 12 areas with low ground sampling distances are known to encumber significantly the vehicle trajectory 13 extraction task. 14 In this paper we present the application of a k-shortest disjoint paths algorithm for multiple-object 15 tracking using a motion-based optimization based on dual graphs. A graph of possible connections 16 between successive candidate positions was built using speed and lane connectivity. Dual graphs were 17 constructed to allow for acceleration and lane-change-based optimization criteria. The k-shortest disjoint 18 paths algorithm was then used to determine the optimal set of trajectories (paths). 19 The proposed algorithm was successfully applied to the vehicle tracking in the A44 suburban 20 motorway, in Portugal. Vehicle positions were detected by image processing and 99.4% of the trajectories 21 were successfully and efficiently extracted using the proposed method.