Automatic estimation of vehicle activity from airborne thermal infrared video of urban areas by trajectory classification

Summary: Analysis of traffic data plays an important role in urban and spatial planning. Thermal Infrared (TIR) video cameras have capabilities to operate at day and night and to acquire the scene sampled with video frame rate. In this paper a strategy for the estimation of vehicle motion and the assessment of traffic activity from airborne TIR video is presented. In contrast to other approaches we handle detecting and tracking vehicles in the video separately, because moving as well as stationary vehicles are intended to be detected. Firstly, vehicles are detected in single frames of the video. Additionally, tie points are detected for co-registration and compensation the sensor movement. Afterwards, a stepwise grouping of image points considering temporal consistence and geometric relation is carried out to determine the vehicle trajectories and classify them into stationary, moving and uncertain dynamical categories. The vehicles are then integrated into the classes "moving," "stationary" and "uncertain" categories. Additionally, in consideration of matching vehicle-related image patches for moving vehicles, the topology of the trajectories are investigated and optimized in order to eliminate disturbances and estimate velocities. The algorithms were tested with video sequence of urban areas in nadir-view and oblique-view. The correctness of the results is achieved higher than 75% for both views.