Vehicle Tracking Using a Multi-scale Bayesian Algorithm for a Perspective Image of a Road

Abstract

Tracking the vehicles in the short distance from the camera using the perspective view of the camera installed above the road requires considering the effect of perspective. Vehicles that are close to the camera are large and easily distinguishable on individual video frames. Moreover, the separation between adjacent vehicles is also high. However, tracking capabilities are deteriorated for longer distances of the vehicle from the camera. The article presents a solution for tracking the vehicles moving away from the camera, based on the image analysis at different scales in order to increase the range of the tracking system. The proposed algorithm utilizes a block matching technique using the correlation coefficients but the size of matched blocks varies for different video frames. Matching can be implemented in several variations depending on the choice of the reference blocks only in the previous frame or averaging of several frames. A particularly useful technique, used in the paper, is the spatio-temporal recursive Track-Before-Detect algorithm, especially for distant objects represented by a small number of pixels.