Study on image processing for video-based traffic measurement and vehicle classification

The accurate classification of vehicles is valuable for the analysis of highway traffic and the improvement of highway performance. Recently, extensive efforts have been devoted to video-based measurement and analysis of traffic flow, which will greatly benefit current and potential applications, including traffic control and analysis, reading and storage, toll-collection violation detection, and license plate finding. This study focuses on adaptive vehicle detection and classification algorithms. A new approach using a restricted background-updating algorithm, has been developed to adaptively obtain the traffic background from which the passing vehicles are identified. A feed-forward, three-layer neural network classifier with the hidden layer being trained by the error back-propagation algorithm is introduced to adapt to the disturbances that affect the classification, such as spatial occlusion, varying illumination conditions, tilt of object, noise induced during the processing course, different kinds of vehicles with different colours and shapes. Moreover, by using the plane-to-plane coordinate transformation technique, a map of planes from the non-vertical angle photography scene to real roadway is achieved, which is essential for the determining of traffic parameters. From the experimental results, we demonstrate that our approach shows good performance both on accuracy and stability.