Real-time Recognition and Reidentification of Vehicles from Video Data with high Reidentification Rate

Abstract

Using video cameras regions of cities can be monitored in order to extract traffic data. Stationary cameras fixed at high buildings can be used which provide data 24 hours a day. The data can be used, e.g., to optimize traffic flow by controlling traffic lights dynamically. In order to minimize the number of cameras it is useful to reidentify vehicles leaving one monitored region and afterwards entering the viewing field of a further camera. From reidentified vehicles travel times can be obtained which are relevant parameters to optimize traffic control. In the present text, a method to reidentify vehicles based on extraction of 3-d-prototype vehicle models and color extraction from the top plane of vehicles is described. Shadows and light reflections on wet street are corrected, and therefore, the high recognition accuracy is achieved which is necessary to find the top plane of the vehicles. Due to the 3-d-model based analysis the cameras can be placed in a broad region of viewing angles. The algorithms are suitable for real-time applications. First results from video data of two cameras are presented which show a high reidentification rate with no false reidentification hypothesis.