Exploiting domain constraints for exemplar based bus detection for traffic scheduling

We describe a computer vision application to adaptive traffic signal control. Adaptive traffic signal control systems allocate green time at an intersection dynamically in response to sensed incoming traffic flows to continually optimize overall throughput. Existing video-based systems are limited to the use of vehicle presence and volume (count) data, and since it is not possible to distinguish between different types of vehicles, optimization opportunities can be missed. We propose to detect specific types of vehicles, such as buses, so that they can be assigned higher priority when appropriate and the overall effectiveness of the adaptive signal system can be improved. We base our system design on current visual recognition technology (HOG SVM) and exploit configuration constraints specific to this application, such as knowledge about the anticipated scale of the vehicles. The challenge is to be robust to varying illumination and weather conditions, occlusions from other vehicles, and large variations in scale while producing recognition results in real-time. We show results on challenging data from traffic cameras under different observation conditions and at varying ranges.