Vehicle detection based on And–Or Graph and Hybrid Image Templates for complex urban traffic conditions

Abstract

In complex urban traffic conditions, occlusions among vehicles and between vehicles and non-vehicle objects are very common, which presents a major challenge to current vehicle detection methods. To circumvent this problem, we have proposed a vehicle detection method based on an And–Or Graph (AOG) and Hybrid Image Templates (HITs). In our AOG, the vehicle object is hierarchically decomposed into multiple vehicle parts by up–down and left–right division to reduce the impacts of vehicle occlusion. Furthermore, the vehicle parts are modeled by HITs to differentiate vehicles from non-vehicle objects effectively. These HITs integrate multiple features including sketch, texture, color and flatness so as to well describe the vehicle features. To test the performance of the proposed method, we have conducted a quantitative experiment and a comparison experiment. The experimental results show that, by combining AOG and HIT for vehicle identification, severe occlusions among vehicles and non-vehicle objects under complex urban traffic environments can be dealt with efficiently. Furthermore, the results also indicated that our method can adapt to variations in vehicle poses and shapes.