Deformable Part Model Based Crossroad Recognition for Unmanned Vehicles

GPS is very popular for locating the unmanned vehicles. However, it cannot solve all the problems about location. As crossroad is very important for unmanned vehicles, it is necessary for unmanned vehicles to recognize the crossroad and to load some other recognition tasks (such as traffic light recognition and traffic sign recognition) at the same time. The precision of GPS can hardly tell us if the traffic lights or traffic signs are shown in images. The best way to obtain this is using image to recognize the crossroads. However, most recent works about scene recognition can not solve the problem of specific crossroad recognition well because of weakening the special structure. The latent spacial structure is a key factor that makes specific crossroad different. There are some differences between their idea and ours. In this paper, a method is proposed to help unmanned vehicles recognize the specific crossroads based on deformable part models (DPM´s) with latent SVM training. In order to validate the method, a mid-large database is built. The excellent performaence of this method is shown in the experiments and illustrates this method to be promising in the future.