Deformable Part-based Model For Specific Crossroad Recognition

Unmanned vehicle technology is an integrated technology which requires the integration of computer vision, machine learning, intelligent planning and control, and many other advanced technologies. Among the numerous works on the unmanned vehicle, little has been done about the speci?c crossroad recognition. However, the accurate identi?cation of the speci?c crossroad is extremely important for the unmanned vehicle, especially when it aims to perform the long-distance driving which comes along with many crossroads.Funded by the National Natural Science Foundation project which aims to complete 2000 km autonomous driving, this thesis mainly tackles the problem of speci?c crossroad recognition and proposes a novel method based on the deformable part model. This method ?rstly utilize the latent support vector machine to train the deformable part model for each speci?c crossroad. The weight for each part is also obtained during the training of the latent support vector machine. The recognition of the speci?c crossroad is achieved by the combination of the holistic image feature matching and the local part feature matching.The experiments con?rmed the validity of the proposed method. To further improve the accuracy, we propose to fuse the information from consecutive frames.In order to make the scores from different crossroads more comparable, we use the sigmoid function to normalize the scores, and propose an approach to optimize the parameters of the sigmoid function. Experiments show that the use of consecutive information plays a signi?cant role in improving accuracy of crossroad recognition.