Research Of Computer Vision-based Vehicle Type Recognition

In recent years, vehicle type recognition plays a more important role in both traffic planning and road monitoring with the increasing number of vehicles. On the one hand, it can provide convenience for the department of vehicles in transportation management, on the other hand, it can also quickly realize the recognition, tracking and locating of a moving vehicle, such as a vehicle that causes a traffic accident and escapes from the scene of the accident.Aiming at the problem of not detecting the moving vehicles accurately and entirely under the complex traffic background, a new way which combines background subtraction method and inter-frame difference method was proposed based on the advantages of the two methods. With the proposed method, the moving vehicle can be detected accurately and entirely in both normal and complex scenes.Aiming at the problem that single feature of vehicle image can not describe the vehicle enough, in this paper both global and local features were considered for vehicle type recognition. The global feature was extracted by the improved Canny edge detection, which could also detect the vehicle's edge in the case of shadow. The local feature was extracted by Gabor transform. For the local feature, vehicle image was firstly divided into multiple non-overlapping patches, and then Gabor filter was used to extract the Gabor features of these patches. Considering the time-consumed computation due to high dimension of Gabor features, grey relational analysis was adapted to choose the most relevant Gabor features of vehicle image, which can improve the real-time performance of vehicle type recognition by reducing the dimension of the input Gabor features.In vehicle type recognition, the vehicle type is classified as big vehicle type and small vehicle type. The big vehicle type contains buses and trucks; the small vehicle type contains vans and sedans. A two-stage classification strategy was proposed in order to reduce the time-consumed computation due to combining all the features as the dictionary directly for sparse representation. In the first stage, the global feature was chosen as the basis of recognition for big vehicle and small vehicle. The pixel grey based image Euclidean distance and K nearest neighbor was employed. In the second stage, the optimized Gabor feature was used to construct the dictionary. The sparse representation combined with the reconstruction error and the Hamming distance to complete vehicle recognition. Experiment shows that the proposed method can recognize the vehicle type accurately and effectively. And it could also solve the occlusion problem in vehicle type recognition.