Robust model for vehicle type identification in video traffic surveillance

Vehicle classification is an inherently difficult problem. Most of researches for vehicle type recognition use images where there are only one vehicle in restricted conditions. In traffic surveillance videos have many different conditions, which increase the degree of difficulty in recognizing the type of vehicle. Thus, the various restrictions in the conventional models make them limited, creating the need of sophisticated models that combine segmentation techniques that allow to extract the information needed to recognize a vehicle within a complex scenario. This work presents a model for vehicle type recognition in traffic surveillance videos. The main obstacle in this kind of videos is the great quantity of information and the constantly variations in the scene. This work presents a model based on local features. Our proposed method is divided into two stages. In first stage, the moving objects are segmented using frame difference techniques, the background image is progressively generated by a heuristic function. In the second stage, each segment(image region with one or more vehicles) is processed, a local descritor is used for feature extraction and this information is organized in a visual vocabulary. A SVM classifier is used for recognizing occlusions and the type of vehicle. We introduce a very simple method to remove occlusions, this method is based on intensity level reduction.