A multi-scale model integrating multiple features for vehicle detection

In traffic video surveillance systems, vehicles with various distances from the camera have different sizes, resolutions, and angles in traffic images. The common multi-scale method, which scales one vehicle template or the input image for detecting vehicles with different sizes, may fail to detect vehicles with various distances from the camera due to the change of the resolution and angle. To deal with this problem, we have proposed a multi-scale model including multiple templates with different scales and features. Our method includes two steps: constructing the multi-scale model and its probability model, and detecting vehicles from traffic images. In the first step, the multi-scale model is constructed by using three templates T₁, T₂, T₃ which represent vehicles with the short, medium, and long distance from the camera respectively. Each template contains one or some combination of sketch, texture, flatness, and color. In the second step, the three templates are applied for vehicle detection by using the template matching with local maximization operations. The main innovation of this paper is that the combination of multi-template and multi-scale method is applied to detect vehicles with various distances from the camera. To test our method, we have done several experiments on various traffic conditions. The experimental results show that our method effectively copes with vehicles with various distances from the camera and provides the detailed vehicle information after vehicle detection. Moreover, our method adapts to various weather conditions, slight pose variance, and slight occlusion.