This paper describes a robust method for recognizing vehicles. The system is based on local-feature configuration, and has already shown that it works very well in infrared images and optical images. The algorithm is based on our previous work, which is a generalization of the eigen-window method. This method has the following three advantages: (1) it can detect even if part of vehicles is occluded; (2) it can detect even if vehicles are translated due to running out of the lanes; and (3) it does not require us to segment vehicle areas from input images. But there is a problem in this method: the system requires large amount of training images to make models of the target vehicles. Collecting training images of the target vehicles is generally a time consuming and hard task. In order to solve the problem, models have been made from computer graphics (CG), not from real images. Because it is easy to obtain various kinds of views for CG vehicles, many training images can be created in a short time. IT has also been confirmed that the system using CG models is effective to real images, performing outdoor experiments. CG models can recognize vehicles in real images without loss of accuracy.
Abstract