Abstract
Collision Avoidance New Car Assessment Program was issued towards reducing traffic accidents in the USA. Cost reduction is essential for spreading these systems and making this program effective. A multi-function front camera system appears promising, which can realize two functions in CA NCAP by one sensor. However, obstacle recognition applying image processing is still a difficult challenge due to the severe illumination condition in general outdoor environment. Furthermore, an application specific IC is necessary because the calculation amount for the recognition is large, which raises the cost. We have developed the novel vehicle recognition algorithm that is a core technology in order to resolve these challenges. The basic concepts is as follows. The image patterns stochastically appeared with a high frequency in the actual environment are preferentially searched evaluating the detection confidence. If enough high confidence is obtained, further calculation is terminated. Thus, unnecessary calculation can be omitted. With this concept, the robust vehicle recognition algorithm for illumination fluctuation with high calculation efficiency can be realized. Furthermore, adjustments for specific recognition errors become easy, which are difficult for the general learning algorithms utilizing general-purpose feature amounts to adjust, because the pattern classification knowledge is structured.
Keywords: forward collision warning, monocular camera, image recognition, CA NCAP