Real-time recognition of multiple pedestrians using car-mounted camera

In ITS, there has recently been emphasis on the importance of developing a driving assistance system which can detect pedestrians with a car-mounted camera and warn of their presence in order to assure pedestrian safety. In particular, the ability to detect multiple pedestrians in real time (at a video rate of 30 fps) is indispensable from the practical point of view. Consequently, this study proposes a method of recognizing multiple moving objects (pedestrians) in real time using dynamic scene input from a car-mounted camera, as a basic technology for the driving assistance system. The presence of a pedestrian is evaluated on the basis of hue information in the HSI color representation system and model-based matching using a circular approximation model. Thus, in terms of the evaluation function, the problem of recognition of multiple pedestrians is replaced by the problem of search for multiple local maxima. However, it is difficult to realize the process in real time if the evaluation value is calculated for the whole image. This paper considers the "attention," which is a function that reduces the burden of human visual information processing, and proposes a method in which a pseudo-attention function is realized by using a hybrid GA, allowing multiple local maxima to be retrieved efficiently in real time. A real-time experiment on the recognition of multiple pedestrians in dynamic images obtained from a car-mounted camera was used to verify the effectiveness of the proposed method. In addition, the practical performance of the system was investigated by an experiment on the recognition of suddenly appearing pedestrians.