Improvement CoHOG Detection and Texture Matching Tracking for Pedestrian's Active Safety

Abstract

The total number of deaths by traffic accidents was 4,863 in 2010 and the percentage of pedestrians in the deaths tended to increase in Japan. In recent years, to provide measures for accident avoidance has been urgently required for the people on the roads or streets including pedestrians and cyclists, so called “vulnerable road users”. In the present study, technologies to detect pedestrians by forward images with cameras installed in vehicles are under development. The purpose in this study is to improve the technologies to detect and track pedestrians with a high degree of accuracy of images with a single camera. In this study, the authors predicted pedestrians' activities by monitoring pedestrians for a given length of time and developed an algorithm to recognize pedestrians and to track the pedestrian movements with a higher degree of accuracy. The effectiveness of the algorithm was investigated by pedestrians' images on real-world roads. Regarding detecting the pedestrian feature descriptor, the authors newly found that an ECoHOG accumulating the integration of gradient intensities was the best method to decrease both the undetectable and excessive detectable ratio. Therefore, the usage of the new method by images captured on the real road was validated.