Pedestrian Detection and Tracking Based on Automotive Vision

GUO Lie,GAO Long,ZHAO Zongyan

DOI: https://doi.org/10.3969/j.issn.0258-2724.2012.01.004

2012-01-01

Abstract:To improve the automotive active safety and guarantee the safety of pedestrians under urban transportation conditions, a pedestrian protection method based on automotive vision was presented. The Adaboost algorithm was utilized to detect pedestrians rapidly, and the Kalman filter principle was adopted to track these pedestrians and obtain their trajectories. With this method, the samples' Haar-like features are calculated and trained by the discrete Adaboost algorithm to obtain the cascaded pedestrian recognition classifiers. These classifiers are exploited to search for pedestrians by scanning those images captured by automotive vision. The Kalman filtering principle is applied to track these pedestrians and build the dynamic region of interest for pedestrian detection. The tracking results are used to analyze their behaviors. The experimental results show that the proposed method can detect pedestrians in about 80 ms per frame with an accuracy of 88%. The time cost can reduce to 55 ms per frame after using the Kalman-based pedestrian tracking method.