Optimized parallel implementation of pedestrian tracking using HOG features on GPU

Accurate pedestrian recognition is required for practical applications such as automotive and security applications. To improve accuracy of recognition, accurate tracking is indispensable just as detection. The authors proposed a novel accurate tracking scheme using HOG features and its parallel implementation on GPU aiming real-time processing. However, the implementation does not have enough performance because the optimization is not sufficient. In this paper, we propose optimized implementation of HOG-based pedestrian tracking on GPU. By the proposed implementation, the total processing speed becomes 2.6 times faster than that of original one and real-time processing is achieved.