A Two-Staged Approach to Vision-Based Pedestrian Recognition Using Haar and Hog Features

Abstract

This article presents a two-staged approach to recognize pedestrians in video sequences on board of a moving vehicle. The system combines the advantages of two feature families by splitting the recognition process into two stages: In the first stage, a fast search mechanism based on simple features is applied to detect interesting regions. The second stage uses a computationally more expensive, but also more accurate set of features on these regions to classify them into pedestrian and non-pedestrian. We compared various feature extraction configurations of different complexities regarding classification performance and speed. The complete system was evaluated on a number of labeled test videos taken from real-world drives and also compared against a publicly available pedestrian detector. This first system version analyzes only single image frames without using any temporal information like tracking. Still, it achieves good recognition performance at reasonable run time.