Hausdorff Distance-based Template Matching For Pedestrian Detection

Pedestrian detection system is a hot issue in the research field of intelligent transportation system, with great practical significance. This article describes a real time pedestrian detection system for night driving with a near infrared camera, which takes advantage of the algorithms in computer vision and pattern recognition.In the framework of the whole system, stationary object detection is the most important and foundational function. Practically, both pedestrians and background appears widely variety, which makes the stationary detection particularly challenging. Based on the filter approach, we use cascaded classifiers to implement the detection. Among them, shape based classification is the central part, which serves the most primary recognition, and directly affects the detection performance of whole system.Based on analysis of the unreliability of candidate segmentation, and pedestrian features, this article introduces shape recognition on edge feature, and the Hausdorff distance-based template matching for the design of shape based classification.There are three main novel contributions in this article.Firstly, based on analyzing the particular features of segmented candidates, invents the two-layer shape based classification on two different feature-extraction methods, and combines the binarizing processed area and edge feature for shape representation. Furthermore, introduces new adaptive edge extraction and binarizing algorithm, which improve the effect of shape feature extraction. The third one is the combination of Hausdorff distance measure and template matching, which is applied to the two sub-classifiers for layered shape based classification.Numbers of testing experiments demonstrate that, comparing with the original shape based classification method, this new approach is more promising, with better applicability and lower false detection rate.