Robust Infrared Pedestrian Detection via GMR and Logistic Regression Based ROIs Extraction

Real-time detection of persons via infrared camera is a potential task for advanced driver assistance systems (ADAS). In this paper, we develop a simple and effective pedestrian detection strategy. To robustly locate the region of interest (ROI), we first design a gradient based descriptor which called gradient magnitude ratio (GMR) to depict the pedestrian target with very low computing cost. Then a logistic regression (LR) classifier is trained to identify the ROIs. After this coarse detection step, about one thousand ROIs is obtained in the experiments. Finally, the detected candidates are feed into a traditional HOG-SVM based classifying framework to gain the fine detection results. To furtherly enhance the real-time performance of the proposed system, PCA based dimension reduction is performed on the original HOG feature. Experiment results show that the proposed method could detect pedestrians effectively with realtime speed under challenging circumstances such as dynamic background, low contrast and various scales.