PEDESTRIAN DETECTION BASED ON HYBRID FEATURES USING NEAR INFRARED IMAGES

摘要

This paper explores a hybrid-based method to fuse multi-slit features and Histograms of Oriented Gradients (HOG) features for pedestrian detection from Near Infrared (NIR) images. The fused feature set utilizes both the multi-slit method's capability of accurately capturing the local spatial layout of body parts (head, torso and legs) in individual frames and the HOG's capability in region information relevant to higher frequency components. The hybrid feature vector describing various types of poses is then constructed and used for detecting the pedestrians. The part based pattern matching analysis indicates that the fused features have much higher feature space separation than the pure features. Experiments with a database of NIR images show that the proposed method achieves a substantial improvement in tackling some difficult cases such as side view, back view which the conventional HOG method cannot handle. Detection and recognition performance is less computationally expensive than existing approaches.