Discriminative Feature Points Distribution in Near-Infrared Pedestrian Images

抄録

Pedestrian detection is an important area in computer vision and plays a key role in intelligent vehicles. The objective of this paper is to provide a method to select the discriminative visual word based on the bag-of-features algorithm for pedestrian detection from near-infrared images. It calculates the difference in the total appearance frequency for each visual word of pedestrian and non-pedestrian images. The visual words which exhibit greater absolute values are considered as more efficient for pedestrian detection and can be selected. The experiment results showed that the proposed method keeps nearly the same detection accuracy even if only 40% of the visual word is selected. In addition, we investigated the distribution of discriminative feature points which belong to the selected visual words from near-infrared images and visible spectrum images.