Pose-specific pedestrian classificatiion using multiple features in par-infrared images

Abstract

We present a multiple feature-based, pose-specific pedestrian classification approach to improve classification performance for fair-infrared (FIR) images. Using pose-specific classifiers and multiple features has proved to be beneficial in visible-spectrum-based classification systems; therefore, we adapt both to an FIR-based classification system. For pose-specific classifiers, we separate poses into sets of front/back and right/left poses and estimate the pose using template matching. For feature extraction, we use histograms of local intensity differences (HLID) and local binary patterns (LBP). Experiments showed that the proposed approaches improve the classification performance of a baseline HLID/linSVM approach.