Pedestrian detection in depth images using framelet regularization

Pedestrian detection in depth images can reduce the effect of clothing in appearance and unpredictable illumination changes, and its applications include robotics and surveillance, etc. Due to the TOF depth images easily contaminated by noise, a parameter-free and framelet-based regularization approach is proposed to remove noise and preserve the object shape in depth images before feature detection and classifying. After noise removal, the histogram of depth difference (HDD) is utilized as a features descriptor and SVM with a linear kernel is adopted as a classifier. Experiments show that the framelet-based approach is adaptive and effective to denoise depth images and pedestrian detection in denoising depth images is feasible. The miss rate decreases from 9.1% of noisy images to 2.2% of denosing images at FPPW=10⁻⁴.