A novel set of pixel difference-based features for pedestrian detection
Xing Liu, Kar‐Ann Toh · 2018
We propose a feature extraction pipeline namely Difference Matrix Projection (DMP)for pedestrian detection. The goal is to design an effective feature set that can be efficiently computed. Our feature consists of pixel differences at different scales and orientations coupled with average pooling and block normalization. We develop a formulation that computes the difference maps and local average using global image projection instead of an iterative filtering. As a result, the computations can be expressed by analytic equations in terms of the input image. The projection matrices are pre-constructed, so they are readily applied to the image for feature extraction. Experiment results on the Daimler Chrysler (Daimler-CB) and NICTA pedestrian datasets show encouraging results, especially for low-resolution samples.