Pedestrian detection from still images based on multi-feature covariances

Abstract

This paper targets the detection of pedestrians from still images, which focuses on developing robust feature representations that encode image regions as covariance matrices to support high accuracy pedestrian/non-pedestrian decisions. Firstly we utilize a fast method for computation of covariances based on integral images. By integrating the advantages of both covariance-based object detection and HOG-and FDF-based pedestrian detection, we then introduce four new feature representations for training a pedestrian detector: Covariance-based first-order Histogram of Oriented Gradient (Cov-HOG1), Covariance-based second-order Histogram of Oriented Gradient (Cov-HOG2), Covariance-based first-order Four Directional Features (Cov-FDF1), and Covariance-based second-order Four Directional Features (Cov-FDF2). To test our feature sets, we adopt a relatively simple learning framework that uses LogitBoost algorithm to classify each possible image region as a pedestrian or as a non-pedestrian. The experimental results show that the proposed algorithm obtains satisfactory pedestrian detection performances on the INRIA person datasets as well as images collected from Google and Flickr websites.