Multi-posture Human Detection Based on Hybrid HOG-BO Feature

Abstract

Human detection in images is a fast growing and challenging area of research in computer vision with its main application in video surveillance, robotics, intelligent vehicle, image retrieval, defense, entertainment, behavior analysis, tracking, forensic science, medicalscience and intelligent transportation. This paper presents a robust multi-posture human detection system in images based on local feature descriptors such as HOG and BO (Block Orientation). The proposed system employs LLE method to achieve dimensionality reduction on the Hog feature descriptors and thus reduce time complexity. Performance of the proposed method is evaluated using feature and classifier based schemes with different datasets. By using classifier based schemes, fast-additive SVM outperforms other SVM classifiers. The combined feature vector can retain precision of HOG as well as improve the detection rate. The experiment results on INRIA person, SDL dataset, and TUDBrussels dataset demonstrate that combined feature vector along with LLE and fast additive SVM significantly improves the performance.