Human detection utilizing adaptive background mixture models and improved histogram of oriented gradients

Abstract

Detecting human is a crux issue in computer vision, with numerous usages especially in human–computer interaction and video surveillance. A framework for human detection with various poses and appearances is proposed in this paper. Initially, a background model is utilized to generate the background image and foreground pixels are classified. Then, HSI color model and color correlogram for removing shadow region and partial occlusion handling are used, respectively. After that, the framework extracts Regions of Interest (ROIs) by analyzing the structure of human body. Finally, features are generated from ROI for classification. A feature descriptor, Improved Histogram of Oriented Gradients (ImHOG), is proposed to alleviate the limitation of HOG. The proposed framework is tested using various videos and the result demonstrates remarkable efficiency with effectiveness.