Detection of Multiple and Partially Occluded Humans Based on Adaptive Background Subtraction and Improved Histogram of Oriented Gradients

Abstract

Human detection in a video surveillance system has vast application areas including suspicious event detection and human activity recognition. In the current environment of our society suspicious event detection is a burning issue. For that reason, this paper proposes a framework for detecting humans in different appearances and poses by generating a human feature vector. Initially, every pixel of a frame is represented as an incorporation of several Gaussians and use a probabilistic method to refurbish the representation. These Gaussian representations are then estimated to classify the background pixels from foreground pixels. Shadow regions are eliminated from foreground by utilizing a Hue-Intensity disparity value between background and current frame. Then morphological operation is used to remove discontinuities in the foreground extracted from the shadow elimination process. Image correlogram is used to label objects within a group. After that, the framework generates ROIs by determining which of the foregrounds represent human by considering conditions related to human body. Finally, features are extracted from ROI for classification. A feature descriptor, improved histogram of oriented gradients (ImHOG) is proposed to alleviate the limitation of Histogram of Oriented Gradients (HOG).Various videos containing moving humans are utilized to test the proposed framework and acquired an over 94% human object detection rate. The proposed framework detected humans from continuous frame sequences with higher adaptability and precision.