Human count estimation in high density crowd images and videos

Abstract

This paper addresses the issue of detection and counting of people in exceedingly swarmed crowd images and video scenes. The identification of individual objects has observed enormous advancements over the late years, but crowd scenes still remain predominantly challenging for detection and counting purposes because of substantial impediments and occlusions, high group densities and large disparities in individuals' sizes and appearances. To address these difficulties, we propose epsilon-Support Value Regression (SVR) fusion-based approach to clout information on the global construction of the crowd scenes and to identify the people in these scenes. We prepared and tested our approach on a new dataset of head pose images and fifty densely crowd images containing over 64000 human head annotations, with counts varying in between 94 and 4543. Our dataset contains jammed crowd scenes with individuals in large numbers, as opposed to other datasets, which encompass very few individuals. The experimental results establish the robustness and viability of the proposed approach by careful assessment of the counting process in terms of Absolute and Normalized Absolute Errors. We also developed a prototype for evaluating the accuracy of proposed system. © 2016 IEEE.