Crowd density estimation in extremely crowded images

In a variety of situations that involve a large number of people, there have always been huge difficulties in identifying an estimation of the number of people involved for the purpose of crowd management and the like. Crowd Management includes the assurance of the safety and security of the crowd, the deployment of law enforcement personnel, and the observance or detection of unusual crowd behavior. With this, many concerned organizations have made used of various manual implementations to accomplish the task of estimating crowd density. However, these implementations sometimes yield to inaccurate results and requires a significantly huge amount of time, effort and manpower. That is why, new implementations and approaches were used. These implementations or approaches take advantage of the technological advancements in the field of Computer Vision. In line with that, we aim to develop a simple crowd density estimation tool for extremely crowded images that makes use of different Computer Vision techniques like Image Processing and Image Segmentation. For this research, we are using an external computer vision tool called MATLAB to implement the simple workflow for the proposed crowd density estimation tool. We have tested our tool on four experiments: (1) whether to include median filter or not in the workflow, (2) which perspective yielded the count nearest to the ground truth, (3) which configuration of the boundary detection algorithm is more appropriate for the tool, and (4) to test the tool on images that needs intricate masks in order to isolate the regions of interest. Our results in these experiments show that the application of a median filter and a bird\u27s eye view or front view perspective with face shots are the ideal perspectives in which the tool can give an estimate nearest the ground truth with only a difference of less than a hundred. © 2015, Mechatronics and Machine Vision in Practice. All rights reserved