An Approach for Crowd Counting and Crowd Density Estimation using Aerial Images

Abstract

When a large number of people gather in a single area, it could lead to a mass stampede which could result in major injuries and deaths. It is important to measure the number of people in a gathering therefore pre-security measures can be taken to avoid such stampedes from rising. High beam cameras or drones can be used to capture images in a large crowded area without any further human involvement. In the present, most of the crowd counting techniques use face and head detection on finding humans in a provided image. One major disadvantage of these approaches is occlusion occurring if objects are near in the images. The goal of this thesis is to perform a bounding box regression algorithm on top of a convolutional neural network based model to detect and count the crowd for a provided input image. The convolutional neural network is trained for head detection and the bounding box algorithm is used to draw bounding boxes on the areas which are detected by the network as heads. Box count is taken as the crowd count. The count is validated against the ground truth count provided by the testing dataset in Classification Accuracy and Mean Approximation Error. The model is implemented using Keras framework, an open source machine library in python. The solution is trained using the dataset “Crowd Detection Model” and has shown a mean approximation error of 0.35 and a classification accuracy of 0.65.