Crowd Estimation via Thermal Images Analyzed by Convolutional Neural Network

Abstract

This research focuses on improving the method of crowd estimation by using a single thermal picture as input. To be more specific, the purpose of this research is to combine state-of-the-art computer vision techniques and convolutional neural network on thermal image inputs to tackle the problem of estimating crowd density. Current crowd estimation methods are either too computationally expensive and privacy-invading like Switch Convolutional Neural Network or too complicated, needing multiple cameras or sensors. Different from preexisting methods, our approach focuses on developing different methods to estimate crowd density that utilized a single thermal image and the knowledge of the environment setting as the only input. Dataset used to train are collected during school events and are accurately labeled. The result shows that our method of using a single thermal camera in a given space to estimate crowd density is a promising approach. This approach offers high accuracy, low computational cost, and privacy protection. Additionally, using thermal images as input can easily avoid errors like counting pictures of faces on clothes or posters. Future improvements for this research include extending the data collection and building an online public database. The neural network will also be fine-tuned, and other state-of-the-art methods will be implemented. Future applications can include a low-cost approach in event reporting and security services in public places like subways and airports.