An Efficient Crowd Estimation Method Using Convolutional Neural Network with Thermal Images
Muhua Xu · 2019 IEEE International Conference on Signal, Information and Data Processing (ICSIDP) · 2019
Crowd estimation, or the task of determining the number of people in a certain area with a high density, has recently regained much attention as the development of Convolutional Neural Network (CNN) based methods. Different from other common techniques that merely estimate the number of people under relatively simple scene, crowd estimation by CNN has capability to handle the high density of people in the image. However, most of the CNN-based crowd estimation methods devote to analyze visible image, which typically results in privacy invasion for the common public. To address these problems, this paper proposes a crowd estimation algorithm that only uses thermal image as input. The proposed method is based on the fact that there typically exists significant temperature difference between the human face and background, which could be easily distinguished with luminance of raw imagery. Additionally, the utilization of thermal images can easily avoid background interface like counting pictures of faces on clothes or posters, which simplified the model complexity in data fitting. With a lightweight CNN framework, the presented method could obtain high accuracy, low computational cost, and privacy protection. Detailed experiments and a base-line image processing mechanism are implemented to demonstrate the effectiveness and accuracy of the proposed method.