Citation

Yuan B, Wang Y, Wang X. Int. J. Model. Identif. Control 2023; 43(2): 145-153.

Copyright

(Copyright © 2023, Inderscience)

DOI

10.1504/IJMIC.2023.132608

PMID

unavailable

Abstract

In order to solve the problems of image blur, uneven illumination and object occlusion in visual monitoring, a human fall detection algorithm based on visible light camera and thermal imager is proposed in this paper. Firstly, the visible light and thermal images are denoised to reduce the interference of noise. Secondly, the skeleton and joint coordinates of the human body are extracted through the lightweight human posture recognition model. Finally, three human posture parameters are designed as recognition features to achieve accurate fall recognition. The method is verified on self-built datasets and public datasets. The experimental results show that the accuracy of the method is 0.93 and 0.94 respectively. Compared with the most advanced algorithms, the proposed method has higher accuracy and better real-time performance.

Language: en

Keywords

computer vision; deep neural network; fall detection; multi-source image fusion