Human presence detection based on gathered thermal datasets through deep learning

• Main Authors: Amoroso, Miguel Carlos Cariño, Atienza, Kenneth Roel Catapang, Ladera, Riff Kurtees Tan, Menodiado, Nico Angelo Mendoza
• Format: text
• Language: English
• Published: Animo Repository 2022
• Subjects: Infrared imaging; Electrical and Computer Engineering
• Online Access: https://animorepository.dlsu.edu.ph/etdb_ece/19; https://animorepository.dlsu.edu.ph/cgi/viewcontent.cgi?article=1020&context=etdb_ece
• Institution: De La Salle University

Human detection using deep learning is often done with colored images rather than thermal images. Lacking integration between existing studies regarding thermal datasets and deep learning networks for human detection can be bridged with a curated dataset and a modified network for the said purpose. With this, the research presents a new dataset and an accompanying network for human detection of elevated thermal images. The created dataset consists of 7101 thermal images of humans and takes into account different parameters such as height, light conditions, number of people, etc. The modified networks were created using either the layer removal method or the replacement of different parts of the network and their algorithms. The proponents trained and tested the created networks named TrimmedYOLO and YOLO-ReT-Mish on the created dataset and were able to achieve mean average precisions as high as 95.66% and 94.45% on Private-Private tests for TrimmedYOLO and Transformed Combined-Private tests for YOLO-ReT-Mish respectively while being more lightweight when compared to the original YOLOv4.