Thermal Imaging Dataset for Human Presence Detection
Jose Martin Z. Maningo, Miguel Carlos C. Amoroso, Kenneth Roel Atienza, Riff Kurtees Ladera, Nico M. Menodiado, Leonard U. Ambata, Melvin K. Cabatuan, Edwin Sybingco, Argel Alejandro Bandala, Jason Espanola, Ryan Rhay P. Vicerra · 2023
In this paper, a created elevated thermal dataset that is designed for training deep learning models for human presence detection will be presented. The dataset consists of 7,101 elevated thermal images of humans captured in different scenarios while walking, running, and clustering. The recordings are captured using Seek Thermal Compact Pro camera under different environments and light conditions, different body positions, and distances from the camera and were preprocessed for the algorithm. The dataset was evaluated using YOLOv7-Tiny regarding performance in mAP scores and loss charts generated when training and testing the captured images. The robustness of the dataset was tested using cross-validation and reducing the number of datasets. The dataset was also combined with other publicly available datasets to compare the difference between the training of the created dataset alone. Combining the private dataset with other public datasets upon training significantly increased the mAP score in detecting the images belonging to the testing datasets that contained publicly available images.