This repository contains all needed data sets for the contribution in Journal of Sensors and Sensor Systems "Enhancing human–robot collaboration with thermal images and deep neural networks: the unique thermal industrial dataset WLRI-HRC and evaluation of convolutional neural networks". You may use this data for scientific, non-commercial purposes, provided that you give credit to the owners when publishing any work based on this data. DOI: 10.5194/jsss-14-37-2025 or as BibTex: @article{sume_enhancing_2025, title = {Enhancing human–robot collaboration with thermal images and deep neural networks: the unique thermal industrial dataset {WLRI}-{HRC} and evaluation of convolutional neural networks}, volume = {14}, issn = {2194-8771}, shorttitle = {Enhancing human–robot collaboration with thermal images and deep neural networks}, url = {https://jsss.copernicus.org/articles/14/37/2025/}, doi = {10.5194/jsss-14-37-2025}, abstract = {This contribution introduces the use of convolutional neural networks to detect humans and collaborative robots (cobots) in human–robot collaboration (HRC) workspaces based on their thermal radiation fingerprint. The unique data acquisition includes an infrared camera, two cobots, and up to two persons walking and interacting with the cobots in real industrial settings. The dataset also includes different thermal distortions from other heat sources. In contrast to data from the public environment, this data collection addresses the challenges of indoor manufacturing, such as heat distortions from the environment, and allows for it to be applicable in indoor manufacturing. The Work-Life Robotics Institute HRC (WLRI-HRC) dataset contains 6485 images with over 20 000 instances to detect. In this research, the dataset is evaluated for implementation by different convolutional neural networks: first, one-stage methods, i.e., You Only Look Once (YOLO v5, v8, v9 and v10) in different model sizes and, secondly, two-stage methods with Faster R-CNN with three variants of backbone structures (ResNet18, ResNet50 and VGG16). The results indicate promising results with the best mean average precision at an intersection over union (IoU) of 50 (mAP50) value achieved by YOLOv9s (99.4 \%), the best mAP50-95 value achieved by YOLOv9s and YOLOv8m (90.2 \%), and the fastest prediction time of 2.2 ms achieved by the YOLOv10n model. Further differences in detection precision and time between the one-stage and multi-stage methods are discussed. Finally, this paper examines the possibility of the Clever Hans phenomenon to verify the validity of the training data and the models’ prediction capabilities.}, language = {English}, number = {1}, journal = {Journal of Sensors and Sensor Systems}, author = {Süme, Sinan and Ponomarjova, Katrin-Misel and Wendt, Thomas M. and Rupitsch, Stefan J.}, month = feb, year = {2025}, note = {Publisher: Copernicus GmbH}, pages = {37--46}, }