ThermalSynth: A Novel Approach for Generating Synthetic Thermal Human Scenarios

In this paper, we propose ThermalSynth, a novel approach for creating synthetic thermal images by mixing 3D characters generated using the Unity game engine with real thermal backgrounds. We use a shader based on the Stefan-Boltzmann law to approximate the appearance in the thermal domain of the synthetic characters. Additionally, we provide a post-processing pipeline to better blend the high-fidelity synthetic data with the lower-resolution real thermal surveillance one. The proposed approach is used to create a dataset for people falling into water near a harbor front. Diverse scenarios of such falls are generated with an ample amount of data to enable the use of deep learning algorithms. To demonstrate the effectiveness of the generated data, we train two standard deep neural networks (AlexNet and ResNet-18) on our synthetic thermal dataset using a supervised learning approach. We test our system on small datasets containing real video footage of actual falls. We observe that training these simple classification networks yields an accuracy of 98.70% at a sensitivity of 100% on the real-world voluntary fall dataset.