Detecting Falls and Slips of Wheelchair Users Using Low-Resolution Thermal Image Analysis

As the aging population grows, there is increasing demand for privacy-preserving monitoring of accidents in private rooms, for which thermal cameras are well suited. This paper focuses specifically on wheelchair users who are more prone to falls. Falls can be classified into high-risk falls that cause injury and low-risk slips where users gradually slide out of the wheelchair. Treating both equally risks generating excessive warnings and missing true emergencies. In this study, we used low-resolution thermal images to detect and distinguish high-risk falls and low-risk slips, thereby reducing caregiver burden and improving safety. However, video analysis requires clear images of the limbs and other details of the body, but thermal cameras cannot consistently capture such detailed features. To overcome this, we proposed a video analysis scheme using torso features and developed a system called “Fall Detection using CNN and Torso Features (FDCTF).” Additionally, to enhance the generalizability of the system to various environmental conditions, we incorporated multi-task learning and a two-layer LSTM network, extending FDCTF. As a result, in a single-environment evaluation, our system achieved an F1 score of 1.0 for fall detection. Furthermore, in a multi-environment evaluation, the F1 score reached 0.82, demonstrating the model’s adaptability to various settings. Moreover, the system significantly reduced caregiver burden compared to conventional methods, with reductions of 0.19 in false alarms, and 0.18 in missed serious falls.