Abstract: Multi-camera, Multi-person, and Real-time Fall Detection using Long Short Term Memory
Christian Heinrich, Samad Koita, Mohammad Taufeeque, Nicolai Spicher, Thomas M. Deserno
Technische Universität Braunschweig, Peter L. Reichertz Institute for Medical Informatics
Falls occurring at home are a high risk for elderly living alone. Several sensorbased methods for detecting falls exist and - in majority - make use of wearable or ambient sensors. The field of video-based fall detection is emerging; however, restricted view of single cameras, distinguishing/tracking of persons, and high false-positive rates pose limitations. Hence, a novel approach for video-based fall detection was proposed : The human pose estimation algorithm openpifpaf was augmented for fall detection by adding multi-camera and multi-person tracking support. For each person, five temporal and spatial features are extracted and processed by a long short-term memory (LSTM) network, classifying each frame as a fall or no fall event. The UP-fall detection dataset was used for evaluation, achieving a F1-score of 92.5%. Still, a trend of false-positives was observed due to imbalance as the dataset contains 36% videos showing falls; contrary to their rare occurrences in real life. We aim to improve this work by integrating it in a smart home laboratory: We will acquire a multi-camera dataset representing falls and everyday activities extending the binary to multi-class classification.
1. Taufeeque M, Koita S, Spicher N, et al. Multi-camera, multi-person, and real-time fall detection using long short term memory. Proc SPIE. 2021;(accepted).