Event based acceleration measurement and fall detection
Niklas Huhs, Jens Kraitl, Christoph Hornberger, Olaf Simanski · Current Directions in Biomedical Engineering · 2025
Abstract Fall detection is essential for elderly and disabled individuals, as undetected falls can be life-threatening. Traditional methods use acceleration sensors and neural networks, but body-worn sensors can be intrusive. This study explores neuromorphic cameras for fall detection with minimal data processing and lightweight neural networks. Denoising techniques were applied to event data, followed by statistical analysis to estimate position, velocity, and acceleration. This method produced patterns similar to accelerometers. Neural network architectures were evaluated, from simple one-dimensional convolutional networks to hybrid models combining convolutional layers with Long-Short-Term-Memory units. Training data were generated by converting video-based fall datasets (le2i, MCFD) into event data using the v2etoolbox. Data augmentation resulted in 2,610 samples (1,314 falls, 1,296 daily activities). The best model, a threelayer 1D convolution combined with a two-layer LSTM (hidden size 64, 125k trainable parameters), achieved 97% accuracy. Live inference on streamed videos and a DVXplorer event camera was possible without noticeable delay. Our approach matches state-of-the-art acceleration sensor methods while offering a non-intrusive, real-time monitoring solution, potentially improving response times and user comfort.