Optimized CNN-Based Fall Detection for Elderly Using Wearable Devices on Resource-Constrained Devices

Fall accidents are a leading cause of injury among elderly individuals, often exacerbated by medical conditions such as Parkinson's disease, insomnia, and sedation, as well as factors like muscle weakness, vitamin deficiencies, high blood pressure, and elevated sugar levels. Continuous monitoring of health conditions can enable real-time fall detection, helping to prevent severe injuries and ensure timely medical intervention. This study focuses on analyzing accelerometer data collected from wearable devices placed on various body parts of elderly individuals. The data is used to train and test a convolutional neural network (CNN) for fall detection. To optimize the model for deployment on resource-constrained devices, such as wearable sensors and embedded systems, we applied optimization techniques like quantization and pruning. Six different models were developed, with a systematic approach that involved the base CNN model, followed by TensorFlow Lite conversion, quantization, pruning, and combinations of these techniques. The final model, which applied both quantization and pruning, obtained the most significant size reduction while maintaining high accuracy compared to the original model. Specifically, the model size was reduced by 90 %, making it well-suited for realtime applications on edge devices. Our results demonstrate that these optimization techniques can improve both the efficiency and performance of fall detection systems.