HAC-POCD: Hardware-Aware Compressed Activity Monitoring and Fall Detector Edge POC Devices
Hasib-Al Rashid, Tinoosh Mohsenin · 2023
Edge Point of Care (POC) devices are crucial for human activity recognition (HAR) and fall detection because they enable real-time analysis and fast intervention, which can greatly improve outcomes in situations of patient and elderly monitoring. The emergence of Artificial Intelligence (AI) has sparked renewed enthusiasm for integrating AI algorithms into low-power embedded systems, broadening the potential applications of the POC devices. This paper introduces HAC-POCD, a system for multimodal human activity recognition and fall detection that processes different modalities of complementary images, designs deep neural network (DNN) models, and employs model compression techniques including knowledge distillation and low bit-width quantization with memory-aware considerations to fit models within lower memory hierarchy levels, reducing latency and enhancing energy efficiency on resource-constrained edge devices. With compact inference model of 58 KB, we achieved 95.6% accuracy on HAR case-study. Our compact inference model, deployed on resource constrained hardware, GAPuino and Raspberry Pi 4, demonstrated low latencies within milliseconds and very high energy efficiency.