AHAR: Adaptive CNN for Energy-Efficient Human Activity Recognition in Low-Power Edge Devices

Human activity recognition (HAR) is one of the key applications health monitoring that requires continuous use wearable devices to track daily activities. This article proposes an adaptive convolutional neural network for energy-efficient HAR (AHAR) suitable low-power edge devices. Unlike traditional (early-exit) architecture makes early-exit decision based on classification confidence, AHAR a novel uses output block predictor select portion baseline during inference phase. The experimental results show suffer from performance loss whereas our provides similar or better as while being energy efficient. We validate methodology in classifying locomotion activities two data sets—1) Opportunity and 2) w-HAR. Compared fog/cloud computing approaches set, architectures comparable weighted F1 score 91.79%, 91.57%, respectively. For w-HAR outperform state-of-the-art works with 97.55%, 97.64%, Evaluation real hardware shows significantly efficient ( $422.38\times $ less) memory-efficient notation="LaTeX">$14.29\times compared set. notation="LaTeX">$2.04\times less notation="LaTeX">$2.18\times memory work. Moreover, 12.32% (Opportunity) 11.14% (w-HAR) than providing no significant overhead.