Human Activity Recognition at the Edge: A Systematic Review of Machine Learning Models, Deployment Constraints, and Latency Energy Accuracy Trade offs
TL;DR
The findings show that accelerometers, gyroscopes, inertial measurement units, smartphones, smartwatches, and wearable sensors are the dominant data sources for HAR at the edge and a gradual shift from accuracy-centered evaluation toward deployment-aware assessment, incorporating latency, inference time, energy consumption, power usage, model size, and memory footprint.
Abstract
: Human Activity Recognition (HAR) has become a key component of intelligent healthcare monitoring, assisted living, sports analytics, smart environments, and wearable computing. However, the migration of HAR models from cloud-centered architectures to edge, embedded, mobile, and wearable platforms introduces constraints related to computational capacity, memory footprint, battery consumption, inference time, and real-time responsiveness. This study presents a systematic literature review of edge-based HAR research, focusing on machine learning models, deployment constraints, and latency energy accuracy trade-offs. Following a PRISMA-based process, 78 studies published between 2021 and 2026 were analyzed from Scopus, Web of Science, IEEE Xplore, and DBLP. The findings show that accelerometers, gyroscopes, inertial measurement units, smartphones, smartwatches, and wearable sensors are the dominant data sources for HAR at the edge. CNN, LSTM, CNN-LSTM, Random Forest, Support Vector Machines, Transformer-based models, and TinyML approaches are frequently used for recognizing daily activities, clinical movements, sports actions, and context-aware behaviors. The results also indicate a gradual shift from accuracy-centered evaluation toward deployment-aware assessment, incorporating latency, inference time, energy consumption, power usage, model size, and memory footprint. Model compression, quantization, pruning, lightweight architectures, feature selection, and microcontroller-based deployment emerge as central strategies for improving edge performance. Nevertheless, the evidence remains fragmented because many studies do not report comparable hardware-level measurements. This review contributes a structured synthesis of edge-based HAR and identifies open challenges related to reproducibility, benchmark standardization, real-world validation, cross-device generalization, and balanced evaluation of latency energy accuracy trade-offs.