Ultra-Lightweight Edge-AI Framework for Real-Time HAR Applications

Abstract

There is an increasing demand for lightweight Human Activity Recognition system on constrained devices due to recent advances in edge intelligence. Conventional deep learning models require high computation, memory, and energy for continuous inference. The development of a very light Edge-AI framework for UHAR on smartphone inertial sensor data the implementation targets a software level of feasibility before being deployed on edge wearable or embedded platforms. All these aspects are integrated into a single pipeline for superior processing efficiency. Classifying the activity is done through lightweight preprocessing and compact temporal representations of segmented sensor windows. The dataset for Human Activity Recognition with Smartphones was sourced from Kaggle and UCI repository. The accelerometer and gyroscope data of six activities of thirty volunteers are there in the dataset. The recognition accuracies produced by the proposed model and chosen baseline were 94.6% and 90.8% respectively. The lag time for inference per processed window dropped from 32.7 ms to 18.4 ms. The model sizes were reduced from 3.9 MB to 1.2 MB making it possible for compact storage. The software-based inference evaluation involved memory utilization, which was brought down from 14.5 MB to 6.8 MB. The energy-related indicator decreased from 9.6 to 4.1 mJ under the stated estimation assumptions. The findings show that the accuracy-efficiency balance of continuous HAR improved under CA conditions. Nonetheless, true validation of the embedded device remains essential before claiming readiness for deployment in practice.