COMPARATIVE STUDY OF DEEP LEARNING MODELS FOR HUMAN ACTIVITY RECOGNITION

Human Activity Recognition (HAR) using wearable sensor data is a cornerstone of mobile health and context- aware computing. While deep learning has significantly advanced HAR accuracy the computational demands of complex architectures often conflict with the limited resources of edge devices like smartphones and wearables . This creates a critical trade off between predictive performance and practical deployability. This paper presents a systematic comparative analysis of five distinct deep learning architectures: a baseline Multi-Layer Perceptron (MLP), a 1D Convolutional Neural Network (1D-CNN), a Long Short-Term Memory (LSTM) network, a hybrid CNN-LSTM model , and a Transformer-based model . We evaluate these models on the public UCI-HAR dataset, focusing not only on classification accuracy and F1-score but also on crucial efficiency metrics: model size (parameters) and inference latency. Our findings reveal that while the Transformer achieves the highest F1- score (0.931), its substantial computational cost makes it less suitable for real-time edge applications. The hybrid CNN-LSTM architecture emerges as the most balanced solution, delivering competitive accuracy (0.925 F1-score) with significantly lower latency and a more compact model size. This study provides a clear, data- driven framework for selecting appropriate HAR models based on specific deployment constraints.