Deep Learning for Human Activity Recognition: Trends, Challenges, and Edge-Centric Deployments
Divesh Kumar, M. Y. Mohamed Parvees · 2025
Human Activity Recognition (HAR) has become pivotal in various domains such as healthcare, surveillance, and human-computer interaction. This paper presents a robust framework that combines advanced deep learning models with edge computing capabilities to enable real-time, accurate, and efficient activity recognition. A hybrid deep neural architecture, comprising Convolutional Neural Networks (CNN) for spatial feature extraction and Bidirectional Long Short-Term Memory (BiLSTM) networks for temporal sequence modeling, is deployed on lightweight edge devices to ensure low-latency performance. The system is optimized for dynamic environments with variable sensor inputs, utilizing a custom pruning and quantization strategy to reduce computational overhead without compromising accuracy. Comprehensive experiments on benchmark datasets demonstrate superior performance in recognition accuracy, inference speed, and energy efficiency compared to existing methods. The proposed edge-deployable HAR solution offers a scalable and practical approach for real-world applications requiring timely and privacy-preserving activity analysis.