Deep Learning for Human Activity Recognition: A Comprehensive Review of Architectures, Performance, and Challenges Across Five Sensory Datasets

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Human activity recognition (HAR) using sensor data allows the automatic detection of human behavior and actions in everyday environments. The development of scalable and privacy-preserving HAR systems is supported by the nonintrusive collection of time-series data using wearable devices and smartphone sensors. Advances in deep learning have further improved recognition accuracy, expanding the applications of HAR across domains, including healthcare and human-computer interaction. Unlike existing surveys that primarily focus on architectural advancements, this study offers the first systematic review that places benchmark datasets at the forefront. This study provides a comprehensive examination of various deep models used in HAR, assessing their relative advantages, limitations, and optimization strategies based on an extensive review of 226 studies. It also delves into the core aspects of HAR system design, including data preprocessing, feature extraction, evaluation metrics, and validation approaches. Through a systematic examination of 160 studies, we analyzed HAR systems built using deep models in terms of architecture, performance, and implementation challenges on five widely adopted benchmark datasets (USC-HAD, PAMAP2, WISDM, UCI-HAR, and OPPORTUNITY). To ensure a balanced comparison across the benchmarks, we analyzed 32 studies for each of the five datasets. Our analysis reveals that hybrid models, particularly CNN-LSTM architectures, consistently achieve superior accuracy across benchmarks. Furthermore, the Adam optimizer and leave-one-subject-out (LOSO) validation protocols are identified as the most effective choices for robustness. We conclude by identifying key obstacles, such as sensor noise, intersubject variability, and real-time processing constraints, and suggest future research directions to improve the robustness, generalizability, and efficiency of HAR systems.

Links and sources

Comments