Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{185758,
        author = {Dr.MONIKA D.ROKADE and Haridas R. Bankar and Aditi A. Borkar and Saurabh Solankar},
        title = {Lightweight Deep Learning Model for Human Action Recognition in Videos},
        journal = {International Journal of Innovative Research in Technology},
        year = {2025},
        volume = {12},
        number = {5},
        pages = {3906-3911},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=185758},
        abstract = {— Human Action Recognition (HAR) is an important branch of computer vision involving the identification and interpretation of human activity in video (e.g., walking, running, or waving). While several deep learning methods such as 3D CNN, and I3D provide reasonable results, they require high computational power to train and are hardware demanding, making them unsuitable or impractical for real-time application, or mobile devices. 

The project will demonstrate a lightweight deep learning model for effective and accurate human action recognition on low-resource devices, or edge devices. The proposed framework represents a lightweight CNN architecture (e.g., MobileNetV2 and EfficientNet-Lite) for spatial feature extraction and temporal modeling techniques such as Bi-LSTM, or Temporal Convolution to capture motion across frames. Model optimization will also be applied based on techniques such as pruning and quantization aware training, in order to reduce model size and latency while preserving levels of accuracy.

The objective is to accomplish real-time action recognition with lowered computational cost for a variety of usages such as surveillance, healthcare, fitness tracking, and smart devices. The model is tested using well established datasets UCF101 and HMDB51, measuring accuracy, F1 score, model size, and processing 

speed. The results show that lightweight models can effectively balance both speed and accuracy and be more efficient for practical applications.},
        keywords = {Human Action Recognition (HAR), Deep Learning, Lightweight CNN, MobileNetV2, EfficientNet-Lite, Bi-LSTM, Temporal Convolution Network (TCN), Model Optimization, Pruning, Quantization, Edge Computing, Real-Time Video Analysis, Computer Vision, Surveillance, Smart Healthcare.},
        month = {October},
        }