Human Activity Recognition (HAR) refers to the process of detecting, identifying, and classifying human activities from sensor data through the application of artificial intelligence techniques. Over the past few decades, HAR has emerged as a rapidly evolving research domain with significant implications in various fields, including video surveillance, identity authentication, smart home automation, healthcare monitoring, and human–computer interaction. In particular, within surveillance systems, timely and accurate recognition of human activities can serve as a preventive measure against incidents such as theft, vandalism, or other suspicious behaviors, thereby enhancing public safety. Among various AI-based approaches, deep learning models especially Convolutional Neural Networks (CNNs) have shown remarkable capabilities in automatically extracting high-level spatiotemporal features and achieving robust classification performance. However, CNN architectures often require extensive hyperparameter tuning to maximize their accuracy and efficiency. To address this challenge, the present study proposes an enhanced CNN model whose parameters are optimized using the Particle Swarm Optimization (PSO) algorithm. The PSO-driven optimization process aims to improve feature extraction quality, reduce overfitting, and enhance generalization capabilities. The proposed framework is implemented and experimentally evaluated on the Wiezmann dataset. Comparative analysis demonstrates that our approach achieves superior recognition accuracy and computational efficiency compared to several existing state-of-the-art methods.