Human Activity Recognition Using Deep Learning and GUI-Based Prediction Tool

Human Activity Recognition (HAR) has emerged as a prominent research area within computer vision and artificial intelligence, driven by its applications in healthcare, surveillance, human–computer interaction, rehabilitation, and sports analytics. Traditional approaches to HAR predominantly rely on video-based recognition systems that depend on temporal information across continuous frames. However, these approaches present several limitations, including high computational costs, dependency on advanced hardware, and challenges in accurately classifying activities from static images. In this project, a novel deep learning–based framework is proposed for recognizing human activities from still images, eliminating the reliance on temporal data. The system integrates an EfficientNet-based preprocessing pipeline with a Transformer architecture to capture contextual dependencies and enhance classification accuracy. To improve accessibility and usability, a graphical user interface (GUI) built with PyQt5 is developed, enabling real-time predictions and visualization of model confidence through bar charts. The methodology encompasses dataset curation, preprocessing, model training, and deployment, with the trained model stored in a portable .keras format for reuse in similar applications. Experimental evaluation demonstrates robust performance, portability, and ease of use, thereby bridging the gap between complex deep learning systems and real-world applications. Furthermore, the system lays the foundation for future enhancements such as webcam integration and video-based recognition, ensuring scalability and adaptability across diverse environments.