Optimizing Human Action Recognition in Still Images Using Deep Learning Models and Grad-CAM++ for Visualization

Human action recognition(HAR) enables the detection and classification of human movements, supporting real-time applications such as surveillance, healthcare monitoring, gesture recognition, hazard avoidance, and sports analysis. However, HAR using still images is challenging due to the lack of temporal information, which makes capturing dynamic actions and distinguishing similar activities difficult. Despite these limitations, we successfully developed a robust HAR model using still images by carefully implementing deep learning techniques. We created an image classification model utilizing transfer learning with EfficientNetB7 to classify human actions into 15 categories based on still images. The model is trained and evaluated on a dataset containing over 12,600 labeled images covering various human activities. Various data augmentation techniques were employed to improve model generalization. A custom architecture was designed to optimize performance, achieving a peak training accuracy of 96.28%. A key contribution of our work is the application of Grad-CAM++. This powerful visualization tool provides interpretability by generating heat maps, indicating which parts of the image contributed the most to the model's predictions. This enhances model transparency, making the decision-making process more transparent. Our model delivers excellent results, showcasing its potential for application across various domains where human activity recognition is essential.