Explainable Human Action Intelligence for Reliable and Transparent Video-based Recognition

The human action recognition discipline has been boosted by the swift developments in Artificial Intelligence to a substantial degree, making it a pillar of the current computer vision studies. The current research contends with essential needs in society, since it uses a self-acquisition paradigm in order to enhance the understanding of intricate physical exercise routines, hence, enabling independent studies. The modern deep-learning models are more likely to promote predictive performance at the loss of interpretability, reliability, and trustworthiness. Their shortcomings are a hindrance in real world application where accountability and open-mindedness are essential. Explainable Human Action Intelligence (EHAI) framework, which seamlessly integrates state-of-the-art deep-feature extraction with thoroughly based explainability modules, which will contribute to putting trust in artificial intelligence. The proposed framework focuses on improving both the recognition accuracy and interpretability of human action recognition. ResNet-50 is refined using the UCF101 dataset, which comprises 101 human action categories. Gradient weighted Class Activation Mapping approach (Grad-CAM) is incorporated to localize and visualize salient regions within video frames, providing insight into the visual evidence influencing the model’s predictions. Transfer learning is employed to adapt ResNet-50 to the target action recognition task, resulting in an accuracy of 95.4%. The explainable AI methods allow generating visual explanations that underline the areas that are the most appropriate to the model predictions and promote transparency, reliability, and trust. As a result, the state-of-the art framework is not only better in current scans of human action recognition, but also improves model accountability significantly, which additionally promotes the clarity of AI-based resolutions.