Action Recognition in Videos Using Block-Based CNN Method

Abstract

Human Action Recognition is an important research area in human behaviour analysis. A reliable action recognition system has many important applications in various fields of computer vision including robotics and human-computer interaction. While researchers have performed analysis of different human actions like dancing, clapping, handwaving, robust recognition remains a significant challenge for machine learning algorithms. Recognizing actions from images and videos is extremely difficult task due to a different kind of actions occurring under different backgrounds. Specific challenges include background clutter, occlusion, camera motion, variations in execution rate, and lighting conditions. To overcome such challenges, we have proposed block-based CNN methodology designed to recognize actions from video data effectively. Our approach extracts temporal features using a CNN model configured with specific temporal block sizes. The proposed model was trained and evaluated using two different block configurations (size 11 and size 5) across five different datasets: KTH, UCF101, UCF Sports Action, YouTube, and MUSTian Action dataset. Using an 80/20 train-test holdout ratio, our Model III achieved near-perfect accuracy, with training and validation performance reaching up to 100% on controlled benchmarks while mainlining competitive results on more complex, real-world datasets.