Multiview Human Action Recognition System

This paper presents an integrated system for human action recognition in videos using computer vision techniques. The system includes stages such as video acquisition, feature extraction, and action recognition. In the feature extraction stage, it performs moving object localization and extraction of points of interest. Moving object localization uses a Gaussian Mixture Model (GMM) for background subtraction and a Prewitt filter for silhouette extraction. Points of interest are extracted using a Gabor filter and temporal differencing. The final stage involves feature reduction and classification using three classifiers: nearest neighbor, nearest mean, and Gaussian mixture model. The system achieves recognition rates of 90.83 % for the nearest mean classifier, 81.34 % for the Gaussian mixture model, and 78.98 % for the nearest neighbor classifier. The results demonstrate the effectiveness of the approach. The system is evaluated on the KTH dataset and shows potential for real-world applications, successfully detecting various human actions in videos.