Human Activity Recognition Based on 2d Texture Signal Pattern Analysis

Human activity recognition is an important research area of computer vision which dictates the need to automatically detect and retrieve semantic events in videos based on video contents. In this paper, we attempt to extract the foreground object from the video clip using color model and generate a unique signal pattern for the detected foreground (human). Signal pattern is generated for the extracted 2D texture features and the most significant features are selected using feature selection method. For each detected object, we can study its corresponding motion pattern, entry/exist points, and behavior patterns. Based on this information, it is efficient to improve the object detection and track the abnormal event occurrence. Experiments were performed on KTH dataset, High-Level Human interaction dataset and real time video dataset. The empirical results show that 85% of accuracy based on precision/recall measure was obtained, and the ability to recognize the activities in real time shows the promise for applied use.