INTER PERSON ACTIVITY RECOGNITION USING RGB-D DATA

The large amount of video data from various sources like CCTV is widely available. Automatic analysis of video for scene understanding is essential and useful in many video surveillances, applications like anomaly detection, activity recognition, patent monitoring. In this paper, we have presented an Activity Recognition System in varying illumination. Proper segmentation and selection of features and classifiers are crucial in such applications. The depth information from the RGB-D sensor and the color cue is used to segment the person and the background. The use of depth information reduced the complexity and improved accuracy in the segmentation. We have used the novel motion feature along with the GEI of the silhouette and person skeletons for describing various activities. KNN, NN, Naive Bayes Classifier, and SVM are used for activity classification. The dataset used for experimentation is prepared with the help of 11 persons for 10 activities in four illumination conditions. Our study shows that the use of the depth information from Kinect sensor reduces the computational complexity in segmentation and motion feature improves the recognition rate.