Human activity recognition based on the blob features

In this paper, we present a novel approach for human activities recognition in the video. We analyze human activities in the sequential frames because human activities can be considered as a temporal object which contains a series of frames. Firstly, we establish a statistical background model and extract foreground object through background subtraction in the video stream. Then, we use foreground blobs of the current frame and a series of frames before the current frame to form a new feature image in certain rules. Finally, we combine the non-zero pixels in the feature image into blobs using the connected component method. Then each blob corresponds to an activity which is characterized by the blob appearance. By recognizing blob features we can recognize activities. We use Gaussian mixture to model features for each type of human activities and employ Mahalanobis distance to measure the similarity.