| Human action recognition is one of the most attractive research areas in computer vision withmany promising applications, such as human-computer interaction and intelligent surveillance.Recent progress in this field points towards the use of Spatio-Temporal Interest Points (STIP) forlocal descriptor-based recognition strategies. The method is robust to partial occlusion and differentviewpoints, since STIP descriptors are local features. This thesis addresses the task of human actionrecognition under moving background based on Spatio-Temporal Interest Points. The main worksare listed as follows:A new method for region of interest (ROI) extraction is proposed. The ROI consists of thedivergence field, vorticity field, motion boundary of the optical flow in space domain, and wherethe flow velocity changes in time domain. Those areas possess a clear physical meaning and reflectthe motion of foreground. Furthermore, in order to enhance the robustness, silhouette is used as asupplement of the ROI, and we make control of their impact through the threshold. Carried STIPdetection in the ROI can effectively suppress the influence of moving background.We propose a point trajectories-based multi-feature fusion method, including point trajectories/HOG/HOF/MBH descriptor. Those descriptors represent the static structure information, motioninformation and relative motion information around the neighborhood of interest points, thus have agood discriminative power in moving background.Bag of words (BoW) model is used to represent action videos. The vocabulary is constructed byK-means clustering first, which is based on the appearance similarity. Then, the optimal number ofvideo-word clusters is discovered by utilizing Maximization of Mutual Information, which is basedon the correlation of feature vectors and actions. After twice clustering, we have a more compactand discriminative code-book.Our experiments show that the proposed method achieves higher recognition accuracy undermoving background. |
PDF Full Text Request