Multiple Video Objects Detection and Classification for Human Activity Analysis
| Autor: | Po-Yen Lee, 李柏諺 |
|---|---|
| Rok vydání: | 2013 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 101 For a variety of applications such as video surveillance and event annotation, the spatial-temporal boundaries in video objects are required for annotating video content with high-level semantics. In this thesis, we present an integrated approach for the detection and recognition of multiple video objects in video clips and its application on human activity analysis. We extract video objects in first the frame using a learnt video object model. A model-based tracking algorithm is then used to search these video objects in successive frames in the video clip. We extract the spatial-temporal features in these tracked video objects. Given trained video clips in an activity class, based on the proposed video object detection, a video object is representation as a pose sequence. A set of class-specific pose codebooks are then trained by clustering patches’ feature, i.e., HOGs and motion vectors (MVs), extracted from video objects. These codebooks are then used to encode each frame in a video clip with multiple Bag-of words (Bow) histograms including HOG Bow histograms for basis objects and salient fragments. Motion vector codebooks are also used to characterize the dynamic property in each frame with MV Bow histograms for individual objects and object pairs. Finally, based on these Bow histograms and a string kernel, SVM classifier is trained to annotate the activity type of the input video clip. Experimental results show that the proposed method gives good performance on publicly available datasets in terms of detection accuracy and recognition rate. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|