| Summary: | 碩士 === 國立臺灣科技大學 === 電子工程系 === 97 === In recent years, the growth of world’s aging population make more elderly people living alone, and elderly care becomes a serious problem. Hence developing an intelligent video surveillance system, which can detect fall incidence or human behaviors, becomes a hot research topic. In this research, we implement a skeleton-based fall detection system for the compressed video. Main issues in this research include image reconstruction from the compressed video domain, multiple objects tracking with occlusion handling, and fall detection. First, we reconstruct pixel values by using DC+2AC values in the compressed video. Then, we use Bayesian classification to discriminate foreground object and background. To enhance the spatial color information, we use triangular geometric histogram to measure the similarity in object tracking. Simultaneously, we convert the multiple objects tracking problem to the problem of finding maximum weight matching on a bi-partite graph, and we use the Hungarian algorithm to solve this problem. Finally, we combine the skeleton analysis, the ellipse of human body, and the change ratio of human shape to detect the fall incident. To verify the performance of the fall detection system, we perform intensive experiments based on videos. The experiment results reveal that the proposed fall detection system can achieve high detection rate and low false positive.
|
|---|
