隨著攝影機普及化，影像資料量大幅提升至無法逐一人工處理的程度。為了自動化的偵測、切割、追蹤物體產生許多演算法，目前最基礎的物體運動檢測演算法是直接比較當前場景與靜態背景的差異來檢測區域。然而此演算法會因為場景光源的變化而產生誤判，為了解決誤判問題需要消耗大量的運算量，故無法低階的運算硬體達到良好的運算效率。
本論文使用單一攝影機搭配低階開發版，藉由圖形處理單元(GPU)進行平行處理，使動態物體的邊界偵測達到即時運算之效果。利用單一影像即可對樣本模型進行初始化。在執行過程中使用當前影像需持續更新初始模型以變達到有效的動態物體偵測，再將動態區域利用連通集標籤進行邊界的計算。於IGS開發板上的執行動態物體偵測在未使用GPU時加速執行時間為14FPS，動態偵測使用GPU加速之執行時間為32.6FPS；加入邊界計算執行時間為27FPS。

As mobile cameras become widely available to the mass,
the amount of video data created with them has reached a point where manually editing each video is impractical. In that wake, many algorithms for automatic object detection, segmentation and tracking were proposed.
As of now, the fundamental algorithm for object tracking uses the difference between current scene and static scene to detect object.
However, the algorithm can fail due to the difference in lighting on the two scene. And to resolve the problem would require a massive computation time,resulting in poor performance of the algorithm on low-end hardware.
In this paper, we propose an algorithm that can detect the boundary of a dynamic object with a single camera. We utilized graphics processing units to improve the algorithm's performance and achieved real-time efficiency on low-end developer hardware. Our sample model is initialized with a single image and constantly updated with new images of current scene during execution to successfully track dynamic objects. And finally, we detect the bounding box of segmented objects with connected component labeling. The performance of dynamic object detection on IGS's developer hardware without GPU acceleration is 14FPS.
After GPU acceleration the performance is improved to 32.6FPS, and 27FPS when boundary calculation is included.

[1] Barnich O., EVS Broadcast Equip., Seraing Belgium , Van Droogenbroeck M. Vibe:A universal background subtraction algorithm for video sequences. IEEE Transactions on Image Processing, pages 1709--1724, June 2011.
[2] G. Gkioxari, P. Arbeláez, L. Bourdev, and J. Malik. Articulated pose estimation using discriminative armlet classifiers. IEEE Computer Vision and Pattern Recognition, pages 3342--3349, 2013.
[3] J. Shotton, A. Fitzgibbon, M. Cook, T. Sharp, M. Finocchio, R. Moore, A. Kipman,and A. Blake. Real-time human pose recognition in parts from single depth images. Communications of the ACM, pages 116--124, January 2013.
[4] Li Liyuan and Huang Weimin and Gu Irene Yu-Hua and Tian Qi. Statistical modeling of complex backgrounds for foreground object detection. IEEE Transactions on Image Processing, pages 1459--1472, November 2004.
[5] M. Sun, M. Telaprolu, H. Lee, and S. Savarese. An efficient branch-and-bound algorithm for optimal human pose estimation. IEEE Computer Vision and Pattern Recognition, page 1616–1623, 2012.
[6] R. Ronfard, C. Schmid, and B. Triggs. Learning to parse pictures of people. In European Conference on Computer Vision, pages 700--714.
[7] Sean M. O’Connell. a gpu implementation of connected component labeling. Illinois State University, 2009.
[8] Y. Yang and D. Ramanan. Articulated human detection with flexible mixtures of parts. IEEE Transactions on Pattern Analysis and Machine Intelligence, (December):2878-2890, 2013.