透過您的圖書館登入
IP:34.231.247.254
  • 學位論文

不穩定更新率影片之物體長程追蹤

Long-term Object Tracking in Video with Unsteady Frame Rate

指導教授 : 陳永昇

摘要


在大型監控環境中,要建置多台攝影機才可以涵蓋整個範圍,但以傳統人工的方式來監視是沒有效率的,所以我們欲發展自動化的長程追蹤系統來增進效率。我們使用網路攝影機(IP camera)的原因是因為透過網路攝影機容易取得影像而且影像解析度高,相對的,使用網路攝影機可能造成影像不穩定更新率(unsteady frame rate)的問題。針對不穩定更新率所造成追蹤失敗,在本論文中,我們提出穩定且有效率的方法來解決。 系統主要分為兩個部份,第一部份為前景偵測,第二部份是物體追蹤。前景偵測主要使用密碼本(codebook)演算法,物體追蹤的方法使用改進均值位移(modified mean shift)演算法。基於目標模型的立方圖(histogram),改進均值位移一步步遞迴尋找下一個目標出現的位置,但不穩定更新率使得物體跳動太大讓改進均值位移追蹤失敗,因此,我們提出結合改進均值位移和移動重置(motion relocation)的方法,除了原始追蹤位置外,在物體出現率最高的前景區域重置另一個起始追蹤視窗,這兩位置分別套用改進均值位移找到收斂位置後,以跟目標模型的相似度來衡量哪個結果為追蹤目標。此外,我們也提出前景擴張(foreground extension)方法來解決物體因停留太久變成背景而追蹤失敗的問題。 本論文中,我們發展了即時追蹤系統,在不穩定更新率的情況下,仍然有穩定的追蹤結果,根據實驗結果顯示,我們提出的改進均值位移比原始均值位移(original mean shift)更據可靠性,而且使用移動重置也提昇了很高的準確率。

關鍵字

長程追蹤

並列摘要


In a large environment, to monitor all targets, we must deploy multiple cameras to cover the whole area. Because manually monitoring is inefficient, we want to develop a long-term object tracking system to improve the efficiency. We use IP cameras because the images are high-resolution and easy to obtain from network. On the other hand, IP cameras may cause unsteady frame rate video. Unsteady frame rate results in a sudden low frame rate and tracking failure. To solve this problem, we propose an efficient and robust method to enhance the long-term object tracking system in unsteady frame rate video. First, we extract the foreground objects using codebook algorithm and build reference histograms for the target model. Next, we use the modified mean shift algorithm to track the object in consecutive frames. Modified mean shift iteratively find next target location with a given target model and current starting region. However, using modified mean shift as progressive tracking is likely to fail when the movement of the object is unstable due to the unsteady frame rate. Thus, we propose the method of combining modified mean shift algorithm and motion relocation to deal with the problem from unsteady frame rate. We create another tracking window located at a foreground region with high consistency in term of object size and moving direction. Then we separately apply modified mean shift at the two window locations as starting points to search their converged points. Similarity measure is used to determine which one is the next target location. Furthermore, we propose the foreground extension method to improve the robustness of tracking when foreground objects become background ones if they remain still for a period of time. In this work, we have developed a real-time tracking system which remains robust in unsteady frame rate video. According to our experimental results, the modified mean shift is more reliable than original mean shift and tracking accuracy has been greatly improved by applying relocation method.

參考文獻


[10] Y. Changjiang, R. Duraiswami, and L. Davis, "Efficient mean-shift tracking via a new similarity measure," in Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on, 2005, pp. 176-183 vol. 1.
[1] D. Comaniciu and P. Meer, "Mean shift analysis and applications," in Computer Vision, 1999. The Proceedings of the Seventh IEEE International Conference on, 1999, pp. 1197-1203 vol.2.
[3] K. Toyama, J. Krumm, B. Brumitt, and B. Meyers, "Wallflower: principles and practice of background maintenance," in Computer Vision, 1999. The Proceedings of the Seventh IEEE International Conference on, 1999, pp. 255-261 vol.1.
[5] R. T. Collins, "Mean-shift blob tracking through scale space," in Computer Vision and Pattern Recognition, 2003. Proceedings. 2003 IEEE Computer Society Conference on, 2003, pp. II-234-40 vol.2.
[7] K. Nummiaro, E. Koller-Meier, and L. Van Gool, "An adaptive color-based particle filter," Image and Vision Computing, vol. 21, pp. 99-110, 2003.

延伸閱讀