多相機影像監控系統之高效率多目標物一致性標籤

Abstract

近年來影像伺服於多相機系統的運用已隨著電腦運算的躍進而吸引大量研究者的目光。基於經濟層面考量，使用有限數量的相機來進行多人追蹤以及一致性標籤已逐漸成為一個重要的研究課題。於本論文中，我們提出了一個能穩健追蹤與即時辨認多人的影像伺服系統，該系統可以架設於一般的建築物結構中。相較於將所有即時影像集中到中央處理伺服器，本系統將追蹤以及影像分割等工作分散於每台獨立相機上面進行，僅將必要的資訊於必要的時刻送到中央處理伺服器進行相機目標物交互關係比對。本篇論文提出一套階層式交互比對多相機目標物的方法，並同時獲得目標物交互比對之信心指標。此信心指標將作為交互分析結果的正確率保證。在進行多相機目標物交互比對後可獲得環境中被追蹤者的人數以及每個人的外觀影像資訊，這些資訊將儲存於中央處理伺服器的目標物資料庫中，以便於判辨是否曾經有人重複出現。在不假設所有相機皆可觀察到共同地板平面的情況下，我們的階層式標籤系統仍可正確的進行目標物交互比對，並利用比對結果去驗證被追蹤者是否重複出現於監控環境中。被追蹤者的外觀資訊將被更新於資料庫中以及強化每台相機上獨立運行的追蹤系統。各相機的多目標物遮蔽處理能有效提升標籤系統的判斷精確度，減少因互相遮蔽所造成的外觀資訊擷取錯誤。經過實驗驗證，本系統即使在目標物互相重疊遮蔽的情況下仍可正確地進行一致性標籤的任務。於本論文最後，作者提出多份追蹤影片截圖並進行詳細的分析與討論。

Visual surveillance in multi-camera system has attracted more interest in recent years. Using limited number of cameras to simultaneously track and correctly label as many people as possible becomes an important topic of research, with low-cost consideration. In this thesis, we propose a surveillance system that can robustly track and identify multiple humans, for general building environments. Rather than gathering all information into a central server every frame, we track and segment each observation from local single camera, and only sending necessary information to the central server for correspondence processing at necessary time. Thus our framework can achieve observation correspondence between multi-cameras with confidence levels as correspondence quality indices. After correspondence process, the tracked object information is stored into the target databases for solving people re-entering problem. Without assuming common ground plane is observed by all cameras, our labeling process, which hierarchically associates objects after correspondence to target databases with matching confidence orders, still can construct relevant and accurate labeling assignment. The people information is then updated to improve target databases and local tracking performance. Occlusion handling for multi-object tracking can effectively enhance labeling accuracy and reduce the error of appearance information extraction due to object overlapping. The proposed labeling system yields robust performance even in most partial occlusion cases. Finally, we conclude with experimental results in several real video sequences and their detailed analysis.