請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92673
標題: | 用於多相機追蹤之資料集與分散式即時系統 Multi-Target Multi-Camera Tracking Dataset and Distributed Real-Time System |
作者: | 陳昱愷 Yu-Kai Chen |
指導教授: | 簡韶逸 Shao-Yi Chien |
關鍵字: | 多相機追蹤,多相機追蹤資料集,分散式多相機系統,即時多相機系統,多相機追蹤系統, Multi-Target Multi-Camera Tracking,Multi-Target Multi-Camera Tracking Dataset,Distributed Multi-Camera System,Real-Time Multi-Camera System,Multi-Target Multi-Camera Tracking System, |
出版年 : | 2024 |
學位: | 碩士 |
摘要: | 近年來,攝影機被廣泛的應用於大規模的監控系統中,提供了像是路徑規劃和行為分析等功能,其中多相機多目標追蹤演算法扮演著重要角色並逐漸受到各界的關注。然而,大部分現有的資料集因為資源考量而規模受到限制,其他的則是僅適用於特定目的,因此這領域長期缺乏適當的資料集而研究發展緩慢。
在本篇論文中,我們提出了一個新的人物資料集名為NTU-MTMC。此資料集具有大規模、真實戶外環境及非演員等特點,同時也包含許多具有挑戰性的場景,像是頻繁的遮擋、劇烈的光影變化和雜亂的背景。特別的是,此資料集中出現行人和自行車手兩種追蹤目標,他們截然不同的移動模式增加了跨相機追蹤的難度。此外,我們提出了一個基礎的線外系統架構搭配上我們設計的兩階段匹配策略及自適應時間函數。最後,我們透過實驗證明此資料集對相關領域有正向的影響,也驗證了我們提出的架構的可行性及效能。 此外,我們還建立了可以高效處理多相機資料的線上系統。與其他傳統系統相比,我們採用分散式架構,由多個中端裝置組成系統。每個中端裝置會同時處理多個相機中的資料並且只傳輸必要資料給彼此來獲得全區的資訊,這種架構可以節省傳輸頻寬並提高系統的靈活性,同時規模也不再受限於單一伺服器的能力。最後,我們透過模擬結果證明了這種分散式系統架構的在實際應用中的可行性,同時也展示出我們系統的準確性及計算效率。 Camera sensors are widely utilized around the world and integrated into large-scale and complex surveillance systems, empowering individuals to manage the information for diverse applications, such as route planning and behavior analysis. Multi-target multi-camera tracking (MTMCT) plays a pivotal role in these applications and has received increasing attention from both academia and industry. However, the development of person-based MTMCT grows slow due to the lack of proper datasets. The scale of most existing MTMCT datasets is limited by high data collection and annotation cost, while others are targeting to specific and limited purposes. In this thesis, we propose a new person-based dataset, NTU-MTMC, which contains several advantageous properties: large-scale, outdoor, real-world, and featuring non-acting individuals. NTU-MTMC consists of challenging scenes, which cause heavy occlusion, severe light variations and noisy backgrounds. Notably, different from other datasets, individuals in this dataset exist in two primary statuses: pedestrians and bicyclists, resulting in completely different transition time distributions and increasing the complexity in inter-camera association. On top of that, we propose a baseline framework for this dataset by adopting a two-stage matching strategy and adaptive time window function. Extensive experiments show the great impact of NTU-MTMC on related tasks and validate the effectiveness of ourproposed baseline framework. In addition, we present an online MTMCT system that can efficiently process real-world multi-camera data. Different from other conventional online frameworks which transmit all data to a central server, our system adopts a decentralized approach, consisting of multiple mid-end gateways. Each gateway concurrently processes several streaming cameras and their sub-tasks of MTMCT, and transmits processed data among themselves. Such distributed structure helps to save communication bandwidth and enhances the scalability and flexibility, while the scale of system is no longer limited to the throughput of the central server. Experiment results demonstrate the feasibility of this structure and show the efficiency and effectiveness of our system in real-time operation for practical applications. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92673 |
DOI: | 10.6342/NTU202401005 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 電子工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-2.pdf 此日期後於網路公開 2025-05-26 | 5.89 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。