請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73876
標題: | 以跨鏡頭多目標追蹤分析建築內使用者行為 Indoor Occupant Behavior Analysis with Multi-Target, Multi-Camera Tracking |
作者: | Shun-Hsiang Hsu 許舜翔 |
指導教授: | 謝尚賢(Shang-Hsien Hsieh) |
關鍵字: | 電腦視覺,深度學習,行人重識別,跨鏡頭多目標追蹤,建築管理, Computer Vision,Deep Learning,Re-identification,Multi-Target Multi-Camera Tracking,Building Management System, |
出版年 : | 2019 |
學位: | 碩士 |
摘要: | 文明發展至今,建築與人類日常活動有密不可分的關係,建築內的環境對生活品質亦有舉足輕重的影響,而建築管理便扮演著監測、維護和改善環境的角色,因此,人們致力於提出有效的管理手段和決策方針,但大多受限於建築的種類、空間範圍過大等因素,只能簡化使用情形的評估或採消極的態度來維護,而隨著物聯網技術發展,管理者得以全天候監測運營中建築物的多項數據資料,進而應用於建築管理中,大幅減少過往因管理而需花費額外的人力成本,且能獲得更符合實際營運情況的決策方向。在眾多影響空間使用狀況的數據中,又以建築使用者的行為模式最具相關性,然而,因使用者行為的複雜性和隨機性,使得現今仍無一有效的方法將行為模式量化和作大範圍的監測,導致預想情形和實際上存有很大的落差。因深度學習技術而於近幾年有很大的進步,對雜訊的適應力變得更加強大,本研究擬利用電腦視覺中的跨鏡頭多目標追蹤方法,透過結合深度學習的追蹤方法,應用於現存於建築各處的監視設備來蒐集足跡資料,進而對使用者於建築內的使用行為進行分析。本研究所採用的方法係先於單鏡頭下進行即時追蹤得到行人軌跡,再利用卷積神經網絡提取的外觀特徵串聯不同鏡頭下監視資料得到的軌跡資料,進而從這些資料得到建築規模(building level)的空間使用率和使用率分佈,且能偵測異常情況並作後續位置追蹤,綜合上述,管理者便能利用這些數據,以檢討和訂定能源策略、維護公共安全和作緊急避難應變等。 During the development of society, buildings and people’s daily activities are inseparable. Therefore, the indoor environment has a great impact on the quality of life and researches on building management systems were devoted to achieve the goal of enhancing energy efficiency and occupant comfort. With the increasing trend of IoT application, data analytics approaches had emerged and they can be applied to understand better indoor environment. However, it is problematic in practice to adopt sensing technology due to stochastic nature of occupant behaviors and large-scale monitoring area. Therefore, a cost-effective and accurate method is required to collect data regarding occupant behavior. This research aims to implement a re-identification system for multi-target, multi-camera tracking with surveillance cameras to obtain more reliable occupancy data. In recent years, tracking combined with deep learning techniques has better performance and more robust to visual obstacles like dim-lighting or being partially obstructed than traditional approaches. The advance in tracking gives the opportunity to develop an application for behavior analysis of building occupants. This research proposes the distributed system for tracking cross non-overlapping cameras. Firstly, multiple object tracking is performed under each camera; then, the probe images of occupants provide appearance and location information. Secondly, feature vectors extracted from the images by the convolutional neural network are used to concatenate trajectory data from different cameras. Finally, the concatenated data are analyzed for usage rate of spaces and their distribution in building levels. Moreover, abnormal situations can be detected and tracked cross multiple cameras. With the analysis, the building manager can not only validate and revise the energy strategy but also enhance public safety and better handle emergency conditions. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73876 |
DOI: | 10.6342/NTU201903037 |
全文授權: | 有償授權 |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-108-1.pdf 目前未授權公開取用 | 5.47 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。