請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/1354
標題: | 藉由多邊界盒多任務學習網路辨識遠距離動作 Recognizing Distant Actions via Multi-box Multi-task Networks |
作者: | Chao-Lun Wu 吳兆倫 |
指導教授: | 陳銘憲(Ming-Syan Chen) |
關鍵字: | 動作辨識,卷積類神經網路,多任務學習, Action Recognition,Convolutional Neural Networks,Multitask Learning, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 無人機的技術在近幾年有著突破性的進展,並且在諸多領域有豐富的應用如監視、救援、運輸以及軍事方面等等。本研究的目標是建立一個能夠在無人機上偵測、辨識事件的模型。這個研究問題困難的部分有兩點:第一,無人機相關的錄影資料非常稀少,要能夠以少量資料訓練出一個泛化能力高的模型十分困難。第二,由於無人機的位置通常離地面較遠,拍攝到的人物動作占畫面的比例很小,會令模型難以辨認人物動作。為了解決這些問題,我們提出了兩步驟的模型。首先先以SSD偵測出人物的位置,之後再藉由多任務學習架構,跟大型人物動作資料庫一起訓練的模型來辨識無人機影像中人物的動作。我們以自己提出的無人機人物動作影像資料來驗證我們的模型。這個影像資料包含14種類型的人物動作。實驗結果說明我們提出的方法可以增加無人機影像的人物動作辨識率。 The technology of drone has advanced significantly during the last few years, which enables drones to be deployed in many tasks including video surveillance, search and rescue, last-mile delivery and military operation. The great potentials attract many researchers to study visual recognition technologies for drone, e.g. object detection in aerial images. However, there is not much research related to action recognition in drone videos. In this thesis, we aim to develop a real-time action detector of drone that can recognize complex human actions such as running, eating, walking, etc. Action recognition in drone is a challenging task due to the following reasons. First, there is no large-scale action dataset of drone, and the scarcity of training data makes learning accurate neural networks difficult. Second, the actions happen at a distance and are hard to be localized. To address this first issue, we propose a multi-box multi-task network architecture for recognizing actions at a distance. The multi-box network is used to generate human location proposal, and the action recognition network is then applied to the proposed locations to detect actions. In terms of the data scarcity, we attach this problem by leveraging the existing large human action databases with multi-task learning. To evaluate the effectiveness of our method, we create a new drone action dataset with 138 videos and 14 different distant actions. Experimental results show that our proposed method can increase the action recognition rate in drone. |
URI: | http://tdr.lib.ntu.edu.tw/handle/123456789/1354 |
DOI: | 10.6342/NTU201803559 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-1.pdf | 5.26 MB | Adobe PDF | 檢視/開啟 |
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。