基於深度學習對空拍影像 進行人物行為分析

Abstract

近年來，無人機因為其優異的機動性與在高空中飛行的優勢，其相關應用已漸趨普及。而人物行為分析，是安全監控中極為重要的議題，但目前針對空拍影像的相關研究較少。利用無人機所拍攝的空拍影像，因為透視投影的關係，影像中的人物會有傾斜情形，使得一般應用於傳統影像的深度學習架構，在空拍影像中的人物偵測效果並不佳。本研究為了克服空拍影像中的透視投影特性，實際利用無人機拍攝大量的空拍影像以建立空拍影像資料集；我們透過修改現有的深度學習架構，利用此空拍影像資料集重新訓練出新的模型，以改善原本深度學習架構在拍影像中表現不佳的情況；本研究也提出了一項基於影像後處理判斷所偵測人物姿態正常與否的方法，藉此達到空拍影像人物行為分析的目標。

Due to the advantages of high mobility and the ability to fly in the sky, drone has inspired more and more applications in recent years. Deep learning-based human activity analysis is an important topic in security surveillance, however, the research works on such analysis with aer-ial images are insufficient so far. Because of perspective projection, people in aerial images are tilted, and degrading the performance of human activity analysis. In order to cope with the issue of perspective projection for aerial images, we use drone to take a large amount of aerial images and build a dataset of human. We also modify the original CNN architecture and use this dataset to retrain the new model for aerial images. Finally, a post-processing method is proposed to classify the pose of a detected person as normal or abnormal, therefore, accomplishing the task of human activity analysis with aerial images.