為公開封閉場域而設計之線上集成式人臉學習與辨識

在一個公開的封閉場域中（如便利商店、圖書館、展場、賣場等）人員流動頻繁，個別身分未知且駐留時間不定，如何針對現場的人進行即時監控是一個重要的議題，為此我們提出一套線上集成式人臉學習與辨識系統（Online Ensemble Face Learning and Recognition system）。此系統混合運用了兩種學習方法包括：「集成學習」與「線上學習」，集成學習運用我們預先訓練的臉部方向辨識器快速的將觀測對象的臉部影像依照臉部方向進行分類達到樣本資料分群的目的，然後再將不同臉部方位的樣本分別訓練成各別的身分辨識器，如此可縮短整體訓練時間；線上學習共有兩個學習階段分別是初始學習階段與接續學習階段，初始學習階段於人員初進入場域時蒐集影像，並使用所有影像進行學習，接續學習階段則透過場域中的視訊監控系統不斷提供觀測對象不同方位的影像，並不斷地抽換新影像進行訓練，逐步地提升辨識效能。經過實驗測試的結果證實，本研究可藉由初始學習階段利用有限數量的初始樣本快速地達到一定的辨識率門檻；而接下來的後續學習階段則透過在場域中不斷捕獲的新樣本，強化訓練，以改善辨識率。由於本系統不需事先累積大量樣本，人臉的學習與辨識採用即時的方式進行，故適用於在公開封閉場域中，針對現場流動人員進行臨時行蹤監控。

關鍵字

電腦視覺；影像辨識；深度學習；卷積神經網路；身分辨識

並列摘要

Due to the maturity of software and hardware, research in the Deep Learning field has developed rapidly. The critical applications, such as face recognition and object detection, have achieved breakthrough performance in conjunction with CNN structure. In order to obtain better identification performance, researchers highlight the need to expect a more significant amount of sample data and longer training time. However, less research shows that whether CNN training data structure and the total training time can improve performance about identification or not. We expect to design a set of Ensemble-like Online Learning methods to understand these two research variables. In the proposed method, it includes ensemble pre-learning and online post-learning. Ensemble pre-learning uses our pre-prepared facial direction, recognizer to quickly classify images of different face directions of observed objects into sample data grouping. This ensemble re-learning can train different face orientation samples to reduce the time required for training of individual trainers and to maintain the identification performance. Online post-learning provides sample images of different orientations of observed objects through the video surveillance system. The online learning method can continuously train new samples to improve the overall identification performance gradually. The research results will show that the efficiency of data training can be improved by pre-integration learning, while post-line learning can break the total training time into several moments. The results of this study will be applied to object tracking in some places, including cashier-free conveniences stores, exhibition halls, and apartment complexes. The accuracy can be upgraded by training uninterrupted data from the Video Surveillance Systems. Finally, comprehensive environment surveillance systems will be developed by means of Ensemble Online Learning and indoor positioning technology.

並列關鍵字

Computer Vision ； Image Detection ； Deep Learning ； Convolutional Neural Network ； Face recognition

國際替代計量

為公開封閉場域而設計之線上集成式人臉學習與辨識

全文下載

主題瀏覽