使用視覺外型字母表在複雜場景下偵測人形

人形偵測系統往往需要已知的背景模型才能得到正確的偵測結果，因此若是光影變化劇烈或是場景過於混亂，其偵測的正確率便會隨之降低。本論文旨在發展一個可在未知背景中偵測人形的方法，這個方法是利用邊緣片段模型 (Boundary Fragment Model) 來描述與偵測人形。在建構邊緣片段模型時，首先是使用background-cut演算法利用顏色與對比資訊切割出人形輪廓，並將輪廓分割成大量的邊緣片段。經由評估機制遴選具有代表性的邊緣片段組成 codebook，最後使用 AdaBoost 演算法學習出一個強偵測器。在執行人形偵測時，可由強偵測器中的每一個弱偵測器比對輸入的影像，並在Hough 投票空間中進行投票。接著再使用 Mean-Shift 演算法找出目標人形的質心位置作為偵測結果之輸出。在實際實驗中，本系統分別於室內外各式未知場景嘗試偵測人形。其中包含人形正面全身、背面和人形部份遮蔽等實驗，並藉由不同的參數觀察辨識率的變化。最後展示一系列辨識的結果並討論對於所遭遇到的困難與未來展望。

關鍵字

背景切割；物件偵測；邊緣片段；圖片切割； AdaBoost

並列摘要

A human detection system usually depends on a precise semi-static background model to achieve accurate detection results. Therefore, when the illumination changes abruptly or the background is very complicated, the detection accuracy will be degraded accordingly. The objective of this thesis is aimed to develop a method that can detect human shape from an unknown and possibly complicated background. The proposed method is based on the boundary fragment model (BFM) to describe and detect a human shape. In the training stage of constructing the BFM for human shape, we applied the background-cut algorithm to extract human silhouettes with both color and contrast information. The extracted silhouettes are subdivided into many boundary fragments. A codebook is constructed by selecting boundary fragments that have high discriminating power of distinguishing human shapes from non-human ones. Furthermore, the AdaBoost algorithm is used to form a strong detector with the codebook. In the detection stage, each weak detector is used to detect boundary fragments in the input image for constructing a Hough voting space. Finally, the mass center of a detected human shape is computed with the mean-shift algorithm as the output of the detection method. Real experiments were conducted to test the proposed method with indoor and outdoor scenes. Each test image contains either a frontal view or a rear view of a single person with different degrees of occlusion. The experimental results obtained with different parameter settings are compared and discussed. Finally, the difficulties and our future work of human detection from unknown and complicated background are presented.