使用分組式部分模型之基於近紅外線夜間行人偵測

Abstract

在智慧型運輸系統領域中，行人偵測是一個重要的議題。由於行人並非光源體，在夜間黑暗的環境下，駕駛者更不容易注意到行人，使得夜間行人偵測對於幫助偵測行人以協助駕駛的視覺系統是十分關鍵的。因此，本論文提出一個在移動式車輛上使用之基於近紅外線光源的夜間行人偵測方法，在近紅外線影像中會顯示輻射近紅外線以及反射經由近紅外線投射燈投出的紅外線的區域。但在某些情況下，行人身上衣物的材質會吸收大部分投射在身上的紅外線，造成這個行人部分區域無法在紅外線影像中顯示。為了解決這類的議題，我們提出了一種部分式的行人偵測方法，根據在人身上標記的關鍵點將行人分成許多部份。由於計算負荷量太高，有效部分的選擇成為當務之急。本研究中首先收集了大量的夜間影片，然後分析偵測率/處理時間和不同數量/類型的部分之間的關係。此外，系統將會學習任兩個部分間的空間關係。藉由這些空間資訊，即使遇偵測的行人部分被遮蔽或者正反左右偵測錯誤，還是可以強化對偵測到的行人部分的信心。在偵測步驟之後我們試著去精鍊其結果，進而提出一種基於塊的分割方法，採用穩健的局部門檻值的方法和兩種過濾器，來驗證行人的偵測框。由實驗可證明，此系統有著令人滿意的結果。

Pedestrian detection is an important issue in the intelligent transportation system field. Since the pedestrian is not an apparent object at nighttime, it is more critical for the vision system to help detecting the pedestrian for assisting driving. Accordingly, this thesis proposes a nighttime pedestrian detection method for a moving vehicle equipped with a camera and the near-infrared lighting. The objects in the nighttime environment will reflect the infrared projected by our emitted spotlight. In some cases, however, the clothes on the pedestrian will absorb most of the infrared and make the pedestrian partially invisible in that part. To deal with this, a part-based pedestrian detection method which divides pedestrian into many parts according to the key points marked on them is used. Due to high computation load, selection of effective parts becomes imperative. In this research work, we first collect a large number of nighttime videos and then analyze the relations between the detection rate/ processing time and different numbers/types of parts. Besides, we will learn the spatial relationship between every pair of two parts. By using this spatial information, we can enhance the confidence of the detected parts even if some parts are occluded or are same types but with different opposite body part. While trying to refine pedestrians after detection, we propose a block-based segmentation method which uses robust local thresholding approach and two filters to verify the bounding boxes. The proposed system is verified by experiments and appealing results have been demonstrated.