背景擷取之研究

Abstract

現今偵測移動的物體通常會先建立當下的背景影像，但背景的影像總會突然受移動的物體、光線變化或是其他干擾的影響，使得背景影像需要不斷地做更新來克服此影響。 我們提出了一個擷取背景影像的方法，先以影片中前N張影像去建立粗略背景影像，再利用第N + 1張影像去取得暫時的背景影像，接著透過背景局部二值模式( BackGround Local Binary Pattern, BGLBP )辨別第N + 1張影像前景和背景的部分，利用此結果及粗略和暫時的背景影像更新粗略的背景影像。之後，對後續的每張影像，利用上述的方法，持續更新背景影像，再針對此背景影像之平均亮度調整至近似目前影像之平均亮度，並辨別出目前影像之前景和背景的部分。

Nowadays, detecting moving objects will first establish a current background image. However, the background image is always affected by moving objects, changes in lighting, or other disturbances, updating the background image constantly can overcome this effect. In this thesis, a method is proposed to extract a background image. First, the first N frames in a video are used to create a rough background image. Next, the (N+1)-th frame is used to create a temporary background image, and then background local binary pattern is used to identify foreground and background parts in the (N+1)-th frame. The result, rough and temporary background images are used to update the rough background image. Then, for each subsequent image, the above method is used to update the background image. Finally, the average lighting of the background image is adjusted to approximate the average lighting of the current frame, and foreground and background parts in the current frame are identified.