使用旋轉矩形解決跌倒偵測誤判之假陽性

雖然目前影像跌倒偵測使用矩形長寬比技術日益成熟，但仍有許多日常行為導致跌倒偵測系統誤判如:彎腰、蹲下、手或腳的擺動等日常行為都有可能讓系統誤判為假陽性，所以本文提出一套使用旋轉矩形解決跌倒偵測誤判之假陽性的跌倒偵測系統，其方法可分為靜態排除法與動態排除法，靜態排除法主要是先透過物件偵測法將前景目標取出，並使用本文所提出的旋轉矩形框住物件輪廓，即可計算物件所旋轉的角度，透過此方式偵測跌倒行為可解決大部分使用矩形長寬比技術進行跌倒偵測所產生的假陽性誤判，但針對部分日常行為本文所提出的靜態排除法也有可能誤判為假陽性，所以本文將透過動態排除法解決靜態排除法可能出現的誤判並且還能提高跌倒偵測的準確率。

關鍵字

假陽性；旋轉矩形；跌倒偵測

並列摘要

Although image fall detection uses the aspect ratio technology to become more mature, there are still many daily behaviors that cause false alarm of fall detection system. For example, bending, squats, hand or foot swings and other daily actions can cause the system to misjudge the false positive. In this paper, we propose a fall detection system that uses a rotating rectangle to resolve false positives in fall detection. Our system is divided into two parts: image exclusion method and video exclusion method. Image exclusion method uses object detection to retrieve foreground objects, and then use a rectangular box to cover the object contours, and use the proposed rotating rectangular wrap object contour, which is calculates the angle at which the object rotates. The detection of the fall behavior in this way can solve the false positive misjudgment caused by most of the rectangular aspect ratio techniques for fall detection. However, the static exclusion method for some daily behaviors may also be misjudged as false positives. Therefore, this paper will solve the possible misjudgment of static exclusion method through dynamic exclusion method and improve the accuracy of fall detection.