隨著人類經濟進步與社會變遷，人們在日常生活的意外事件發生的種類中，跌倒占了相當大比例。本篇論文以Viterbi演算法為基礎，透過kinect擷取圖形的人體資料，達成跌倒分析之依據，以通報相關單位，避免嚴重後果。 維特比(Viterbi)演算法可以運用在各種不同領域中，例如通訊領域中接收端的資料更正。在有限的雜訊通道傳輸時，藉由機率累積，可以找出最大可能﹙maximum likelihood﹚的資料路徑﹙survivor path﹚，以找出最有可能的原始傳送資料，進而達到資料更正之目的。 相同的觀念我們也可以把這個演算法運用在不同的領域。當某種資料與已知的資料庫做比對，若與資料庫的資料有相似的變化特性，則我們可以把這種規則當成判讀之依據。以跌倒偵測而言，目前已知的方法大約可分為攜帶感測器在身上與影像分析。攜帶感測在身上其缺點為通常使用者會遺忘攜帶感測器，影像分析通常因為資料處理使精準度不高。 本篇論文藉由kinect擷取圖像進行偵測人體骨架，藉由紀錄骨架移動位置，以Viterbi演算法進行跌倒資料庫比對，以圖形來分析人們是否跌倒。經過實驗的分析，我們得到相當高的精準度。Fall detection is an important issue in action recognition field because falls usually imply that subjects might be in danger. This thesis developed a system which detects fall events base on the configuration of human body model using the Viterbi algorithm. The configuration of human body is obtained via Kinect. The Viterbi algorithm is widely used in various applications such as the data correction mechanism in communication field. When transmit data through a noisy channel, correct data stream can be successfully traced back. Same idea can be adopted in the pattern recognition field. A system detects falls if the input data sequence and the database template have similar trends. Fall detection systems can be divided into two categories; sensor-carrying and image-based. The disadvantage of sensor-based method is that users sometimes forget to bring sensors. The recognize rate of image-based methods are usually affected by the change of illuminations in the environment. This work provides a method to detect fall by analyzing the images obtained from Kinect. First, the skeleton of subject is modeled. Second, the trajectories of each joint are recorded and analyzed using the Viterbi algorithm. Finally, the state transition obtained from Viterbi algorithm is used to determine the occurrences of fall. Experiments demonstrate that 95% of the fall events are successfully recognized using the proposed method.
