利用影像處理與類神經網路進行人體坐姿判讀

Abstract

本研究之目的在於使用人體影像進行坐姿的判讀。利用相機拍攝人體坐姿影像，以正規化與高斯濾波法消除背景，之後進行二元化處理。為避免坐姿的判讀會受到影響，嘗試將圖形中小腿刪除後之結果視為第一類目標圖型，將手與下半身刪除的結果視為第二類目標圖型。計算目標圖型的特徵參數－慣性矩（moment of inertia）與慣性積（product of inertia），並將參數輸入倒傳遞類神經網路（Back Propagation Neural Networks，BPNN）進行分析運算，最後再將網路輸出值透過分類器來判讀坐姿。 本研究中有10位受試者，共拍攝708張坐姿影像，其中屬於坐姿正的影像有317張，屬於坐姿不正的影像有391張。隨機挑選其中388張坐姿影像訓練類神經網路，剩餘的320張坐姿影像則進行後來的測試。其測試結果中第一類目標圖型之靈敏度（sensitivity）為80%，準確率（accuracy）為79.38%；第二類目標圖型之靈敏度（sensitivity）為78.13%，準確率（accuracy）為82.82%。因此由本研究的結果可知，以慣性矩與慣性積作為圖形特徵參數所建立起的坐姿判讀法具有辨識坐姿好壞的能力。

The purpose of this study is to identify human sitting postures by using images. We use camera to capture the body's sitting images, remove the background by image normalization and Gaussian Smoothing filter, and then make images binaries. To avoid the identification of sitting postures being affected, we try to remove lower leg of posture images for first target pattern, remove hands and lower body for second target pattern. We calculate the target patterns characteristic parameters - moment of inertia and product of inertia, and then we make those parameters input back-propagation neural network (BPNN) for analysis and computing, finally, the output value of neural network through a classifier to identify the sitting postures. In this study, there are 10 subjects and 708 sitting images, which belong to the good posture were 317 images, belong to the bad posture were 391 images. Randomly selected 388 sitting images to train the neural network, the remaining 320 sitting images do test later. The test results of the images of first target pattern, the sensitivity is 80%, and the accuracy is 79.38%; the sensitivity of second target pattern is 78.13%,and the accuracy is 82.82%. From the results of this study, the posture identify system which we set up by using moment of inertia and product of inertia has the ability to recognize good and bad posture.