使用延伸形狀探勘之撿取手勢辨識

Abstract

穿戴式裝置識別人體手勢的使用越來越多，在現場還有許多工作能產生良好的識別結果。然而，現有的研究仍然無法顯示特徵和特徵能夠對手勢進行分類的關鍵特徵的形狀。在這項工作中，我們提出了SenseShape，一種基於Shapelet Mining的算法。這是一種強大的算法，由於本地功能的使用，噪聲和失真都很健壯。我們修改算法以處理數據為多維的慣性傳感器，以識別物品採集手勢中的微量活動。 SenseShape通過將距離計算以外的統計特徵實現為慣性數據，改進了傳統的形狀挖掘。為了探索我們視野的可行性，我們使用用戶腕部佩戴的6軸傳感器進行實驗，收集物品在3個不同高度的貨架上放置物品採集手勢數據。新提出的方法表明，無論貨架位置或移動速度如何，精確度和回歸率分別為94.5％和91.6％，可以確定採摘姿勢的微觀活動。

The usage of wearable device to recognize human gestures are getting more exposed, there are many work in the field which yield good recognition result. However, existing study still can not show the shape of a key feature which special and distinct-able to classify gestures. In this work we propose SenseShape, an algorithm based on Shapelet Mining. It is a powerful algorithm which robust to noise and distortions because of local features usage. We modify the algorithm to work with inertial sensor whose data are multidimensional to recognize micro activity from item-picking gesture. SenseShape improved the traditional shapelet mining by implementing statistical features other than distance calculation into inertial data. To explore the feasibility of our vision, we conducted experiments using 6-axis sensor worn at wrist of the user to collect item-picking gesture data where objects are put at 3 different height levels of shelf. The new proposed method shows that it can identify micro activity of item-picking gesture regardless of shelf-location or movement speed with 94.7\% accuracy while precision and recall rate is at 94.5\% and 91.6\%, respectively.