透過您的圖書館登入
IP:3.21.104.72
  • 期刊
  • OpenAccess

基於資料擴增之行為辨識

Data Augmentation for Human Activity Recognition

摘要


由於科技的進步,現在許多硬體設備有著非常好的運算能力,而深度學習也因此有著突破性的發展,成為目前熱門的研究與發展方向,因為使用深度學習架構,只需要足夠的資料量,和良好的網路架構,與傳統的機器學習方法相比,往往都有較好的結果。若深度學習模型缺乏資料量,則其效能就不會那麼理想。本文以行為辨識為基礎,結合感測器資料與深度學習方法,利用感測器資料去進行資料擴增改善辨識結果。由於一般的影像資料擴增方法,並不適用於感測器資料,所以我們親自蒐集部份資料進行實驗,並且提出了五種方法做資料擴增︰混入雜訊、時間序列置換、感測器的資料旋轉以及兩種深度學習生成模型,使用少量的訓練資料,建立更高效能的模型。本研究所使用的方法針對UCI數據集其準確率可以達到98.32%,與過去未使用資料擴增方式的行為辨識相同分類器比起來,最高可以提升1.5%。

並列摘要


Due to the advances in technology and hardware development, deep learning has become a popular research topic in recent years. In comparison with traditional machine learning algorithms, deep learning based approaches have better performance in many tasks. However, the application of deep learning requires ample data and a well -designed neural network architecture. When the training data is not enough or limited, we do not normally obtain good results using deep learning approaches. This thesis proposed an approach based on the fusion of deep learning and data augmentation to improve the activity recognition accuracy. We proposed five methods for data augmentation, including noise injection, permutation, rotation, and two different generative models to reach a high -performance model with limited training data. The experimental results showed that the proposed data augmentation approach could reach an accuracy of 98.32% on a UCI dataset, and boost the accuracy up to 1.5% when compared with data without augmentation.

並列關鍵字

Data augmentation GAN VAE Deep learning Activity Recognition

延伸閱讀