Human behavior classification based on CEEMDAN-WT denoising and cascade forest
Chen Zhiqiang, Minmin Miao, Wenjun Hu · 2021 China Automation Congress (CAC) · 2021
Human Behavior Recognition (HAR) is a hot research topic in the fields of machine learning, human-computer interaction and pattern recognition, which has a wide range of social application value. Considering poor generalization ability of machine learning model and the long training time of deep learning model, a human behavior classification method based on cascade forest model is proposed. Firstly, original data set is screened and divided into several segments by using human behavioral acceleration data set provided by wireless data mining laboratory of Fordham University. Then, the filtering process is completed by combining complete EEMD with adaptive noise(CEEMDAN) and wavelet transform. Taking 10s as the window size, the average composite acceleration, grouping proportion and various statistics are extracted from the triaxial acceleration data as features. The six behaviors (walking, jogging, upstairs, downstairs, sitting, standing) are classified by using the Cascade Forest Algorithm in comparison with traditional machine learning method and Deep Neural Network (DNN) algorithm. Experimental results indicate that the classification indexes of cascade forest are above 96%, which is better than traditional machine learning and deep neural network algorithm and shows the proposed model effectiveness for human behavior recognition.