Activity Recognition Technology Based On HF-SVM And Feature Selection Using Hierarchical Clustering

Intelligent computing based on human activities is an important research direction in the field of artificial intelligence.Its purpose is to provide users with intelligent application services by acquiring user's state and environment data information.With the rapid development of mobile intelligent devices(such as intelligent wearable devices)and related sensors,activity recognition technology based on the intelligent equipment has become the research hotspot.Due to the limitation of hardware resources,such as computing power,storage space and energy,and the traditional machine learning model needs huge computing power,mobile intelligent device-based activity recognition technology can not be widely used.In order to solve the above problems,in this paper,the traditional feature selection and support vector machine techniques are studied,and then proposes a multi-category classification method based on hierarchical clustering feature selection and hardware-friendly kernel function.The main algorithm improvements and achievements are as follows:(1)An improved feature selection algorithm based on hierarchical clustering algorithm is proposed.The traditional hierarchical clustering-based feature selection algorithm uses an evaluation function based on mutual information and correlation coefficients.This does not apply to continuous data in the field of motion recognition.This paper improves the evaluation function based on Pearson correlation coefficient and shared Nearest Neighbor.Using the improved feature selection algorithm based on hierarchical clustering,the feature extraction is completed and the complexity of the activity recognition technology in the process of model training is reduced.(2)The SVM algorithm based on hardware-friendly kernel function is proposed.Traditional SVM algorithm needs a lot of exponential operations in model training and algorithm application.Based on the Gauss kernel function and the Laplacian kernel function,this paper proposes a hardware-friendly kernel function,which not only has the advantage of preserving the anti-noise of Gaussian kernel function,but also has a small computational cost.(3)Propose the activity recognition technology based on combinatorial classifier.Activity identification is a multi-class classification problem.The traditional multi-class classification algorithm has the problem of large classification error.In this paper,the SV algorithm based on OVO strategy is proposed.The output of multiple classifiers is applied to the Sigmoid function,and the final output is obtained according to the vote of each binary classifier.The method proposed in this paper avoids using Sigmoid function to get the maximum value,and it is easy to be affected by the noise.It also avoids the errors caused by the same classifier weight.In this paper,the activity recognition technology is studied from the above three aspects,and the experimental analysis shows that the proposed activity recognition technology has high accuracy and practical value.