Algorithm Research Of Human Activity Recognition Based On Time Series

The development of artificial intelligence and the demand of industry application promote the research on human activity recognition(HAR).At present,the main research methods are based on video data or sensor data.However,the practical effect of the former is not satisfactory in consideration of the cost of deployment and the complexity of environment.Thanks to the development of wearable sensors,many studies focus on deploying multiple sensors on human body to achieve good results.However,different from previous studies,this paper tries to study HAR by using a single triaxial accelerometer worn on the wrist to minimize the interference to individual and reduce the cost of sensor deployment.A self-collected Sanitation dataset is collected from the open environment via a triaxial accelerometer integrated into a smart watch.This dataset includes seven types of daily work activity data of sanitation workers.The seven types of activity are:walk,run,sweep,bweep(sweep with big broom),clean,dump and daily activities.After data calibration,the reading drift of acceleration time series is eliminated.And noise filtering is realized by using a third-order Butterworth low-pass filter afterwards.A feature extraction algorithm is proposed in this paper.There are 90 features,including 57 time-domain features and 33 frequency-domain features,are extracted to form a sample set of human activity recognition.In the research of human activity recognition algorithm,traditional machine learning algorithms are mostly used.Support vector machine(SVM)and k-nearest neighbor(kNN)usually achieve better recognition result.And some other researches adopt deep learning algorithm,but very few researches focus on ensemble learning algorithm.In this paper,an improved sub-window based ensemble learning algorithm is used for accurate recognition.The experimental results show that the recognition accuracy is improved significantly compared with traditional algorithms,and the validity of the research results is verified.Traditional HAR based on time series adopts sliding window analysis method.This method is faced with the multi-class window problem which mistakenly labels different classes of sampling points within a window as one class.In this paper,a HAR algorithm based on U-Net is proposed to perform activity labelling and prediction at each sampling point which could be called as dense labelling and dense prediction.The time series of triaxial accelerometer are mapped into images with single pixel column and multi-channel which are input into the U-Net network for training and recognition.Thus the accurate activity recognition of each sample can be achieved.A verification experiment is conducted on the self-collected Sanitation dataset and four open datasets.The experimental results show that compared with other five algorithms,our proposal has the highest accuracy and Fl-socre in each dataset,and the performance is stable enough which shows high robustness.