Unsupervised personalized human activity identification method based on multi-sensor data alignment

Abstract

The invention discloses an unsupervised personalized human activity recognition method based on multi-sensor data alignment, and the method comprises the steps: 1) carrying out the preprocessing of activity data, collected by a plurality of wearable sensors, of a training user and a new user, and respectively constructing a source domain data set and a target domain data set; 2) optimizing a humanactivity recognition model by utilizing the labeled samples in the source domain data set, so a good activity classification effect is realized on training user data; 3) selecting samples from the source domain data set and the target domain data set, adding domain labels, and using an adversarial learning strategy to alternately maximize / minimize domain discrimination loss so as to align distribution of multi-sensor data of the training user and the new user in the feature space; 4) inputting the preprocessed new user data into the activity identification model with determined parameters toobtain an activity identification result. The method can improve the activity identification accuracy of the new user.

Description

technical field [0001] The invention relates to the field of human activity recognition, in particular to an unsupervised personalized human activity recognition method based on multi-sensor data alignment. Background technique [0002] Human activity recognition based on wearable sensors uses the data collected by the sensors worn by the user to infer the user's activity information. support. Commonly used sensors include accelerometers, gyroscopes, magnetometers, etc. According to the number of sensors used, they can be divided into human activity recognition based on a single wearable sensor and human activity recognition based on multiple wearable sensors. Due to the small amount of information collected by a single wearable sensor, it is generally only used to identify simple human activities, such as walking and running. With the development of wearable sensors, human activity recognition based on multiple wearable sensors has received extensive attention. Multiple w...

AI Technical Summary

Problems solved by technology

Multiple wearable sensors placed on different body parts of the user have great differences in data modes, placement orientations, etc., and have different data distribution rules. It is difficult to take care of all sensors by directly using the unsupervised domain adaptation method designed for a single source of data. The data distribution law of the data makes it impossible to achieve fine-grained distribution alignment

Images