The Research Of Human Activity State Recognition Base On Accelerometers

With the development of society and internet, people formed the habit of working with computer and playing with mobile phone gradually. People always enveloped in working and playing on computer and mobile phones. They spend less and less time on physical exercise, due to the lack of scientific activity assessment tools and effective way to remind. Activity recognition base on accelerometer makes it possible for people to quantify the daily movement and make sense of their self-health. It can recognize the state of activity by wearing some sensors on the body. This paper described activity recognition base on Android mobile phone inner accelerometer. The mainly work of this paper are as following,Give a solution to solve the problem of collect acceleration data set on unstable sensor device by add an asynchronous data queue to ensure the sampling frequency of the data is maintained at about 20 HZ. Because the working frequency of Android mobile phone is not stable and deferent mobile phones has different performance.To process the acceleration time series data set, we raise a method that consideration both the number of sampling points and span of times to split the slide window. This method can make the number of different activities in a single slide window as less as possible. When face the impact of different carrying style of mobile phone. This paper forward a solution that breaking down gravity acceleration part from the raw acceleration time series data before doing feature generation. By this way, the classification result reached 94.2%.We compared different slide window size with the recognition precision and conclude that the best slide windows size at 20 HZ is 6 second or 120 sampling point. We also proofed the effectiveness of consideration both the number of sample point and span time when split the windows.This paper didn’t end their work with recognizing the activity of single window. We conducted an in-depth discussion on correct, summary and statistic the long time activity state. To solve the problem that it is easy to classify wrong when switch the activity, we proposed the forward-backward scan algorithm to smooth continuous window activity which improved the activity summary report effectively.