International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 10 Issue: 07 | July 2023
p-ISSN: 2395-0072
www.irjet.net
Human Activity Recognition Using AccelerometerData 1
2
3
Aishwarya G Shenoy , Rohit Keswani , Subhadeep Das
------------------------------------------------------------------------***-----------------------------------------------------------------------Abstract - Human Activity Recognition (HAR) has a II. APPROACH wide range of applications due to the widespread usage of acquisition devices such as smartphones and its abilityto capture human activity data. The ability to retrieve deeply embedded information for precise detection and its interpretation has been transformed by breakthroughs in Artificial Intelligence (AI). In this paper, the time series dataset, acquired from Wireless Sensor Data Mining Lab (WISDM) Lab, is used to extract features of common human activities from a raw signal data of smartphone accelerometer. A 2D convolutional neural network is used to visualize the data.
A. Pre-processing and Data Exploration The raw dataset, containing over a million rows, has a lot of noise that needs to be eliminated. In order to improve the accuracy, we first clean the dataset by condensing the dataframe. We are shortening the dataframe by reducing the number of participants to five. Then, the existence of any null values is checked. After checking for null values, we determine the data distribution of five participants performing several activities. Through this data distribution we can deduce that the data is unbalanced. Standing examples are fewer in number as compared to Walking examples. As a result, using this data directly will cause overfitting, which will skew the results in favour of walking. To avoid this issue, we need to balance the data by eliminating unnecessary data that is covered in the further section of this paper. Now, we plot the tri-axial orientation to observe the variation in the data and convert all the data from string type to float type. Figure 1 represents the time taken to perform different activities by the participants.
Keywords: Human Activity Recognition, StratifiedKfold cross validation, CNN
I. INTRODUCTION Human Activity Recognition (HAR) is the technique of utilizing Artificial Intelligence (AI) to recognize and classify human activities from raw activity data collected by a range of devices. A few examples of such devices include smartphones and smartwatches. Smartphones and smartwatches consist of accelerometer sensors (tri-axial accelerometers) that is used to measure acceleration in three different dimensions. The orien- tation of the device can be determined by these accelerometers, which can be helpful information for activity detection. The time series dataset used in this paper consists of raw data that is collected from 36 different participants performing different activities such as walking, jogging, sitting, standing, ascending and descending for specific time periods. The data is collected using a sample rate of 20 Hz (1 sample every 50 milliseconds) which is equivalent to 20 samples per second. There are six different attributes in the dataset namely user, activity, timestamp, x-axis, y-axis and z-axis. The dataset contains overa million rows and 6 columns that needs to be cleaned and processed. The challenge of significant feature extraction from the raw sensor data has become easier with the introduction of Deep Learning (DL) in the HAR domain. A 2D Convolutional Neural Network (CNN) is used to classify the data and a StratifiedKfold cross validation is used to split the data into train and test data.
© 2023, IRJET
|
Impact Factor value: 8.226
Fig. 1. Representation of data in a bar graph In order to determine the ratio between training and test set, the variation in the signal values of time taken by the participants while engaging in different activities has to
|
ISO 9001:2008 Certified Journal
|
Page 445