Human Activity Recognition using Distributed Computing

​

Technologies used:

​ - Apache PySpark - Spark ML - H2O - AWS EMR - AWS S3 ​ ​

Contributors

​

• Shreejaya Bharathan
• Ivette Sulca
• Roja Immani
• Aakanksha Nallabothula Surya
• Sakshi Singla
• Akansha Shrivastava ​

Goal

Determine which device-meter combination is best at recognizing Human activities among the following available devices and sensors per device: - Devices: Phone, watch - Sensors: Accelerometer, Gyrometer

Dataset

​ source : https://archive.ics.uci.edu/ml/datasets/WISDM+Smartphone+and+Smartwatch+Activity+and+Biometrics+Dataset+ ​ Data (X,Y,Z coordinates) has been collected for 51 subjects for about 18 activities at a rate of 20Hz for 3 mins each: ​ • Non hand-oriented:walking, jogging, stairs, standing, kicking • Hand-oriented (G): dribbling, playing catch, typing, writing, clapping, brushing teeth, folding clothes • Hand-oriented (E): eating pasta, eating soup, eating sandwich, eating chips, drinking

2 devices & 2 sensor types : ​ device : phone & watch
sensor : gyroscope & accelerometer

Number or observations: 15630426 ​ ​

Process

​

1. Data merged into one dataframe and stored to S3 bucket with the following columns: user, device, sensor_type, activity, timestamp, x, y, z ​
2. Features extracted for each user-device-meter-activity combination: mean(x), mean(y), mean(z), std(x), std(y), std(z), covariance(x,y), covariance(y,z), covariance(z,x), histogram(x,4), hostogram(y,4), histogram(z,4), range(x), range(y), range(z) ​
3. Fit 10 models on the extracted features with different device-sensor combinations and compared accuracies ​ ​ ​
4. Found XGBoost with all four sensors to be the best at classifying the activities, ran it on various AWS EMR cluster settings to compare runtimes ​ ​ ​

Conclusions:

​

1. All four sensors together classify human activities best ​
2. XGBoost model using H2O gives the model with best accuracy (70%) ​
3. Hand-oriented sitting activities like eating pasta, soup, sandwich etc are the most difficult to correctly identify ​ ​

Description of Files

1. Preprocessing.ipynb - code used for preprocessing and feature extraction
2. modelling_notebooks - code for the 10 models fitted
3. visualization/preprocess_for_viz - preprocessing code used for generating time series and aggregate value visualizations of the various activities
4. visualization/visualizations on Plotly - code for generating time series and aggregate value visualizations of the various activities