How & Why?

How did we do what we did

The trials were conducted on a group of 30 volunteers ranging in age from 19 to 48. Each participant used a smartphone (Samsung Galaxy S II) while doing six tasks (walking, climbing stairs, walking down stairs, sitting, standing, and lying). Using its inbuilt accelerometer and gyroscope, we acquired 3-axial linear acceleration and 3-axial angular velocity at a constant rate of 50Hz. The experiments were videotaped so that the data could be manually labeled. The resulting dataset was divided into two sets at random, with 30% of the volunteers chosen to create test data and 70% of the participants chosen to create training data.

After applying noise filters as a pre-processing step, the accelerometer and gyroscope sensor data were sampled with fixed-width sliding windows of 2.56 seconds and 50% overlap (128 readings/window). A Butterworth low-pass filter was used to separate the gravitational and body motion components of the sensor acceleration data into body acceleration and gravity. Since it is believed that the gravitational force solely consists of low frequency components, a filter with a cutoff frequency of 0.3 Hz was employed. A vector of features was generated from each window by calculating variables in the time and frequency domain.

Our Approach

In order to learn from the data and predict human activity, we first employ a pre-engineered dataset with traditional machine learning (ML) model. Then, using a RAW dataset and a deep learning model, we could anticipate human activity by learning from the data.

To put it simple terms, you could say that it is a multi-class classification, time-series analysis issue because, given a fresh data point, we must forecast human activity. Additionally, each data point relates to one of the six activities.

Motivation of our project

Smart phones, which use cutting-edge technology to give users intelligent aid in their daily operations, have emerged as one of the most useful tools in our daily lives for communication. Mobile phones have features like cameras, GPS, web browsers, and other embedded sensors, which enables the improved performance of applications in light of the client's specific location, movement, and context. Portable working frameworks with type of machine learning and connectedness, interfaces for application development for executing third-party tools and applications. We, as a team, were very intrigued by the kind of data collection that is used for this project and also wanted to learn sensor data cleaning based on digital signal processing , how it can be used to predict an activity based on the training data.

Page updated

Google Sites

Report abuse