Human activity recognition (HAR) is a tool that has become increasingly common and easier to perform over the last few years with the huge growth of people using smartphones equipped with a wide range of different sensors and access to the internet. The uses for HAR are countless including healthcare, fitness and advertising. However, HAR algorithms are not always perfect and smartphone sensors produce a huge amount of data every second meaning there is a lot of interest around improving these classification algorithms not only for increased accuracy but also for speed and efficiency, particularly important for lower-power devices with small batteries like smartphones.
The aim of this report was to test and evaluate machine learning and data science methods to classify the activity in a given time window from sensor data generated from a smartphone (particularly accelerometer and gyroscope data). Different approaches to this problem are tested with the aim of maximising classification accuracy while keeping speed and efficiency in mind. The algorithms must be robust and generic enough classify unseen or unusual data as this will be vital in a real-world implementation.
The analysis found that very high performing classification algorithms could be created to not only classify activities but extended to detect falls with extremely high accuracy as well. A wide range of algorithms were tested but overall it seemed that the most information could be inferred from the frequency-domain representation of the accelerometer signals as these were the algorithms that performed best throughout. The main achievements of this report involved producing a fall detection classifier which performed 100% precision with minimal amounts of data in training/testing as well as an activity classifier that performed 99% precision classifying between five different activities.
