Abstract:
Many developments have been observed from research into activity recognition. Alongside these developments, many challenges have also been identified which affect the design, implementation and evaluation of the activity recognition systems performance. One such challenge is the successful inclusion of contextual awareness in order to improve the system’s performance. This research seeks to examine the effect of localising a wearable device, in the activity recognition problem. Three machine learning models were implemented, which make use of the on-body device location in different ways. The first model contains no knowledge of the on-body device location, the second model contains the encoded location of the device as a feature in the dataset, the third model separates each dataset according to their corresponding location, with each location being treated as an independent problem. A final fourth model was proposed and implemented which attempts to closely emulate the best performing model of the previous three, while being fully autonomous. The autonomy is achieved by applying another classification step to determine the device location and then performing activity recognition. The performance of each model was tested using various combinations of feature selection algorithms and classifiers. When using no location information, model 1 generated a classification accuracy of 89%; using the location as an encoded feature inserted into the dataset, model 2 yielded a classification accuracy of 90.2%. Classification of the activities when considering training data only from the location of the wearable device, model 3 generated an average accuracy of 95.5%. The fully autonomous model 4, which was based on the activity recognition in model 3, managed to achieve a 94.5% classification accuracy. These results show that using the location of the device to give the system added context, makes a statically significant impact on the performance of the system.