Abstract

The recognition of physical activities performed by humans is first identified as a classification problem that when solved provides an automated approach to recognising what is being done at any given time with respect to motion. The primary aim is to recognise activities using wearable sensors. The ability to recognise human physical activity is important to many sectors such as health care, security, rehabilitation and general well-being. Information extracted from the activities performed enables the systems to learn and perform better. In general, this concept is important especially in healthcare as it could be used as an indicator of health status and lifestyle choices therefore the ability to correctly predict activities is very important.

The design of this study is based on the use of three traditional machine learning algorithms in comparison with a deep learning algorithm to determine which algorithm achieves higher prediction accuracy using each model’s base parameters as the benchmark for evaluation. Random Forest, Support Vector Machines and K Nearest Neighbor were the traditional algorithms used while Convolutional Neural Network was the deep learning algorithm. All models were cross-validated and evaluated with classification evaluation metrics.

CNN achieved the best classification accuracy of 99.9%. Significant results were also obtained from RF and SVM. All models' hyperparameters were tuned with the best parameters to achieve these results. KNN however, performed the least when compared to the other models.

Finally, the thesis is concluded with discussions on the entire research and identifying limitations, and ideas for possible research areas to be investigated for future work.

Predicting Human Activity using Motion Tracking with Machine Learning