Wellness monitoring through wearable devices

Human activity recognition is still a popular area of research in recent years. Many researchers have developed various types of wearables sensors that could be worn in different positions (hip, thigh, sole etc.) with varying degrees of success. With the amount of research done in this areas, many research works have studied optimizations techniques in the different parameters of activity recognition such as window size, idle prediction etc. to further improve their classification accuracy. For this project, the framework built from previous work [5], the Robust Activity Recognition using Smartphone and Smartwatch (RARSS) would be extensively tested and further improved. The two test methodologies will also be reused, 5-fold Cross Validation and Leave-one-person-out (LOPO) testing. Newly collected dataset from 11 subjects is also added to the previous dataset which amounts to a total of 26 subjects. Modifications were also made to the order of pre-processing steps, namely data augmentation (mean deduction) and data sampling. These modifications were also tested to study the effects of such modifications. Similar to previous work, extracted features from RARSS will also be compared with two other feature sets. However, instead of only RARSS data, two other datasets would be used for evaluation. The main findings presented in the report are as follows: i. Support Vector Machine models using the default Cost and gamma parameters perform much poorer than previously reported when previous experiments were replicated. ii. The addition of new data subjects increases the Random Forest model scores for LOPO testing while a negligible decrease for 5-fold Cross Validation. iii. RARSS features are shown to be still better than other benchmarking feature sets even when other datasets are used. iv. Using both original and mean deducted data significantly improved the model scores over other features. v. The accuracy of inactive device classification is 90% over 1200 data window samples. vi. RARSS testing using inactive devices reduces the overall F1 scores to 35% The real-time human activity recognition system built by previous work [5] is also adapted and improved in this project. This system shares some of its components from the RARSS system such as pre-processing steps. Hence the modification made to the pre-processing pipeline as mentioned earlier is also effected on the real-time system. Also, a pre-determination phase is added in an effort to improve the robustness of the system. A pre-determination phase determines if the devices are being worn (active) during real time monitoring. Inactive devices are filtered from the activity prediction, thus resulting in a more robust activity prediction.Bachelor of Engineering (Computer Science