Personification of Bag-of-Features Dataset for Real Time Activity Recognition

• Published in: 2016 3rd International Conference on Soft Computing & Machine Intelligence (ISCMI) pp. 73 - 78
• Main Authors: Gadebe, Moses L., Kogeda, Okuthe P.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.11.2016
• Subjects: Accelerometers; Activity recognition; bag-of-features; BMI; Classification algorithms; dataset; Euclidean distance; Feature extraction; Human Activity Recognition; K Nearest Neighbour; new user; personalization; personification algorithm; Real-time systems; Smart phones; subject
• DOI: 10.1109/ISCMI.2016.27

Personalization of activity recognition is possible and important, when existing public dataset collected from large group of subjects can be tailored and be used as training and testing dataset for new users (subjects) who have similar personal traits. However, due to shortage of personalized dataset and techniques to tailor public dataset for new users weakens the personalization of human activity. To address shortage of personalized dataset, we propose a personification algorithm that extracts and tailor-make bag-of-features dataset to support new users from publicly available Human Activity Recognition dataset (PAMAP2 and USC-HAD). Studies indicate that BMI can be used to profile user's weight as either normal weight or overweight or obese, which could be used to predict cardiovascular diseases. For that purpose our personification algorithm uses height, weight and BMI to generate human activity bag-of-features. The personification algorithm is implemented in Scala and Java programming languages and is deployed on Apache Spark Server. We validated our algorithm, by running three set trials of experiments for each 5 K threshold values using 2 randomly selected new user's profile against two publicly available Human Activity Recognition dataset PAMAP2 and USC-HAD. The results indicate that it is possible to tailor bag-of-features from public dataset. Overall performance of our algorithm shows precision, recall and F-score of 0.70%, 0.50% and 0.60% respectively.