Anomaly detection and regime searching in fitness-tracker data

In our project, we solve the problem of human activity monitoring based on data from sensors attached to the hands of various workers. First of all, the recognition results help to increase labor productivity and optimize production processes at a building site. Also, the analysis of the behavior of workers allows us to track a person's well-being, compliance with safety measures and accident prevention. Data collected from the fitness tracker, require careful preprocessing. The Gaussian Process model was applied to fill in the gaps in time series and extract outliers, that increase metrics of the models. The comparison of several models for activity recognition was performed if form of supervised learning. An anomaly detection approach was applied and provided useful results for activity monitoring during construction work. In addition, the neural network based on the architecture of variational autoencoder allowed us to extract main work regimes. The fitness tracker time series data set was collected, tagged and published for further research.