Passive and non-obtrusive health monitoring using wearables can potentially bring new insights into the user's health status throughout the day and may support clinical diagnosis and treatment. However, identifying segments of free-living data that sufficiently reflect the user's health is challenging. In this work we have studied the problem of modelling real-life gait which is a very indicative behaviour for multiple movement disorders including Parkinson's disease (PD). We have developed a probabilistic framework for unsupervised analysis of the gait, clustering it into different types, which can be used to evaluate gait abnormalities occurring in daily life. Using a unique dataset which contains sensor and video recordings of people with and without PD in their own living environment, we show that our model driven approach achieves high accuracy gait detection and can capture clinical improvement after medication intake.