Freezing of Gait (FoG) is a debilitating walking problem affecting over 50\% of Parkinson's Disease (PD) patients. Rhythmic cues during FoG can help patients to resume walking. FoG can be detected from lower limb acceleration data, but current detection algorithms lack context, resulting in false positive detection and initiation of cues when these are not required. Cues are only required during ambulatory motion. Adding a task classification model to an FoG detection device can help to increase specificity. Furthermore, using this model, an activity log of PD patients can be kept in the home environment, thereby facilitating disease evaluation. A decision tree model using Bayesian classifiers was developed for task classification. Predictors were extracted from raw data of an ankle-worn tri-axial accelerometer and the best predictors were determined at each decision node using feature selection methods. The proposed decision tree model (trained using healthy subject data) was tested on PD patient data to evaluate transferability. Furthermore, to compare performance and computational speed, our Bayesian decision node classifiers were compared to Naive Bayes classifiers. The proposed model was over 30 times faster, compared to the computational speed of using Naive Bayes classifiers at the decision nodes. Ambulatory windows were identified with a sensitivity over 91% for both healthy subjects and PD patients, showing that the Bayesian decision tree model developed in this study can provide context for an FoG detection wearable, to enable effective cueing. This will help patients to walk more confidently and to keep active.