Freezing of Gait (FoG) is an episodic motor disturbance in Parkinson disease (PD) that causes patients to be unable to initiate or maintain their locomotion. Prior work that used simple and easy-to-learn algorithms based on a singular feature and rule-based classifiers are not sufficient to learn variations in patient walking styles and freezing patterns. Efforts to use machine-learning algorithms suffer from challenges caused by imbalanced datasets. Here, we propose a new approach for FoG detection using a wide set of online calculable features and an instance-based and non-parametric Possibilistic K-Nearest-Neighbor (KNN) classifier. The issue of imbalanced datasets is addressed using the Self-Organizing-Map (SOM) algorithm.