Falls are a common cause of injuries and traumas for elderly and could be life threatening. Delivering a prompt medical support after a fall is essential to prevent lasting injuries. Therefore, effective fall detection could provide urgent support and dramatically reduce the risk of such mishaps. In this paper, we propose a hierarchical classification framework based on a novel anatomical-plane-based representation for elderly fall detection. The framework obtains human skeletal joints, using Microsoft Kinect sensors, and transforms them to a human representation. The representation is then utilized to classify the sensor input sequences and provide a semantic meaning of different human activities. Evaluation results of the proposed framework, using real case scenarios, demonstrate the efficacy of the framework in providing a feasible approach towards accurately detecting elderly falls.