Falls are one of the leading causes of death in the elderly people aged 65 and above. In order to prevent death by sending prompt fall detection alarms, non-invasive radio-frequency (RF) based fall detection has attracted significant attention, due to its wide coverage and privacy preserving nature. Existing RF-based fall detection systems process fall as an activity classification problem and assume that human falls introduce reproducible patterns to the RF signals. We, however, argue that the fall is essentially an accident, hence, its impact is uncontrollable and unforeseeable. We propose to solve the fall detection problem in a fundamentally different manner. Instead of directly identifying the human falls which are difficult to quantify, we recognize the normal repeatable human activities and then identify the fall as abnormal activities out of the normal activity distribution. We implement our idea and build a prototype based on commercial Wi-Fi. We conduct extensive experiments with 16 human subjects. The experiment results show that our system can achieve high fall detection accuracy and adapt to different environments for real-time fall detection.