A real-time fall detection system monitors the daily activity of especially elderly people to enlist someone's help as fast as possible in case of emergency. This paper presents a new real-time fall detection algorithm using a single commercial accelerometer. After transforming the acceleration data from Cartesian coordinates to spherical coordinates, the main part of the algorithm is based on a fuzzy logic inference system and a neural network. These methods allow both the integration of specific expert knowledge about typical falls as well as generalization ability. In order to compare the achieved performance of the method to those of literature, four fall scenarios (forward, backward, sideward and collapse) were performed and evaluated in a laboratory trial with, in the first instance, 5 test subjects. The average sensitivity of those four fall scenarios reached 94% and the false positive rate was about 0.35%. These results show that one single accelerometer is completely sufficient to implement a reliable fall detection system and, furthermore, that knowledge based methods are a suitable alternative to standard pattern recognition methods.