Motion detection aims at discriminating between moving objects and a static environment. This task can be seen as the grouping of local motion observations into moving objects. The framework we propose is derived from a perceptual grouping principle, namely the Helmholtz principle. It consists in defining an image model in the absence of moving objects instead of modeling the moving objects. This prevents from any complex model design while enforcing the generality of the approach, since there is no prior to specify on the objects to be detected. Detections are then said to be performed a contrario moving regions appear as low probability events in the "no motion" or a contrario model. The modeling framework induced by this approach is compact and handy, since it is simply built on independant identically distributed random variables. Furthermore, computing automatic detection thresholds and attaching a confidence level to each detected moving region is possible through the probalistic setting of the framework. The resulting detection algorithm is thus truly generic and avoids parameter tuning. The method performance is assessed on various real image sequences.