In this paper we present a novel computer vision-based movement detection algorithm which can be used for applications, such as human detection with unmanned aerial vehicles. The algorithm uses the deviation of all pixels from the anticipated geometry between 2 or more succeeding images to distinguish between moving and static scenes. This assumption is valid because only pixels which correspond to moving objects can violate the epipolar geometry. For the estimation of the fundamental matrix we present a new method for rejecting outliers which has, contrary to RANSAC, a predictable runtime and still delivers reliable results. To determine movement, especially in difficult areas, we introduce a novel local adaptive threshold method, a combined temporal smoothing strategy and further outlier elimination techniques. All this leads to promising results where more than 60% of all moving persons in our own recorded test set have been detected.