The video images for this ride across bumpy cobblestone pavement were stabilized by recognizing and compensating unwanted translational camera displacements.


The original data as well as the stabilized sequences are placed next to each other.

Download sequence in high resolution
Duration: 20 s
File size: app. 8 MB


These data have been processed on a standard PC in real-time.

In this scene of a busy cross-road the velocities and the directions of various objects in motion were determined.


The evaluated motion data are displayed in the image at fixed grid-points by means of arrows. The arrows point into the direction of the detected motion and the length of the arrows is proportional to the velocity. For a better visibility the velocities are also displayed as color values.

Download sequence in high resolution
Duration: 45 s
File size: app. 19 MB

These data have been processed on a standard PC in real-time.

In these two examples the mass flow from a spreader was analyzed under different light conditions. For this, the video images were stabilized, the mass flow was separated from the moving background and the Optical Flow of the falling masses was determined.


The sequences were stabilized in a way that the high-frequency translational displacements of the camera were compensated, while the panning of the spreader was preserved.

The evaluated motion data of the mass flow are displayed in the images at fixed grid-points by means of arrows. The arrows point into the direction of the detected motion and the length of the arrows is proportional to the velocity. For a better visibility the velocities are also displayed as color values.


In the first video the mass flow is very dark and of low contrast while the background has a large overexposed area. In the second video however the mass flow itself is overexposed in the lower area while the upper part is superposed with the shadow of the spreader. The two sequences show that the falling masses can be recognized and analyzed reliably even under different and very difficult light conditions.

The original data as well as the stabilized sequences with the Optical Flow data are placed next to each other, respectively.

Download sequence 1 in high resolution
Duration: 30 s
File size: app. 49 MB

Download sequence 2 in high resolution
Duration: 30 s
File size: app. 49 MB

These data have been processed on a standard PC in real-time.

For many applications in industry it is useful to be able to detect and to analyze motions in video data robustly and with high accuracy. Such recognition systems are essential e. g. for the development of autonomous vehicles. But also in other fields, such as in the area of security technology or in mining industry, many processes could be optimized or even completely automated by means of a reliable analysis of video data.

“Teaching” computers human vision is a complex task, which is still under research worldwide. Even though many problems in the field of video analysis are solved in principle, there is still a great need for optimization for being able to create solutions that are suitable for industrial needs. This necessity also becomes evident from the large amount of publications in trade journals dealing with this topic.

Due to our enhancements of up-to-date mathematical models in the field of image correlation we at Syperion have come a big step closer to the “seeing computer” by developing fast, precise and robust algorithms for the analysis of object motion in video data. For demonstration purposes we have developed software which allows to separate moving image contents reliably from the background, even in presence of disturbing camera motion, and to determine the absolute values as well as the directions of the object movements. In the context of video processing, these parameters are also referred to as the Optical Flow.

We have tested the performance of the developed algorithms intensely on computer animated standard-sequences as well as on real video data with complex motion patterns. The investigated scenes contain e. g. bumpy rides on cobblestone pavement, crowds in motion and dumped soil in open cast mining. In all cases the object motion could be determined reliably and with high precision.