Collision recognition from a video

Results

Recognition and Differentiation Between Static and Moving Objects

Under certain conditions, we get very good results - all moving objects are correctly classified, and only a few static objects are wrongly classified as moving.
Here is an example, taken from a camera set on top a Roomba robotic vacuum cleaner -

Collision Detection

We couldn't get to testing collision detection on real movies. Instead, we've tested the algorithm on a synthetic environment, which should emulate realistic optimal conditions.
The results we got on this synthetic environment can be summarized by this graph -

Where scenario 4 is a collision scenario, and the rest of the scenarios are non-collision scenarios.
On the bottom line, we get only 2% false negatives and 12% false positives.