The widespread use of multimedia technologies gives several opportunities for the applications of video processing, an important area of which is motion detection. The processing of high resolution videos would require a lot of resources however. If we know where to look for the moving object on the next frame, we can reduce the area we need to examine, saving a significant amount of resources.
The goal of my thesis was to implement the tracking of a potentially fast moving object on a video recording, and to examine if it is possible to give an estimation of the object’s next position. In this thesis I gave an overview of some of the most common methods of object tracking and motion estimation, and then created the programs implementing these using OpenCV-Python. The goal of the estimation was to give an approximation of the objects coordinates on the next frame, reducing the area where I need to look for it.
For the implementation I used simple recurrent neural networks, the Elman network and the Jordan network. I trained these with a smaller part of the full recording, and then tested their behavior with a different part of it. During the testing I calculated the errors of the network outputs, and compared the two based on these.