A frame differencing algorithm that allows for small camera movements.

Abstract: The field of computer vision is ever-changing, and with more video being captured every day than before. This thesis examined motion detection through frame differencing. More specifically, this thesis examined what effect small camera movements had to the motion detection and if there was an algorithm that could mitigate some of the unwanted camera movements from the resulting motion detection. It was examined through having different test case with varying amounts of movement inflicted onto the camera. Test cases with and without moving objects in the frame were examined. The moving objects were annotated and used as a metric to determine the F1-score. The two algorithms that were compared both managed to mitigate some vibrations and movement of the camera. One aspect that would benefit further research within this area is pixel perfect annotations of the moving objects within the frame. Potential future research could be on a heuristic-based approach or if a deep neural network can perform frame differencing and mitigate small camera movements.