Motion detection and temporal filtering of noisy image sequences

When capturing digital video there is always some amount of noise in the resulting signal. This is more pronounced in low-light conditions. In this thesis we have evaluated five algorithms for motion detection and noise reduction. All algorithms produce a filtered image and a motion mask. One of the algorithms is based on block-matching, one on fuzzy logic and one on low-rank matrix completion. The final two are much simpler. The first of these estimates the standard deviation of the noise and thresholds the grey-level differences based on that. The last one is a novel approach, relying on spatial smoothing to create motion masks. The generated motion masks are then used to determine what parts of the image that can be filtered temporally. Temporal filtering is done using weighted averaging. Matlab is used for implementation and evaluation of all algorithms. Some are then implemented in OpenCL for testing in real-time using actual cameras. The methods have been analysed regarding the quality of the filtered output as well as the accuracy of the generated motion masks. This has been done using a synthetic image sequence, where a noise-free reference exists. Real-world captured sequences have been judged subjectively. The results indicate that the two simplest methods are the most efficient and the best of the evaluated algorithms for light and moderate noise. For images corrupted by heavier noise, the method based on low-rank matrix completion seems the most promising