A probabilistic approach for video stabilisation in compressed domain

Machine vision systems, which are being extensively used for intelligent transportation applications, such as traffic monitoring and automatic navigation, suffer from image instability caused by environment unstable conditions. On the other hand, by increasing the use of home video cameras which sometimes need to remove unwanted camera movement, which is created by cameraman shaking hands, video stabilisation algorithms are being considered. The video stabilisation process consists of three essential phases: global motion estimation, intentional motion estimation and motion compensation. Motion estimation process is the main time consuming part of global motion estimation phase. Using motion vectors extracted directly from MPEG compressed video, instead of any other special feature, can increase the algorithm generality. In addition, it provides the facility for integrating video stabilisation and video compression subsystems and removing the block matching phase from video stabilisation procedure. Elimination of any iterative outlier removal preprocessing and adaptive selection of motion vectors has increased speed of the algorithm. Although deterministic approaches are faster than the related probabilistic methods, they have essential problems in escaping from local optima. For this purpose, particle filters, the ability of which is considerable when submitted to non-linear systems with non-Gaussian noises, are utilised. Setting the parameters of the particle filter using a fuzzy control system reduces the incorrect intentional camera motion removal. The proposed method is simulated and applied to video stabilisation problem and its high performance on various video sequences is demonstrated.