Estimation of Camera Ego-Motion for Real-Time Computer Vision Applications

How can we distinguish the scene motion from the camera motion? In most of the computer vision applications, camera movements avoid true detection of events. For instance, if the camera is trembling due to wind in outdoor environments, it creates high frequency motion in the scene motion as well. To detect flame, we use high frequency motion. If camera trembles, then non-flame regions can also be detected as flame due to high frequency camera motion. Consequently, it is essential to detect the camera motion and avoid event detection (e.g., flame detection) when the camera is moving. In this paper, we have suggested a camera ego-motion estimation algorithm, which uses a feature based approach for motion estimation. Feature selection limits the number of feature points to be tracked and henceforth lowers the computational complexity. Besides, it has a low dependency on structure in the video frame. For video frames with unknown motion fields, displaced frame difference has been used as the criteria for evaluating the algorithm, whereas vector difference between known and estimated displacements has been operated as an error measure for video frames with known motion fields