- Large object tracking
- Make use of ‘Image Moments’ – it is useful based on the assumption that the interactive applications only need a very coarse summary of global averages of orientation and position”;
- Two methods: motion-based calculates the differences between frames, shape-based analyze at a frame-by-frame basis.
- Applications: coarse control e.g., controlling the moving directions of a robot, controlling the movement of a game character.
- Shape Recognition
- Orientation histogram – a histogram summarizing the orientations of an image’s edges;
- Example-based applications involved two phases: training and running. In the training process, the user provides the system with one or more examples of a particular hand shape. The computer stores these hand shapes and calculate their orientation histograms. In run time, it compares the orientation histogram of the input hand shape with the pre-stored ones and determine the closest match.
- Motion Analysis
- Usually achieved using optical flow. But given certain application context, the algorithm can be tremendously simplified, e.g., by measuring horizontal/vertical pixel offsetting frequency one can estimate how fast (and how big the range) the object is moving.
- Small object tracking
- Example: how to track a hand that appears in a much larger image. Form a template of what a hand looks like; use the template to scan the image to determine at what location a hand is most likely to occur.