Object region detection algorithm has been studied in many fields of computer vision. It is essential for surveillance systems, tracking systems and recognition systems. This thesis proposes an algorithm for detecting object region in image sequences using multi-sensor fusion and background estimation. The proposed algorithm includes following topics.
First, detection using background estimation is supplemented via a new adaptation rule. Background estimation algorithm is widely used because it is robust to illumination changes and is not sensitive to the velocity of objects. But it is sensitive to background noise because not considering spatial correlation in an image. So this is recovered through the techniques such as estimation using two thresholds and update with two adaptation rates.
Second, multi-sensor fusion scheme increases capacity of detection. For example, when the color of objects is similar to that of background, the method using CCD sensor doesn't work well. But, it is possible to solve this problem using fusion with another sensor like IR one.
Third, dynamic sprite is used to apply detection algorithm using background estimation to the movable camera. When the camera moves, the global motion compensation must be done at first. But it is not enough for detection because newly appeared region doesn't have any stochastic information. So dynamic sprite is used for fast detection using previous stochastic information when the camera moves to the previous position again.
Finally, edge-based segmentation is applied to the result of detection. More exact object region is extracted with simple cost function because the result of detection is roughly segmented.
Simulation results show that the proposed algorithm detects object region successfully even with disguise and movable camera.