Visual surveillance system in the mobile-based platform is one of the most active research topics both in computer vision and in mobile robotics. Among them, an abandoned object detection problem in the public place has been significantly dealt with in order to prevent latent terrors and accidents. Previous researches related to the abandoned object detection have been frequently studied on a static camera that uses difference information between images.
This thesis concerns an abandoned object detection based on a mobile robot platform. For this purpose, a mobile visual surveillance system is proposed and implemented. Firstly, camera motion is analyzed. Generally, mobile robot has specific motions such as pitch, yaw, roll and dolly. Using optical flow, camera motions are estimated and are compensated using the image difference between background image and current image.
It is insufficient to extract the foreground image which is regarded as an abandoned object with just simple difference information between current image and background image in database. The proposed method employs image segmentation technique in advance to the main algorithm to get a better foreground image. To explain in detail, each segmented image with proper size is to be compared with an background image in database to find a matching part. If there is no matching segment, it is chosen as an abandoned object candidate. The matching criteria are the ratio of the number of corresponding feature to the total number of features. Among the abandoned object candidates, some candidates are regard as abandoned objects which have higher similarities measured by image difference.
The proposed framework implemented and verified in mobile robot which was developed for educational and entertainment purpose a house.