| Summary: | 碩士 === 國立成功大學 === 電腦與通信工程研究所 === 105 === Spaceborne remote sensing videos are important resources for us to observe earth. With the advance of science and technology, the sources from satellites are not just low resolution images but high resolution videos. The objective of this Thesis is to do multi-object tracking with high resolution satellite videos.
The proposed algorithm is separated into two parts: object detection and object tracking. In object detection, continuous mode is used to build the background model. After getting the foreground with the background model, foreground after noise elimination is classified into three possible categories: big objects, small water objects, and small ground objects. For small water objects, a ground mask is built to separate water and ground. For small ground objects, a building mask is constructed to avoid the error that causes by buildings. Then the foreground is further processed through a mathematical morphology filter.
In object tracking, the speeds of objects are calculated to predict their locations. Occlusion is a problem of object tracking. When the occluded objects are detected, local extremums are computed to separate objects.
|
|---|
