| Summary: | 碩士 === 大同大學 === 資訊工程學系(所) === 95 === In this paper, we designed a fast intelligent visual surveillance system installed in front of a public entrance. The main functions are to extract moving objects like pedestrians, to track their locations, and to determine if any abnormal behaviors like wall climbing and falling happened. To save computational cost, frame difference is utilized to produce motion masks which indicate moving regions. By taking both time difference and background difference into consideration, illumination effects can be greatly reduced. Usually, in real situations, the raw motion masks are fragmented and may contain a significant amount of holes inside. By referring to original frames to fill the holes, we can obtain much more reliable motion masks. Then, connected-components are used to extract motion masks. However, some motion masks are connected due to occlusion. As long as those objects are not fully occluded, they can be segmented by proposed multi-modal thresholding on vertical projection of motion masks. Location estimation and weighted block-based matching are combined for the purpose of object tracking. The weight calculated according to the amount of overlapping pixels is assigned to each block. Measurement for similarity is then defined to recognize semi-rigid objects like human. According to the experimental results, the moving objects can be extracted and tracked accurately by means of proposed methods. Occlusion examples are also given to demonstrate the robustness of our system. Finally, motion masks are analyzed by size, position, time and horizontal projection to classify whether they stop, disappear, climb, or fall.
|
|---|
