Surveillance Video Abstract Extraction And Display Technique

Surveillance cameras have been installed all over every corner of the city, protecting our safety. However, the large amount of video data would cause big trouble on video browsing and querying. Video abstraction could compress a lengthy surveillance video into a short one, while keeping most of the original video information and presenting users in a hierarchical type. This thesis made some researches on dynamic background reconstruction, moving objects detection, moving objects tracing in the surveillance video, together with the highly efficient displaying of multiple objects in space and time.For background reconstruction, to improve reconstruction speed and accuracy, this thesis took the mean value method of and designed an adaptive way to automatically adjust the learning rate, that is, to select suitable learning rate based on the object area in current frame. Through experiments, background images reconstructed with the improved scheme is more pure than the fixed learning rate methods, and the computing cost is not obviously increased.For object detection, this thesis selected the background differentiating method. The object features can be obtained by connected component labeling after binarizing the differentiating image and morphological filtering. This paper proposed a multi-scale connected component labeling algorithm, which can detect object with different scales on different levels. It could find objects which are not too small quickly so is suitable for our task.For object tracing, based on the standard CamShift algorithm, this thesis proposed a new tracing method based on motion consistency. Since one object could easily be broken into multiple sub-objects in the object detection step, which will effect tracking accuracy and speed. Merging the sub-objects with strong movement consistency and spatial adjacency, we could avoid the problem of “one object, multi tracing”, and alleviate the influence of incomplete objects.To display multiple objects more efficiently in space and time, this thesis wouldn’t display all frames of the objects as before. We only display the most representative patterns of the video. Thus, the video displayed will be simpler and clearer, and the compression ratio will be higher.The system designed and implemented in this thesis achieves the main functions of surveillance video abstraction and displaying, with satisfying results.