| In this thesis, there are three procedures of video object detection having been studied: object discovery, segmentation and classification. The main work of this thesis are as follows:Firstly, a few example systems on video object deteciton research from overseas and domestic are introduced, then the mainstream and evolution of motion detection, image segmentation and feature based classification methods involved in detection.Secondly, for object detection under complex background, this thesis made comparisons among MOG, codebook model, and frame difference detection methods. Meanwhile it proposed a new noise dropping method based on the silhouette structure differences between noise and foreground object. It alse introduced a morphologic gradients based method to discard foreground shadow of objects. Then the moving object discovery work finished.Now the available motion detection methods can hardly avoid producing incomplete binary foregrounds. To draw objects completely and precisely, this thesis used K-means clustering algorithm to fixup objects foregrounds, making them closer to objects'real appearance. The fixed binary foregrounds are used as proof when judging whether each watershed-segmented fragments belongs to foreground objects or background. Finally, it tells exactly where an object is.Intelligent Video Surveillance Technology comstomarily focuses itself on the tracking and analysis for some certain kind object, so it is necessary for video detection to classify the moving objects. Taking pedestrian for example, this thesis studied classification methods based on statistical machine learning theory using invariant image features. A pedestrian recognition experiment was carried out by HOG descriptor and SVM classifier.This thesis established a verifying detection system also based on codebook background model and SVM using HOG descriptor, and showed its corresponding results. In the end, this thesis made summary and outlooks for video object detction algorithm. |
PDF Full Text Request