Research Of Video Search Based On Target Recognition

In recent years,the application of target recognition from video is very extensive,such as video surveillance in public places,license plate recognition in intelligent transportation,and sighting of military weapons.Using an artificial observation of the abnormal movement of an object in the video can be very time-consuming and labor-intensive and the accuracy is not high.Therefore,the video recognition technology based on target recognition has a strong research value.This paper designs and implements a video search system based on target recognition.The user needs to provide video and the target image library that needs to be searched.If the target exists in the video,the corresponding video frame will be finally output.Firstly,using the improved background subtraction method to find the video frames of the scene change as candidate frames;then extract the candidate frame and the sift feature points of the target image library to be searched;finally,the optimized VP forest algorithm is proposed,and the candidate frames are needed.The searched target image library performs feature matching and outputs a successfully matched frame.The main contents are as follows:(1)Video framing,using scene detection methods,extracts scene change frames from the video as candidate frames.Firstly,by analyzing the commonly used scene detection methods,the background difference method is selected for scene detection.Then,an improved background subtraction method is proposed.The Euclidean distance algorithm is used instead of the difference method to judge the similarity between the current video frame and the background template.In the case of many noise points,the accuracy of scene detection is increased.Before the scene detection,the preprocessing of the video frames is increased to reduce the influence of noise points and illumination on the detection accuracy.(2)Extract sift features of candidate frames and target image libraries.Due to the stability and uniqueness of sift features,this paper uses sift feature points for image feature matching.Firstly,the sift feature points of the target image library are extracted to be spared,and then the sift feature points of the candidate frames are extracted in real time for subsequent feature matching operations.(3)Match the candidate frame with the target image library.An optimized VP forest algorithm is proposed.By constructing multiple VP-trees,the self-similarity of the near subtrees in VP-tree is reduced,and the matching accuracy is greatly improved.And in the process of constructing VP-tree,the point with large variance is selected as the advantage point,and the probability of randomly selecting the edge point is reduced.A large number of experiments have proved that the optimized VP forest algorithm performs feature matching under different data set sizes,different data dimensions,different distance functions,etc.The accuracy rate obtained is higher than the current popular random KD forest algorithm.The improved background subtraction method and the optimized VP forest algorithm are applied to the video recognition system based on target recognition,and good results are obtained.