Research On Uav Target Recognition Method Based On Video Stream

With the development of UAV technology,UAV photography has broad application prospects in aerial photography,terrain mapping,military reconnaissance,etc.This paper mainly studies the problem of target recognition in UAV overhead video,analyzes the existing moving target recognition methods,and proposes a moving target recognition method that combines target tracking and image recognition.Taking the moving target in the aerial video of the drone as the research object,tracking calibration and classification recognition are carried out.Aiming at the real-time problem of the recognition method,an improved method based on data de-redundancy is proposed,and the video stream-based UAV target recognition method is verified through simulation experiments.The main research contents of this paper are as follows:(1)Aiming at the real-time problem of target tracking process and target recognition process,a UAV target recognition method based on video stream is proposed,and a tracking and recognition strategy is designed.The UAV target recognition method consists of three parts:target tracking and calibration,target feature extraction and target image recognition;it takes the steps of real-time tracking first,and then classification and recognition to avoid the target image recognition method and the target tracking method from occupying time each other.(2)Aiming at the problem of light interference in UAV target tracking,the video stream-based target tracking method is adopted,the video stream target tracking strategy is designed,and the target tracking is performed through the mean shift algorithm;for the time-consuming problem of target tracking,a fusion key frame is proposed The de-redundant video stream target tracking method,combined with an adaptive key frame algorithm,filters out redundant video frames by setting a threshold.(3)Aiming at the small target problem of target feature extraction,the HOG-based target feature extraction method is adopted,the HOG feature extraction strategy is designed,the sliding window parameters are set,and the HOG feature of the target image is extracted;for the redundancy problem of the target feature,a fusion is proposed The HOG interval optimization feature extraction method of PCA filters out redundant optimization features.(4)Aiming at the time-consuming problem of target image recognition,the target image feature is classified and recognized by the SVM-based target image recognition method;for the problem of de-redundant feature recognition of target image,the target feature based on nonlinear kernel function is used to reduce the dimensionality Recognition method,adjust the kernel function parameters,shorten the time-consuming training of the classifier,and efficiently recognize the target.