Research On Recognition Algorithms Based On Unmanned Aerial Vehicle Image

Since the increase in the air vehicle violations and the improvement of the air traffic management,the automatic analysis of specific environmental traffic activities becomes the top priority.The method based on computer vision has aroused wide concern in many air traffic monitoring technology.The UAV(Unmanned Aerial Vehicle)identification is a research direction in the field of computer vision,which has the extensive research significance and value in both military and civilian aspects.The UAV identification technology is mainly to analyze and deal with the images captured by the monitoring system and identify the UAV target correctly.This paper mainly studied the UAV recognition algorithm design and implementation based on the shallow machine learning and depth model of the convolutional neural network model,so as to fulfill the task of UAV image classification and testing to identify.The research on the UAV identification mainly includes the feature extraction and the classifier design.This paper presents a method based on the H(Histogram of riented radients)feature of target identification algorithm and the LBP(Local Binary Pattern)to extract the image feature,combined with the SVM(Support Vector Machines)algorithm to realize the image recognition and classification.The experiment shows that the target recognition algorithm based on H feature,LBP operator and the image shape feature can classify the UAV target accurately.In addition,the influence of the flight attitude and the light intensity on the target recognition of the UAV is also analyzed in this paper.In the complex scenarios,the UAV detection is also a challenging task.The paper presents a cascade deep multi-task network framework,which effectively combines the UAV detection and the wing terminal feature points,and uses the internal relation between them to improve the recognition performance.The framework consists of two convolutional neural networks: a shallow full convolution region recommendation network and a precise output network for the extraction of the UAV.The whole network framework is trained separately,but it is completed by two networks in the test phase.Firstly,generate a set of candidate boxes for the input image with the full convolution shallow network region recommendation algorithm and the candidate boxes is determined through local non-maximum valuesuppression and border regression.Secondly,preliminarily screen the candidate box,and obtaining a small number of candidate boxes.At last,input them into the second stage of deep network refinement processing,obtaining the optimized test results are and the wing marks are output.The experimental results show that the model proposed in this paper can detect the different attitude of the UAV and solve the problem of occlusion.At the same time,the UAV with a certain Angle and size can also accurately locate its wing position and have good recognition performance.