Target Recognition And Processing Based On Quadrotor Aircraft

In recent years,drones have become more and more noticed by people.Drones can be used to detect enemy targets and launch bombs in military field.What’s more,drones also can be use for aerial tracking in our daily life,the realization of these tasks can not be separated from the detection and identification of the target.Traditional target recognition relies on corners,textures,and color detection to distinguish targets.This kind of method has lower requirements on the equipment and less calculation,but it has low recognition rate and is easy to be interfered.It is difficult to use this method for target recognition on an aircraft.With the development of deep learning theory and the increase in computing power of equipment,neural networks can be applied to various fields.In the field of computer vision,convolutional neural networks are often used to classify objects.Therefore,a series of target recognition algorithms based on traditional methods and neural networks are compared in this paper.Finally,the experimental scheme for target recognition and processing on quadrotor aircraft is determined.In this experiment,a four-rotor aircraft was used as an experimental platform,and a visual image processing system was set up on the platform.At the same time,two kinds of experimental schemes are designed.The first one is the feature point matching method based on traditional image processing.The purpose of target recognition is achieved by using certain features of the target image itself to match in the acquired new image.This article compares the three commonly used matching methods.SIFT,SURF and ORB.then selected ORB as the experimental method of the program.The second one is based on the CNN convolutional neural network identification scheme,which can be constructed using the Caffe framework.The implementation of this solution requires the following two steps:candidate area screening and target classification.For candidate region screening,FCN,RCNN series and SSD models were compared,and a candidate region extraction method based on improved SSD network model was determined.AlexNet was used to classify the image and design corresponding image classification network.The target in the candidate area is further identified to improve the recognition accuracy and improve the classification effect.Then the paper conducts an experimental study on the above two schemes.In the GPU mode,for the first ORB feature point matching scheme,it can process images of about 10 frames per second,and the target detection height range is between 1.5m to 5 m.The accuracy is about 85%.The second method is based on the convolutional neural network identification method.The improved SSD model is used to select candidate regions.The recognition height is between 1.5m to 10m,the image processing speed per frame is 35ms,and the target recognition accuracy is about 97%.After classifying images with AlexNet,the recognition accuracy is above 97.2%,and the processing speed is around 20ms.By comparing the experimental results,in the complex high-altitude environment,the first scheme lacks a certain recognition rate,robustness,and real-time performance.The processed data can only be used as an auxiliary reference and is applicable to some inexpensive aircraft that do not have GPU support.The second method based on the convolutional neural network has higher recognition rate,stronger robustness,and certain real-time performance,The innovation of this paper lies in the application of the improved SSD model in the recognition of aircraft landmarks.It is proved that the target recognition scheme based on convolutional neural network can be used in fields with certain requirements on real-time performance.which is greatly improved compared with the traditional algorithm.