Research On The Key Technology Of Single Optical Aperture UAV Detection-tracking-recognition System

In recent years,the development of UAV is very rapid.All kinds of low-cost small UAVs have strong professional functions such as flight,transportation,remote sensing,shooting,etc.,which have brought great changes to people’s production and life.However,the incidents of UAV illegally flying frequently arise,which brings serious security risks to the country and the people.So the effective supervision of UAV has become a hot research field.A complete UAV supervision task mainly involves three parts: detection,tracking and identification.The existing technical schemes realize all functions through the interaction of multiple devices,resulting in a very complex system and high cost.Therefore,this thesis proposes a single optical aperture UAV detection,tracking and recognition system,which integrates the detection,tracking and recognition tasks into a single aperture photoelectric device,and realizes the integration of all functions of the system.More specifically,the system uses a variable focus ball machine with a two-axis rotating platform.Firstly,it detects the small moving targets of the suspected UAV in the sky in the wide-angle field of view under the low magnification state of the camera.Secondly,the platform attitude control technology and camera level zoom technology are used to locate and magnify the detected target in real time.Finally,by optimizing the tracking algorithm and recognition algorithm,the target can be quickly tracked and accurately classified,and the alarm can be sent out at the same time.The key technologies involved in this thesis mainly include UAV moving small target detection,target tracking under dynamic camera and UAV target recognition.The main research contents are as follows:Aiming at the detection of small moving targets,this thesis proposes a two-stage small moving target detection algorithm based on Extraction and Evaluation.Firstly,in the process of foreground information extraction,an improved Gaussian mixture background model is proposed,which can adaptively adjust the sensitivity of target detection according to the complexity of background and suppress a lot of noise;Secondly,in the stage of target trajectory evaluation,the extracted continuous multi frame foreground information is used to cluster and fit the target trajectory,and the trajectory with the highest confidence is selected as the detection target to further suppress the discrete point noise.This method divides the detection of small moving objects into two stages: foreground extraction and trajectory evaluation.It has double suppression of noise,strong anti-interference ability and robustness,and can be applied to complex urban low-altitude scene.Aiming at the target tracking under the dynamic camera,this thesis proposes a tracking algorithm of UAV based on PTZ cooperative control.In the whole tracking process,the target,camera and background are all in dynamic changes,which makes the tracking algorithm have higher requirements for accuracy,real-time and hardware control speed.On the one hand,a tracking algorithm based on kernelized correlation filter with adaptive matching region is proposed,which improves the tracking stability of the system;on the other hand,this thesis introduces real-time attitude control technology and adaptive hierarchical zoom technology for two-axis rotating platform and camera respectively,which ensures that the UAV is always in the center of vision with appropriate scale,making the tracking process fast,stable and accurate.For UAV target recognition,this thesis adopts the UAV target recognition algorithm based on Lightweight Networks and Transfer Learning.In order to solve the problem that the sample size of UAV data set is insufficient,the Transfer Learning method is used to ensure the effective learning;in accordance with the requirements of the actual system for memory,the Lightweight Networks structure is used to train the classification model to reduce the memory occupation of the model.Experiments indicate that this approach can effectively learn UAV target under insufficient sample set,and has high recognition accuracy.At the same time,the model occupies less memory,which meets the needs of flexible deployment of the system.