Research On UAV Visual Tracking Under Low Visibility

As a kind of intelligent aircraft,Unmanned Aerial Vehicle(UAV)has a wide range of application prospects,and has been widely concerned in civil and military fields.At present,the rapid development of computer vision technology has greatly improved the ability of UAV to complete various tasks through visual system.Among them,object tracking technology has become the core means for UAV to perceive and understand the scene through vision.However,the performance of visual system can be disturbed due to the bad weather such as fog and rain,and it is difficult to accomplish the object tracking task if it just depends on the low quality image signal.In order to achieve a better object tracking task in bad weather,an UAV-oriented visual tracking method is presented in this paper.Based on the pretreatment of low quality image data,the UAV navigation parameters are introduced as the assistant means of environmental perception to improve the performance of UAV visual tracking.From this,the main research work and results are as follows:(1)Aiming at the low resolution、loss details、poor quality of aerial images under fog,a new aerial image haze removal method based on deep learning is proposed.Firstly,the atmospheric scattering model is introduced to explain the degradation mechanism of aerial image,then auto-encoder network is designed to extract the haze-related features of the image by unsupervised learning mode,and a convolution neural network is designed to regress the scene transmittance corresponding to the haze-related features,and lastly the clear aerial image is restored based on the scene transmission.A database consisting of multiple scene foggy images and corresponding transmission maps is established as the training set,and the effectiveness of the proposed method is validated by the experimental results of indoor、outdoor and aerial images.(2)Aiming at the low precision of navigation system,a visual aided UAV navigation method is proposed.Because of the occlusion of satellite signals and the accumulation of Inertial Measurement Unit(IMU)errors over time,the traditional navigation method is difficult to meet the accuracy and flexibility of the navigation system.It is an irreplaceable advantage to introduce machine vision technology into navigation system as an auxiliary method of error correction.Firstly,the classical Lucas-Kanade optical flow method is improved,and the UAV speed is estimated by epipolar geometric constraint,which is used as the observed value for the next fusion algorithm.Then a nonlinear observer model is designed to fuse the estimated velocity with satellite and inertial sensor data.Finally,the accurate position、speed、posture and so on are obtained.The feasibility and effectiveness of the proposed method are verified by the experimental results of simulation data and the Zurich-urban flight data.(3)Based on the reconstructed aerial image and the modified navigation parameters,an UAV target tracking algorithm for low visibility environment is proposed.Firstly,the theoretical optical flow of the scene is estimated by using the navigation parameters,and the motion object is detected by the difference between theoretical optical flow and real optical flow.Then,a tracking algorithm based on discrete Kalman filter is developed to track the object’s trajectory and transform it into the world coordinate system.The effectiveness of the proposed method is verified by the experimental results of simulation three dimensional data and surface unmanned ship voyage data.(4)The UAV hardware platform is built and the UAV object tracking test under low visibility is designed.The effectiveness of the tracking method in low visibility environment is validated by the experimental results,and the influence of haze removal method on navigation system and the effect of navigation correction on object tracking are analyzed.