Deep Learning Based Detection And Localization For Fixed-wing UAV’s Autolanding

With rapidly development of Technologies in Unmanned aerial vehicle（UAV）,UAV has been broadly used in many fields,such as industry,Agriculture,Military and even now,logistics.There is great progress in aspect of stability,loading,and duration of flight,but automatic taking off and landing is still the bottleneck of UAV application.Ground based vision guidance system has abundant computing resources,not limited by mission loading,and,what’s more,vision sensor passively image,has no problem of s_si_tg_un_da_yl _od_fi _this^{sturb ing}_arti._{c l}G_e i _cho^ven _ot_sh_ie_{n g}a _ground^dvantag _ster^es ofs of _eg_or ^{o und}_{visi o} b_n-a_bse_asd_{e d}v ⁱ _guid^sion _ag_nu_ci_ed _syst^ance _e s_my ^s _ass ^tem _s,_{t u}t_dh_ye object,position UAV Space trajectory with EKF filter.The study emphasis is UAV target detection and positioning.Taking advantage of abundant computing resources,large labeled image sources,deep learning algorithm is used as UAV target detection and tracking framework.In order to improve the instantaneity and robust of the detection algorithm,target detection,target tracking and target tracking failure is combine in a whole frame to do UAV target detection,making high real-time performance and strong robustness.The major focus and innovation are as follows:（1）YOLO based UAV target detection and image source auto labeled in simulation environment.Choosing regression based deep learning algorithm as UAV target detection algorithm.The algorithm only look once and the image once,and UAV target position can be regressed.After using deep learning target detection algorithm,difficulty of detection was transferred into deep learning network training and image source labeling.A method of UAV target auto label was used in Xplane simulation environment according to the ground truth of UAV pose and camera pose.（2）Tracking failure judgment based on correlation of targets in image.Correlation of two images reflects the similarity of two images.According to theory of adaptive correlation filters,calculate target tracked at present and target tracked previously,Evaluating target tracking performance based on Peak to Sidelope Ratio （PSR）of correlation output.（3）Robust detection algorithm of UAV targetIn order to improve the instantaneity and robust of the detection algorithm,target detection,target tracking and target tracking failure is combine in a whole frame to do UAV target detection,making high real-time performance and strong robustness.