Target Tracking And Landing Pose Estimation Based On Uav Platform

Autonomous landing of UAV is one of the key technologies in the future,and visual landing as the assistant of inertial navigation and other landing technologies is hotspot in the current research.Traditional UAV visual landing algorithm is designed in the form of traditional iconography and artificial landing icon.This kind of algorithm design has two limitations.On the one hand,the landing icon is not always imaged in an ideal way in the actual environment,and light changes,occlusion and other situations are very common.However,this kind of algorithm is difficult to guarantee its generalization only through the manually designed feature extractor;on the other hand,when the UAV needs to make a forced landing in an unknown area in a special situation,it is only necessary to The landing algorithm of cooperation target which is designed by human is not competent.In view of these two points,this paper studies landing and detection tracking technology in known and unknown environment,mainly including the following contents:1.In the research of UAV cooperative target landing,by introducing deep learning detection algorithm to replace the traditional feature design and classifier algorithm,Cascade R-CNN which can detect both landing cooperative target and vehicle target uses the vehicle position information detected by the neural network model as the prior information of the follow-up vehicle target tracking and matching algorithm,and uses SIFT features to match the template and panorama frame by frame,so as to achieve the tracking of vehicle target.2.In the case of landing of UAV cooperative target,first calibrate the Airborne Camera with Zhang Zhengyou's internal parameter calibration algorithm.Because of the pre detection process and the detection of the specific landing icon position,the original image can be cut,and the corner feature can be extracted from the local image.According to the coordinate system and mapping formula,the unmanned position and attitude can be estimated to provide guidance for landing For reference.3.In the research of UAV's non cooperative target landing,through the introduction of deep learning semantic segmentation algorithm to classify the aerial images at the pixel level,obtain the landing area with categ ory information,select the appropriate category and give different priority.At the same time,according to the current height of UAV,the imaging area of landing area in the image is estimated,the landing frame of corresponding size is set,and the doma in entropy of panoramic image is calculated as the evaluation reference of regional smoothness.Using the idea of sliding window for reference,the landing frame is traversed in the candidate landing area,and the optimal landing frame is evaluated by comb ining the entropy of the field and the position information,which can be used as a reference for the landing of UAV in the position environment.