Research On UAV Ground Target Recognition And Landform Image Updating

The problems of illegal construction and geological disasters have always been concerned.However,the efficiency of traditional manual inspection is low due to the limitation of human resources and information access.Fortunately,Unmanned Aerial Vehicles(UAVs)are widely used in target inspection for their advantages such as flexibility,high resolution and high efficiency.In this thesis,it is focused on the recognition of dynamic targets such as pedestrians and ground vehicles,and the method of geomorphic image updating based on the aerial photography of the multi-rotor UAVs.Aiming at the target recognition using multi-rotor UAVs,an improved learning algorithm is proposed based on the classical network YOLOv3.Although YOLOv3 has high speed and precision,it has too many parameters and takes a long time to load the model in practical applications.Therefore,the classical YOLOv3 used in the UAVs with limited computing resources has to be further improved,and it is necessary to compress the network model to guarantee the real-time capability of the system.In this thesis,based on the python implementation of YOLOv3,the sparse training and model pruning are performed for the network,and the size is compressed to achieve the lightweight goal with satisfied detection accuracy.In addition,GIoU,mAP and F1 are introduced to evaluate the performance of target detection in the test of the improved algorithm and the corresponding pruning model.For the updating of geomorphic image,the orthophoto images are firstly obtained through the aerial photography of the multi-rotor UAV,and the designated task of the aerial survey area is manually performed using the ground station.After setting the area and shape of the aerial survey,the UAV is guided to carry out the inspection task according to the given trajectory.Then,the images with digital elevation information can be embedded into satellite maps,the visual feature comparison is performed between the detected images and the original satellite maps.On this basis,the terrain video frame collected by UAV is matched with panorama by FL ANN homography matching device,and then the collected image is updated to the panorama generated by the last task.The research on UAV ground target recognition and landform image update has realized the ground target recognition and landform orthophoto successive update.In this thesis,a high-definition visual camera on multi-rotor UAV is used as the experimental platform to detect the dynamic targets such as pedestrians and ground vehicles,and the improved model of the network YOLOv3 can enhance the efficiency of the target recognition.By contrast experiment,the model training and testing time are reduced by about 57%to verify the correctness of the algorithm,and the detection accuracy is almost the same.In addition,the panorama matching experiment shows that the method proposed in this paper can be used to update the landscape image.