Research On Highly Realistic 3D Terrain Construction Technology For Level D Flight Simulator Visual System

Recently,the rapid development of the civil aviation industry has brought great challenges to the operational safety of air traffic.As the essential equipment for the airworthiness training of pilots,the Level D flight simulator plays an important and indispensable role in improving flight safety.The terrain data is able to describe the texture of the earth's surface on a global scale,which is highly compatible with the characteristics of flight simulation training.Therefore,the real terrain data is naturally applied to construct the visual system of the flight simulator,which simulates the real-time scene outside the aircraft cockpit to provide effective visual information for building a realistic environment of the flight operation during the pilot training.The 3D terrain is fundamental to simulate the flight environment,in which a real-time and high-quality terrain simulation makes great contributions to improve the effectiveness,fidelity and immersion of the simulator scene.In this work,the approaches for improving the quality of the terrain data are studied to meet the requirements of the visual system of the Level D flight simulator,aiming at building a realistic and immersible simulation environment for pilot training.The details are summarized below:1)By analyzing the defects of preprocessing the terrain data and building the 3D terrain,a novel framework is proposed to preprocess the terrain data in an automatic manner for the Level D flight simulator.The proposed framework comprises basic data management,the digital elevation model(DEM)preprocessing,remote sensing image preprocessing,data quality evaluation,and terrain construction modules.In succession,the requirements of each module are analyzed from the perspective of the real-world application.This work mainly focuses on the dehazing and super-resolution of the remote sensing image,DEM preprocessing and 3D terrain scheduling rendering,so as to lay a preliminary foundation to implement a high realistic 3D terrain.2)To solve the problems of color distortion and poor visibility caused by the hazy terrain data,this work firstly focuses on the dehazing of the remote sensing image.To this end,an encoder-decoder based architecture is proposed to achieve the dehazing task by considering the local contextual information.In the proposed model,both the pixel-to-pixel and the region-to-region conversions are utilized to extract the low-level and high-level features to enhance the quality of the remote sensing image.A multi-scale convolutional block is proposed to meet the multi-resolution requirement of the visual system.Similarly,the attention mechanism is also studied to mine the local semantic features from diversified terrain and landform.In addition,the model architecture is also improved to achieve better performance,including the group normalization,FRe LU activation,and multi-task loss function,etc.Experimental results on the real-world dataset demonstrate that the proposed model is able to achieve superior performance on both the remote sensing image and natural scene image dataset,which further supports the construction of the 3D terrain for the visual system of the flight simulator.3)To cope with the low-accuracy terrain and poor discrimination of reference information caused by low-resolution remote sensing images,a generative adversarial network(GAN)based super-resolution framework is proposed in this work.To achieve the model optimization,a new dataset is built to achieve the super-resolution task for the visual system of the flight simulator,called Airport80.In the GAN-based super-resolution framework,the generator is finally applied to reconstruct the high-resolution image from its low-resolution counterpart,while the discriminator is to distinguish the generated images from real samples to improve the final performance.This work also focuses on modifying the neural network architecture of the generator model.The attention mechanism is applied to extract low-frequency terrain information,which improves the generalization for diversified terrain and landform.The deformable convolutional network-based feature fusion block is proposed to learn the deformable receptive fields to focus on the high-frequency details,such as edge,contour,and other details.Experimental results on the newly constructed dataset demonstrate that the proposed approach can enhance the resolution of the remote sensing image in an accurate and realistic manner,which further meets the requirements of the visual system of the flight simulator.4)Based on the procedure of constructing 3D terrain for the visual system of the flight simulator,the whole pipeline is implemented in this work to validate the practical application of the proposed pre-processing framework.Firstly,the DEM is re-matched to improve its accuracy based on the airborne navigation database.In succession,an Infini Band-based synchronous terrain scheduling rendering architecture is proposed to promote the high-performance rendering engine to further improve the realism of the terrain.Finally,to achieve the high-performance rendering for constructing the 3D terrain in a global manner,the scheduling method of massive terrain data,high-precision texture rendering method and the coordinate accuracy control method are applied to improve the performance.At the end of this work,the proposed framework is validated on a real B737-300 Level D flight simulator,and the results confirm the efficiency and effectiveness of the proposed approach for building a high-performance database.A realistic visual system is obtained by optimized terrain data,which is also achieved in a real-time application.Meanwhile,the author believes that the proposed framework is able to provide some reference techniques and approaches for the construction of 3D terrain in the visual system of the flight simulator.