Research On 3D Martian Terrain Reconstruction Method Based On Monocular Image

With the development of Mars exploration technology,rovers from various countries have successfully landed on Mars,opening the exploration of the Martian surface.The rover is expensive and difficult to set up multiple at the same time for distributed exploration,which reduces the efficiency of information collection.To improve the efficiency of information collection,multiple low-cost small robots can be arranged for exploration tasks at the same time.In the process of detection,the robot’s environmental perception greatly affects the quality of the detection task.Due to the size problem,small robots are difficult to carry too many sensors for environmental perception.At the same time,they do not have enough computing power and storage capacity to process and store complex sensor information.Using a monocular camera as the environment perception sensor of a small robot,only a single monocular scene image is processed and stored,which can reduce the robot’s load burden and reduce the cost of the robot.Among the existing methods of environmental perception using the monocular camera,there has been a relatively mature development in structured scenes such as indoor environments and urban roads.It is difficult to transfer the perception scheme of structured scenes to unstructured ones in unstructured natural scenes due to the irregularity of landforms and obstacles.Environmental perception requires obtaining as much scene information as possible from sensors.Monocular images are rich in scene color information and texture information,but lack terrain semantic information and depth information.The above information can be obtained from monocular images by using the deep learning method,but the improvement of network accuracy is accompanied by a decline in real-time performance.Therefore,considering the environmental perception requirements of small robots,this paper studies an environmental perception scheme based on monocular image semantic segmentation and three-dimensional generation,to realize a Mars three-dimensional reconstruction method that takes into account both accuracy and real-time.The key research content of this paper is divided into two parts: real-time semantic terrain segmentation and realtime 3D information generation.The specific research contents are as follows:1.Semantic segmentation method based on Short-Term Dense Concatenate network: semantic segmentation technology can obtain more scene information,such as geomorphic category and geomorphic boundary information.The trained semantic segmentation model can predict the semantic label of each pixel in the scene.The existing models with high accuracy have a large amount of calculation and low realtime performance.According to the requirements of network accuracy and real-time,the Short-Term Dense Concatenate network is used to complete the task of semantic segmentation.Short-Term Dense Concatenate network adds guidance modules in the training process and removes these modules in reasoning.Through this strategy,the network accuracy is improved and the amount of network calculation is reduced.2.3D terrain generation method based on dual encoder pix2 pix network:conditional generation countermeasure network combined with supervised training mode can complete the specified generation task through specific input.Using the powerful generation ability of the conditional generation network,the corresponding three-dimensional information is generated by inputting monocular images into the network.In the process of network training,the discriminator network judges the authenticity of the output result of the generator network and uses the judgment result to adjust the generator network parameters to enhance the generation ability of the generator.After the network training,only the generator model is used for reasoning,and the generator network adopts a lightweight structure to obtain the real-time threedimensional information generator model.3.Construction of Mars plane stereo dataset and environment perception experiment based on semantic segmentation and 3D reconstruction: to realize the environment perception scheme based on semantic segmentation and 3D generation,this paper constructs a Mars plane stereo dataset based on Mars semantic segmentation dataset AI4 MARS.The dataset has matched the monocular close range map of Mars,semantic segmentation map,and depth.The data set is used to train the semantic segmentation network and three-dimensional generation network.The accuracy of the semantic segmentation model can reach 58% m Io U and the speed can reach 152 FPS;The absolute relative error of pixels between the depth map generated by the threedimensional generation model and the ground truth is 0.273,and the root mean square error is 0.454,and the model speed is 150 FPS.The speed of the overall environmental awareness scheme is 7 FPS.