Research On Stereo Matching Algorithm And Evaluation Method Of Martian Scene Based On Virtual Reality

In 2021,our country has carried out its first Mars exploration mission.The complex and changeable terrain of the Martian surface poses a great safety risk to the Mars rover.The rover mainly relies on binocular stereo vision for environmental perception,and the accuracy of the stereo matching algorithm is an important guarantee for the safe driving of the rover.The traditional stereo matching algorithm does not work well in the Martian terrain,while the deep learning-based stereo matching method has poor generalization to unknown scenes.In addition,there is currently a lack of real measurement data of the Martian environment as an algorithm evaluation standard.Therefore,in order to evaluate and realize the high-precision Mars stereo matching algorithm,the work done in this paper is as follows:(1)In view of the lack of real measurement data as the basis for algorithm evaluation,and the deep learning network needs a large amount of real-world binocular data for training,this paper builds a Mars realistic terrain scene based on virtual reality technology,and projects the three-dimensional virtual Martian terrain scene into two-dimensional terrain through MVP transformation.The real disparity map is obtained through the conversion relationship between the depth map and the disparity map,which is used as the evaluation standard of the binocular matching algorithm in this paper and the source of the data set for algorithm training.(2)There are many repetitive texture areas in the Martian terrain scene,and the colors are single and similar.It is difficult for traditional stereo matching methods to process such data.Aiming at this problem,a stereo matching algorithm based on pyramid disparity fusion of Mars scene is proposed,which optimizes the matching cost by fusing low-resolution disparity information in the cost calculation part,and reduces the blurring matching problem of single use of Census transform value as the matching cost.(3)Aiming at the problem of poor generalization of the binocular matching algorithm for Martian terrain scenes and the inability to obtain real disparity labels for algorithm training,a self-supervised learning Martian stereo disparity estimation network is proposed,which can be performed without the need for real disparity labels.The algorithm training greatly improves the generalization of the algorithm.In order to verify the effectiveness of the two proposed stereo matching algorithms,the binocular data of the Martian terrain scene with real disparity maps generated based on virtual reality are used for evaluation.The experimental results show that the stereo matching algorithm based on the pyramid disparity fusion of the Mars scene has a better matching effect on the Mars scene than other traditional stereo matching algorithms;the stereo disparity estimation network based on self-supervised Mars scene training uses binocular data-set of Mars terrain scene for self-supervised training,and the effect is significantly improved.