Precise visual localization and terrain reconstruction for China's Zhurong Mars rover on orbit

During the three month movement of China's Zhurong Mars rover, it is a very important task to obtain the precise position of the rover and the terrain around the stations, which provides basic information for subsequent path planning, mechanism movement and scientific exploration. At the visual localization step, a bundle adjustment model by using the stereo navigation camera (Navcam) images with overlapping areas between adjacent stations can achieve continuous relative pose for the rover. which can effectively reduce the impact of wheel slip and inertial navigation system errors on positioning accuracy. Firstly, the initial position and attitude information of the Zhurong Mars rover and the rotation angle of the mast can be contained from the telemetry data. Then, through the inverse coordinate transformation, the Navcam pose in the landing north-east-down (NED) system can be obtained, which is regarded as the initial value of the Navcam's external orientation element. Secondly. the speeded-up robust features (SURF) matching algorithm is used to match stereo images at adjacent stations, and some sparse connection points are obtained. Thirdly, combined with the Navcam's internal orientation elements and relative pose parameters, a bundle adjustment model is constructed. A least square solution is used in the proposed bundle adjustment model, and the optimized Navcam pose at the current station can be calculated. Then, through the coordinate transformation. the Zhurong Mars rover's position in the NED system is obtained. At the terrain reconstruction step, a visual closed block network adjustment model and the stereo matching technology are used to obtain a digital elevation model (DEM) and orthophoto map (DOM). Firstly, an epipolar line correction method is performed on the Navcam's stereo images at the current camera station, and tie points are obtained by SURF. Secondly, the proposed closed block network adjustment model is constructed. When a total least squares solution is used in the proposed model, the optimized relative pose of the sequence images is calculated. Taking one pair of stereo images as a bridge, the sequence camera pose is integrated into the NED System. Thirdly, dense matching and forward intersection are performed in the current sequence images to obtain dense point clouds in the NED system. After point cloud dcnoising, interpolation and back projection. the DEM and DOM products are obtained, and the terrain reconstruction mission at the current station is completed. According to the ground experiment of the Zhurong Mars rover, the visual localization precision is better than 2%, the terrain reconstruction accuracy is 5 mm@3 m. In the orbit mission_ taking the visual localization results of the Zhurong Mars rover at the end of time months as an example, the cumulative error of the visual localization results is 5.1%. Based on the Navcam images sequence on the lander, Mars DOM images and one image from the descent camera with a height of about 200 m, the proposed technology is used to obtain the precise landing position (109.925 degrees E, 25.066 degrees N) of the Zhurong Mars rover on May 22, 2021. The landing position result corrects the estimated landing position of the rover based on the orbital parameters. The deviation of the two measured positions is about 2.1 km. At the same time, the localization and terrain reconstruction results effectively support on-orbit departure, path planning, and coordinate measurement of scientific target points, etc.