Global Localization of the Perseverance Rover via Orbiter-UAV-Rover Collaborative Matching

The precise global localization of the Mars rover serves as a fundamental prerequisite for long‑distance scientific traverses and in‑situ geological investigation. As Mars represents a typical GNSS‑denied environment, accurate positioning is typically accomplished through the registration of rover‑acquired imagery with orbital maps. Mainstream methodologies address the substantial perspective and scale differences between ground‑level and orbital images by first generating orthophotos from rover imagery, which are then aligned with satellite‑based imagery for localization.

The successful deployment of the Mars Helicopter (Ingenuity) enables the use of acquired UAV imagery as an intermediate bridge for the rapid and accurate global localization of the Perseverance rover. Accordingly, this study proposes an orbiter-UAV-rover collaborative matching framework, as illustrated in Fig.1. This framework sequentially performs three core steps: (1) cross-perspective matching between rover and UAV imagery, (2) cross-scale matching between UAV and orbiter imagery, and (3) a matching connection strategy that integrates the two matching sets to establish a continuous geometric transformation chain.

Figure 1. Schematic diagram of the proposed global localization framework.

Specifically, the rover-UAV image matching procedure is implemented through the following sequential steps, and the efficacy of this approach is demonstrated in Fig. 2.

(1) Horizon-based Pose Estimation: The visual horizon within the rover image is segmented using a Mask R-CNN model. This horizon line is then analytically processed to derive the pitch and roll angles of the rover camera.

(2) Cross-Perspective Image Rectification and Matching: Leveraging the estimated orientation angles, the rover image is orthographically rectified to approximate a nadir view, thereby aligning its perspective with that of the UAV imagery. A deep learning-based feature matching network is subsequently applied between the rectified rover image and the UAV image to establish dense, pixel-wise correspondences.

(3) Correspondence Projection: The matched feature points from the rectified image pair are back-projected onto their original coordinates in the raw rover image.

Figure 2. Comparison of cross-view feature matching results before and after orthographic rectification.

Following the establishment of correspondences between rover and UAV imagery, the matching results between the UAV and orbital data are subsequently derived using our previously proposed method [1]. This process culminates in the formation of a two-tier correspondence chain, effectively linking the rover, UAV, and orbiter, as visually summarized in Fig. 3.

Figure 3. Visualization of cross-platform feature matching results.

Figure 4. Results of collaborative matching and localization.

Table 1. Localization error of the Perseverance rover for different sites.

Localization experiments were conducted at multiple sites along the Perseverance rover's traverse. As shown in Fig. 4 and Table 1, multi-platform images were well-associated, achieving an average accuracy of 0.4 m (resolution of the orbital image is 0.25m). High-precision rover positioning information enables the precise fusion of multi-site local geological mapping products and ensures the accurate integration of rover and orbital-scale geological mapping products.

 

Reference:

[1]   CAO Z, FU H, XU X, et al. A Novel Template Matching Method Incorporating a Multi-Candidate Region Optimization Strategy for the Initial Localization of Mars Helicopter. Transactions in GIS, 2025, 29(2): e70052.