SCREP: Scene Coordinate Regression and Evidential Learning-based Perception-Aware Trajectory Generation

TL;DR

This work presents a perception-aware trajectory planner that couples an evidential learning-based SCR poseestimator with a receding-horizon trajectory optimizer, and hardware-in-the-loop experiments validate the feasibility of the proposed trajectory planner under close-to-real deployment conditions.

Abstract

Autonomous flight in GPS-denied indoor spaces requires trajectories that keep visual-localization error tightly bounded across varied missions. Map-based visual localization methods such as feature matching require computationally intensive map reconstruction and have feature-storage scalability issues, especially for large environments. Scene coordinate regression (SCR) provides an efficient learning-based alternative that directly predicts3D coordinates for every pixel, enabling absolute pose estimation with significant potential for onboard roboticsapplications. We present a perception-aware trajectory planner that couples an evidential learning-based SCR poseestimator with a receding-horizon trajectory optimizer. The optimizer steers the onboard camera toward reliablescene coordinates with low uncertainty, while a fixed-lag smoother fuses the low-rate SCR pose estimates with high-rate IMU data to provide a high-quality, high-rate pose estimate. In simulation, our planner reduces translationand rotation RMSE by at least 4.9% and 30.8% relative to baselines, respectively. Hardware-in-the-loop experiments validate the feasibility of our proposed trajectory planner under close-to-real deployment conditions.

Figures (7)

• Figure 1: Top‑down 3‑D reconstructions of the two sites produced by SCR. (left) RGB-colored points (right) Normalized entropy map. The accumulated scene coordinates and their uncertainties from multiple viewpoints reveal spatially consistent uncertainty distributions across the environment.
• Figure 2: Illustration of the proposed receding-horizon perception‑aware trajectory generation. (blue) Camera orientations are optimized to steer the camera toward low-entropy coordinates. (red) Position trajectory (black) Current pose.
• Figure 3: System architecture. (1) A pose estimation block consisting of E‑SCRNet, PnP‑RANSAC, and a fixed‑lag smoother; (2) A two‑stage trajectory optimization block for position and yaw.
• Figure 4: Plots showing normalized uncertainty and scene coordinate error of E-SCRNet predictions. (left) Mean and standard deviation of L2 error (right) Density histogram.
• Figure 5: Qualitative results of E-SCRNet for a test image. (upper left) Aleatoric Uncertainty (upper middle) Epistemic Uncertainty (upper right) Entropy (lower left) L2 Error of predicted scene coordinates (lower middle) Predicted scene coordinates (lower right) Input image.