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

Juyeop Han, Lukas Lao Beyer, Guilherme Venturelli Cavalheiro, Sertac KaramanarXiv 2025 · DOI · arXiv

Abstract

Autonomous drone navigation in GPS-denied indoor environments remains a significant challenge due to the limitations of passive visual localization. This paper introduces SCREP, a novel framework that combines Evidential Scene Coordinate Regression (SCR) with a perception-aware trajectory optimizer. By quantifying predictive uncertainty through entropy, SCREP enables drones to actively "look" at high-confidence landmarks. Experimental results demonstrate that this active perception approach reduces rotation errors by over 30% compared to standard navigation methods while maintaining real-time feasibility.

Introduction

Navigating complex indoor spaces without GPS requires robust visual localization. While traditional feature-matching methods are common, they often rely on computationally expensive maps that struggle to scale. Scene Coordinate Regression (SCR) has emerged as a powerful alternative, using neural networks to predict 3D coordinates directly from single images. However, SCR is prone to outliers and high latency, and its performance degrades significantly when the camera faces featureless surfaces like white walls.

SCREP addresses these flaws by transforming the drone from a passive observer into an active agent. By integrating uncertainty estimation directly into the navigation pipeline, the system can plan trajectories that prioritize visual information, ensuring the drone always keeps reliable landmarks within its field of view.

System Architecture

The SCREP pipeline is designed for real-time operation on resource-constrained hardware. It consists of four primary modules that bridge the gap between raw visual data and optimized flight control.

The SCREP system architecture, illustrating the flow from UAV sensors through E-SCRNet and Pose Estimation to the Trajectory Optimizer.
  1. E-SCRNet (Evidential SCR Network): Utilizing evidential regression, this network outputs a probability distribution for each pixel rather than a single coordinate. This allows the system to distinguish between aleatoric uncertainty (data noise) and epistemic uncertainty (model ignorance) in a single forward pass.
  2. Entropy-Based Cost Function: The system computes the total entropy of the predictions. Low entropy indicates high-certainty landmarks, which the planner uses as "visual anchors."
  3. Perception-Aware Trajectory Optimizer: Using a receding-horizon optimizer based on B-splines, the system generates paths that satisfy kinodynamic constraints while adjusting the drone's yaw to focus on low-entropy regions.
  4. Fixed-Lag Smoother: To mitigate SCR latency, a smoother fuses high-rate IMU data with the more accurate but slower SCR pose estimates, providing a continuous and stable pose for the flight controller.

Active Perception and Trajectory Optimization

The core innovation of SCREP lies in its ability to perform active yawing. Most autonomous planners point the camera forward along the velocity vector. In contrast, SCREP evaluates the environment's "informativeness" in real-time.

A perception-aware trajectory showing the drone adjusting its yaw (blue frustums) to track low-entropy, high-certainty landmarks in the environment.

As shown in the figure above, the optimizer steers the camera toward areas with low normalized entropy (represented by darker points). This behavior ensures that even if the shortest path to a goal involves passing a featureless wall, the drone will rotate its camera toward a known corner or textured surface to maintain localization accuracy.

Experimental Results

The framework was evaluated using photorealistic simulations in the FlightGoggles environment and validated with Hardware-in-the-loop (HITL) testing. The researchers compared SCREP against several baselines, including standard forward-facing cameras and Fisher Information-based planners.

Key performance metrics include:

  • Localization Accuracy: SCREP achieved a reduction in translation Root Mean Square Error (RMSE) of at least 4.9% and a substantial 30.8% reduction in rotation RMSE.
  • Robustness: In challenging trajectories like the "ampersand" test, SCREP maintained a translation error of ~18cm, significantly outperforming baselines that drifted up to 26cm.
  • Computational Efficiency: The entire pipeline—from inference to optimization—runs in under 100ms, making it suitable for high-speed autonomous flight.

Conclusion

SCREP demonstrates that uncertainty is not just a byproduct of neural networks but a valuable signal for robotic navigation. By leveraging evidential learning for real-time entropy estimation, drones can navigate GPS-denied environments with unprecedented reliability. Future work will focus on integrating dynamic obstacle avoidance and extending the system to handle non-static environments with changing lighting and moving objects.

Summary updated

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.
BibTeX
@misc{han2025screp,
  author = {Juyeop Han and Lukas Lao Beyer and Guilherme Venturelli Cavalheiro and Sertac Karaman},
  title = {{SCREP: Scene Coordinate Regression and Evidential Learning-based Perception-Aware Trajectory Generation}},
  year = {2025},
  eprint = {2507.07467},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.07467},
}
Figure