Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning-Based Robust Optical Guidance for Hypersonic Platforms

Adrien Chan-Hon-Tong, Aurélien Plyer, Baptiste Cadalen, Laurent Serre

TL;DR

This work investigates robust optical guidance for hypersonic platforms by replacing traditional single-reference image registration with a direct learning approach that maps a stack of scene images to target localization. During training, a CNN learns the mapping $f_w(x)$ from sampled views of reference stacks to corresponding target positions $l$, optimized via the regression loss $L(w,x,l)=(f_w(x)-l)^2$, enabling real-time inference. A compact ConvNext Tiny-based network achieves about $60$ FPS on CPU for $256\times256$ inputs, offering a practical onboard solution. Experiments across weak and strong appearance changes, including bimodal snow/no-snow scenarios and trajectory-based evaluations, show improved robustness over single-reference methods, with clear gains when leveraging multiple reference images, though data collection and target-specific training remain challenges. The approach provides a fast, mode-robust alternative to traditional registration, with potential for real-time guidance in GNSS-denied, high-speed flight, while underscoring the need for further scaling and safety analyses $($e.g.$, stack acquisition, per-target training$)$.

Abstract

Sensor-based guidance is required for long-range platforms. To bypass the structural limitation of classical registration on reference image framework, we offer in this paper to encode a stack of images of the scene into a deep network. Relying on a stack is showed to be relevant on bimodal scene (e.g. when the scene can or can not be snowy).

Deep Learning-Based Robust Optical Guidance for Hypersonic Platforms

TL;DR

This work investigates robust optical guidance for hypersonic platforms by replacing traditional single-reference image registration with a direct learning approach that maps a stack of scene images to target localization. During training, a CNN learns the mapping from sampled views of reference stacks to corresponding target positions , optimized via the regression loss , enabling real-time inference. A compact ConvNext Tiny-based network achieves about FPS on CPU for inputs, offering a practical onboard solution. Experiments across weak and strong appearance changes, including bimodal snow/no-snow scenarios and trajectory-based evaluations, show improved robustness over single-reference methods, with clear gains when leveraging multiple reference images, though data collection and target-specific training remain challenges. The approach provides a fast, mode-robust alternative to traditional registration, with potential for real-time guidance in GNSS-denied, high-speed flight, while underscoring the need for further scaling and safety analyses e.g.)$.

Abstract

Sensor-based guidance is required for long-range platforms. To bypass the structural limitation of classical registration on reference image framework, we offer in this paper to encode a stack of images of the scene into a deep network. Relying on a stack is showed to be relevant on bimodal scene (e.g. when the scene can or can not be snowy).
Paper Structure (9 sections, 4 figures, 2 tables)

This paper contains 9 sections, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Direct Guidance Learning
  4. Experiments
  5. Implementation Details
  6. Guidance under Weak Change
  7. Guidance under Strong Change
  8. Guidance across Trajectories
  9. Conclusion

Figures (4)

  • Figure 1: Overview of the offered framework: guidance is cast as the problem of learning the target localization $y_r$ from an image $x_r$ - at training time, the deep network is trained by sampling views from the stack of reference images - at runtime, the network directly predicts a localization from the current image.
  • Figure 2: Illustration of the image of this first experiment. Hypothetical target (red arrow) is not really visible in first image, yet, the surrounding is sufficient to know where it is. This explains how our model is able to learn a mapping image-to-target at any resolution.
  • Figure 3: Illustration of limits of registration on a single reference image in presence of strong change: the image displays the same region in two S2 images with green dot being the SIFT keypoints. Due to important appearance change only 20 SIFT will then be matched while the pair is already registered. Adding sub-sampling or geometric deformation frequently makes SIFT unable to perform registration while the offered baseline just learn to predict the target. Illustration done with github.com/Vincentqyw/image-matching-webui.
  • Figure 4: Outputs along a trajectory: all 8 images represent an image and an output mask (red dot is the location of the target, yellow is the pixel-wise predicted likelihood of being the target location). One can again notice the ground resolution difference between first en final image, yet the algorithm can coarsely predict the location of the target in all those situations.