VANGUARD: Vehicle-Anchored Ground Sample Distance Estimation for UAVs in GPS-Denied Environments

This paper introduces VANGUARD, a lightweight geometric perception tool that enables LLM-based UAV agents operating in GPS-denied environments to accurately estimate Ground Sample Distance and recover metric scale by leveraging detected vehicles as environmental anchors, thereby significantly reducing spatial hallucinations and catastrophic failures compared to state-of-the-art vision-language models.

The Gist

Imagine you are flying a drone over a city, but you've lost your GPS signal and your camera's "ID card" (metadata) is missing. You can see the world below, but you have no idea how big anything actually is. Is that parking lot 10 feet wide or 100 feet? Is that car 15 feet long or 150 feet?

Without this "size sense," your drone is like a person trying to catch a ball in the dark—they might guess, but they'll likely drop it.

This paper introduces VANGUARD, a smart tool designed to give drones their "size sense" back, even in the dark.

The Problem: The "Guessing Game" of AI

Recently, engineers started using super-smart AI models (called VLMs or LLMs) to help drones plan their missions. These AIs are great at understanding language and recognizing objects. "I see a swimming pool," they might say.

But when asked, "How big is it?" these AIs start hallucinating.

The researchers tested five of the smartest AI models available. When asked to measure the area of a field just by looking at a picture, they were wrong 50% of the time. Sometimes they thought a small pond was the size of a lake, or a tiny car was the size of a bus. In the real world, if a drone thinks a landing spot is huge when it's actually tiny, it could crash. This is called Spatial Scale Hallucination.

The Solution: VANGUARD (The "Car Ruler")

Instead of asking the AI to "guess" the size, the researchers built a simple, math-based tool called VANGUARD.

Here is the clever trick: Cars.

No matter where you are in the world (mostly in cities), cars are roughly the same size. A standard sedan is about 16 to 17 feet (5 meters) long.

VANGUARD works like this:

1. Spot the Cars: The drone takes a picture and uses a detector to find all the cars. It draws a box around them.
2. Count the Pixels: It measures how many pixels long those cars are in the image.
3. Do the Math: If the AI sees a car that is 100 pixels long, and it knows a real car is 5 meters long, it can instantly calculate: "1 pixel = 0.05 meters."
4. The "Ruler" is Born: Now the drone has a ruler. It can measure anything else in the picture (a field, a building, a pool) by counting pixels and multiplying by that ruler.

Why is this better than the "Smart" AI?

The researchers compared their method to the "guessing" AI models.

• The Guessing AI: Tries to use its "common sense" to estimate size. It fails because it doesn't truly understand physics; it just predicts what words usually go together.
• VANGUARD: Uses deterministic math. It doesn't guess; it calculates. It's like using a tape measure instead of guessing how long a rope is by looking at it.

The Results:

• VANGUARD was wrong by only 6.87% on average.
• The Guessing AI was wrong by 38% to 52% on average.
• When measuring a swimming pool, the Guessing AI thought it was 50% smaller than it really was. VANGUARD got it within 1.5% of the real size.

The "Safety Switch"

VANGUARD is also smart about when not to trust itself.

• If the image is too blurry (cars look like tiny dots), VANGUARD says, "I can't measure this accurately."
• It gives the drone a confidence score. If the score is low, the drone knows, "Okay, I shouldn't try to land here based on this measurement. I'll wait for better data or use a different strategy."

The Big Picture

This paper teaches us a valuable lesson for the future of robotics: Don't rely on a general "smart" brain to do everything.

Instead, give the robot a specialized tool for specific jobs.

• Let the "Smart Brain" (LLM) decide where to go and what to do.
• Let the "Math Tool" (VANGUARD) handle the exact measurements.

By combining the two, we get a drone that is both smart enough to plan a mission and precise enough to execute it safely without crashing into things it thinks are bigger or smaller than they really are.

In short: VANGUARD turns a drone with a broken ruler into one with a perfect, invisible tape measure, using nothing but the cars parked on the street.

Technical Summary

Here is a detailed technical summary of the paper "VANGUARD: Vehicle-Anchored Ground Sample Distance Estimation for UAVs in GPS-Denied Environments."

1. Problem Statement

Autonomous aerial robots (UAVs/MAVs) operating in GPS-denied or communication-degraded environments often lose access to camera metadata and telemetry. Without this data, monocular imagery lacks Ground Sample Distance (GSD)—the physical size of a single pixel in meters. Consequently, the system cannot convert pixel-level measurements into real-world metric dimensions.

The paper identifies a critical failure mode in modern Vision-Language Models (VLMs) and Large Language Models (LLMs) used as high-level planners: Spatial Scale Hallucination. Even when provided with hints about object sizes, state-of-the-art VLMs exhibit median area estimation errors exceeding 50% (ranging from 38% to 52%), leading to potentially catastrophic decisions (e.g., attempting to land on a surface too small for the drone).

2. Methodology: VANGUARD

The authors propose VANGUARD (Vehicle-ANchored Geometric Understanding And Resolution Determination), a lightweight, deterministic Geometric Perception Skill designed as a callable tool for LLM-based agents. Instead of relying on the LLM to "guess" the scale, the tool computes it mathematically using ubiquitous environmental anchors.

The pipeline consists of five stages:

1. Vehicle Detection:

   • Uses an Oriented Bounding Box (OBB) detector (YOLO11l-OBB) to identify small vehicles in the aerial image.
   • OBBs are used instead of axis-aligned boxes to accurately capture vehicle length regardless of orientation.
   • Detections with low confidence are discarded.

2. Outlier Filtering:

   • Applies a median-based filter to remove false positives (e.g., road markings) and atypical large vehicles (buses/trucks).
   • A detection is kept if its pixel length $P_i \leq \alpha \cdot \text{median}(\{P_j\})$, where $\alpha = 1.5$.

3. Kernel Density Estimation (KDE) Mode Finding:

   • Instead of using a simple mean or median, the system applies KDE to the distribution of filtered pixel lengths.
   • This identifies the modal pixel length ($P_{mode}$), which is robust against skewed distributions caused by outliers.
   • Fallback: If fewer than 5 vehicles are detected, the system defaults to the simple median.

4. GSD Computation:

   • The system uses a pre-calibrated reference vehicle length ($L_{ref}$). Based on the DOTA v1.5 dataset, the modal physical length of small vehicles is determined to be 5.045 meters.
   • GSD is calculated as: $GSD_{pred} = L_{ref} / P_{mode}$.

5. Confidence Evaluation & Safety Fallback:

   • A composite confidence score ($C \in [0, 1]$) is generated based on sample sufficiency, distribution concentration, detection quality, and anomaly detection.
   • Resolution Guard: A hard threshold is applied. If the estimated pixel length is too small (indicating low resolution, $GSD > 0.3$ m/px), the confidence is capped at 0.3.
   • If $C < 0.5$, the agent is instructed to fall back to alternative strategies (e.g., visual odometry) rather than trusting the metric estimate.

3. Key Contributions

• Metadata-Free GSD Estimation: A novel method to recover absolute metric scale using small vehicles as geometric anchors, achieving 6.87% median error on the DOTA v1.5 benchmark.
• Empirical Evidence of VLM Limitations: A 100-entry benchmark demonstrating that even with explicit hints (e.g., "cars are ~5m"), VLMs suffer from Spatial Scale Hallucination, exhibiting 2.6× higher category dependence and 4× more catastrophic failures compared to the deterministic VANGUARD pipeline.
• Tool-Augmented Agent Paradigm: The integration of VANGUARD as a stateless API for LLM planners. This allows agents to delegate metric-critical perception to a deterministic geometric tool, ensuring safety without requiring the LLM to learn complex 3D geometry.
• Robust Statistical Approach: The use of KDE for mode estimation provides a 17% improvement in accuracy over simple mean aggregation, effectively handling noise and outliers in vehicle detection.

4. Experimental Results

The system was evaluated on the DOTA v1.5 dataset and a custom RS-GSD Benchmark v5.0.

• GSD Estimation Accuracy:

  • Achieved 6.87% median error on 306 images (67% coverage).
  • 83.3% of images had errors under 20%.
  • The method performs best in sub-meter resolution urban imagery ($GSD < 0.3$ m/px).

• Area Estimation (Downstream Task):

  • Integrated with SAM (Segment Anything Model) for segmentation, the pipeline achieved 19.7% median error on a 100-entry benchmark.
  • Comparison with VLMs:
    • VANGUARD: 19.7% median error, 97% of predictions within 100% error.
    • Zero-shot VLMs (e.g., GPT-4o, Qwen): 38–52% median error, with frequent order-of-magnitude deviations.
    • VLMs with Hints: Even with hints, VLMs only improved to ~35% error (GPT-4o) or remained high, whereas VANGUARD remained consistent.

• Failure Analysis:

  • VLMs showed high "category dependence" (performance varied wildly by object type), whereas VANGUARD showed 2.6× lower dependence.
  • VANGUARD reduced catastrophic failures (errors >100%) by a factor of 4 compared to the best VLM baseline.

5. Significance and Conclusion

The paper argues that deterministic geometric tools are essential for safe autonomous spatial reasoning in embodied agents. While LLMs/VLMs excel at high-level reasoning and semantic understanding, they currently lack the precision required for metric-scale physical tasks.

VANGUARD bridges this gap by acting as a "perception skill" that grounds the agent's decisions in physical reality. By offloading GSD estimation to a robust, mathematically grounded pipeline, autonomous UAVs can operate safely in GPS-denied environments without relying on hallucinated spatial estimates from generative models. Future work aims to extend this to multiple reference object classes (e.g., shipping containers, road lanes) and diverse geographic regions.