Typography-Based Monocular Distance Estimation for Advanced Driver-Assistance Systems

Technical Summary: Typography-Based Monocular Distance Estimation (T-MDE)

Problem Statement
Accurate distance estimation to a leading vehicle is a fundamental input for Advanced Driver-Assistance Systems (ADAS), including forward collision warning (FCW), adaptive cruise control (ACC), and automated emergency braking. Production systems typically rely on radar, LiDAR, or stereo cameras. While effective, these sensors introduce significant cost, power consumption, thermal load, and packaging constraints. Furthermore, radar often lacks lateral resolution, and stereo cameras require rigid mechanical baselines that are difficult to maintain calibration over a vehicle's service life. Single cameras are inexpensive and ubiquitous but suffer from a fundamental weakness: a single image lacks absolute scale. Without an external assumption of a real-world length, a single camera cannot distinguish between a small nearby object and a large distant one. Existing monocular solutions that assume vehicle width are limited by the high variability of vehicle dimensions across the fleet, leading to significant systematic errors.

Methodology
The paper proposes Typography-Based Monocular Distance Estimation (T-MDE), a method that recovers metric distance from a single camera by exploiting the regulated geometry of the rear license plate. The method operates on the principle that U.S. license plates have a fixed outer size (305 × 152 mm) and character heights strictly defined by state regulations (ranging from 63 to 72 mm).

The system processes video frames through a seven-stage pipeline:

1. Detection and Rectification: The plate is detected using a parallel ensemble of classical image processing techniques (adaptive thresholding, edge detection) to ensure robustness against glare and motion blur. The plate is then warped to a frontal view via homography.
2. Optical Character Recognition (OCR) and Segmentation: The plate serial is read using a primary engine (EasyOCR) with a fallback (Tesseract). Crucially, the system explicitly segments individual characters using vertical projection profiles to measure their pixel height ($\bar{h}$) accurately, rather than relying solely on the text string.
3. State Identification: The issuing state is identified via a three-stage cascade (text matching, color distribution analysis, and a lightweight classifier) to retrieve the specific legal character height ($H_s$) for that state. If identification fails, the system defaults to the national mean height to avoid large systematic biases.
4. Multi-Channel Distance Estimation: The method computes distance using three independent typographic measurements derived from the same plate:
   • Character Height: The primary metric, derived from $D = f_{px} H_s / \bar{h}_{corr}$.
   • Stroke Width: Fixed at 1/6 of the character height by regulation.
   • Character Spacing: Fixed at ~1/5 of the character height.
     These are combined via inverse-variance weighting. Additionally, the system measures the pixel spacing of the two mounting holes (fixed at 279.6 mm by AAMVA standards) as a fourth, recognition-independent channel.
5. Pose Correction: The system estimates the 6-DoF pose of the plate to correct for camera pitch, roll, and plate yaw. This prevents foreshortening from being misinterpreted as increased distance.
6. Depth Network Check: A single-image depth network provides a relative depth map, which is scaled to meters using a factor anchored to the geometric distance. This acts as a consistency check but is weighted down if it disagrees with the geometric measurement, ensuring the physical scale is never overridden by the learned model.
7. Filtering and Warning Logic: A constant-velocity Kalman filter smooths the distance, closing speed, and time-to-collision (TTC). The filter's measurement noise is modeled to grow with distance, matching the geometric error budget. The output is a warning state (caution or imminent threat) aligned with U.S. FCW regulations.

Key Contributions
The paper distinguishes this work from prior reports and related literature through five specific contributions:

1. Multi-Channel Robustness: Instead of a single measurement, the method fuses three typographic features (height, stroke, spacing) and two geometric checks (mounting holes, depth network). Channels are weighted by their measured variance, allowing the system to automatically down-weight corrupted data (e.g., motion blur affecting thin strokes) without losing the distance estimate.
2. Per-State Calibration: The system dynamically identifies the issuing state to select the correct character height, eliminating the systematic scale error (up to ~12%) that would result from using a single national constant.
3. Full Pose Recovery: The method recovers the 6-DoF pose of the plate, correcting for perspective foreshortening and providing bearing and yaw angles essential for lane-level association and reliability flagging.
4. Safety-Oriented Filtering: The output is not a raw frame-by-frame distance but a smoothed state (distance, speed, TTC) derived from a Kalman filter with distance-dependent noise, designed specifically to drive driver-assistance warnings without flickering.
5. Field Evaluation: The method is evaluated on real-world data (58 sessions, 2,265 measurements across 15 states) rather than synthetic or staged inputs, reporting internal consistency and detection rates alongside the design error model.

Results
Evaluated on a low-cost automotive camera across 15 states:

• Detection: The system achieved a mean plate detection rate of approximately 99%.
• Recognition: A confident serial was read in 63.4% of detections; however, distance estimation remained continuous even when the serial was unreadable, relying on the mounting holes or the depth check.
• Accuracy: The system reported a median per-frame uncertainty of 0.128 m (mean 0.173 m) at a mean distance of 1.24 m. The geometric and fused distances agreed to within 1.12%.
• Consistency: The measured error followed the theoretical model where relative distance error equals relative character height error (coefficient of variation ~2.3%). The Kalman filter residual was near zero (mean -0.033 m), indicating no systematic lag.
• Pose: The system successfully handled plates with yaw angles up to 87.8°, correcting for foreshortening to prevent distance underestimation.
• Runtime: On a consumer GPU, the system sustained approximately 1.1 frames per second, with the majority of compute time consumed by the recognition and depth networks, not the geometric calculations.

Significance and Claims
The paper claims that T-MDE offers a cost-effective, low-maintenance complement to existing active sensors (radar, LiDAR) and stereo cameras. Its primary significance lies in fault tolerance: because it derives scale from a printed standard rather than time-of-flight, parallax, or learned priors, it fails under different conditions (e.g., missing plates, extreme yaw) than radar or LiDAR. This diversity of failure modes enhances the safety of the perception stack.

The method provides a "metric anchor" from a passive sensor, returning a distance, bearing, and vehicle identity (via the plate serial) simultaneously. The authors emphasize that while the system achieves high internal consistency and a design error of less than 0.13 m, it does not claim externally verified accuracy for all conditions, noting that a controlled campaign with surveyed references is the necessary next step for certification. The method is presented not as a replacement for active sensors, but as a scalable, passive layer that improves the robustness of ADAS perception stacks.