💻 computer science

A Geometry-Informed Computer Vision Method for Detecting and Examining Overtaking Vehicles From A Bicycle

This paper presents a geometry-informed computer vision pipeline that automates the detection and analysis of overtaking vehicles from single bicycle-mounted cameras, achieving high accuracy in event identification and lateral distance estimation to overcome manual annotation bottlenecks in cycling safety research.

Original authors: Gandhimathi Padmanaban, Rayane Moustafa, Fred Feng

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gandhimathi Padmanaban, Rayane Moustafa, Fred Feng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are riding your bike down a busy city street. You know that feeling when a car zooms past you from behind? Sometimes they give you plenty of room, and sometimes they clip right by your handlebars. For years, researchers have wanted to study exactly how often these "close calls" happen, but they hit a major roadblock: the "Human Review" Bottleneck.

To study this, they used special bikes with cameras and sensors. But to find the actual moments when a car passed the bike, a human had to watch hours and hours of video, frame-by-frame, like a detective looking for a specific clue. It was slow, expensive, and limited how much data they could collect.

This paper introduces a smart, automated "digital detective" that does the heavy lifting for them. Here is how it works, broken down into simple concepts:

1. The Three-Stage Detective System

Instead of just looking at a picture and guessing, the system uses a three-step logic process, much like a security guard checking an ID card:

Step 1: The Spotter (RT-DETR): First, a powerful AI looks at the video and says, "Hey, that's a car!" It finds every vehicle in the frame.
Step 2: The Tracker (ByteTrack): Next, it follows that car as it moves. It keeps the car's identity even if it gets briefly hidden behind a tree or another car.
Step 3: The Geometry Judge (The Magic Part): This is the paper's big innovation. The system doesn't just rely on what the car looks like; it checks the geometry (the math of how things look in a camera).
- The "Approach" Test: As a car passes you from behind, it starts small on the left side of your rear-view camera and gets bigger as it moves to the right. The system checks: "Is the car moving from left to right? Is it getting bigger? Is it staying on a straight path?"
- The "Pass" Test: If the car does all that and crosses a specific line on the screen, the system confirms: "Yes, this is a real overtaking event."

Why is this cool? It doesn't need expensive radar, lasers, or complex camera calibration. It just needs a standard camera and the laws of perspective. It's like knowing a friend is walking toward you not because you recognize their face, but because they are getting larger in your vision and moving from your left to your right.

2. How Well Did It Work?

The researchers tested this "digital detective" on 315 real-world passing events in Ann Arbor, Michigan.

The Score: It found 97.8% of the actual passing events.
The Mistakes: It made zero false alarms. It didn't mistake a parked car or an oncoming car for a passing one.
The Misses: It missed 7 events. Why? Mostly because another car blocked the view (like a friend standing in front of you at a concert) or the AI got confused for a split second.

3. The "Early Warning" Superpower

One of the most exciting findings is timing.

The system spotted the car's intent to pass an average of 2.44 seconds before the car actually reached the cyclist.
Think of it like a head start in a race. Since it takes a human about 1.5 seconds to react to danger, this system gives the cyclist a "safety buffer" of nearly a full second to move over or brace themselves.
In 84% of the cases, the system gave enough warning time to be useful for an active safety alarm.

4. How Close Were the Cars? (The "5-Foot" Rule)

Many places have a law saying cars must pass cyclists with at least 5 feet (1.5 meters) of space.

The researchers used a special sensor on the bike to measure the actual distance for 96 of these events.
The Result: 33% of the cars passed closer than the 5-foot rule. That's one out of every three cars.
The closest pass was only about 2.9 feet away—dangerously close.
This proves that even with bike lanes painted on the road, drivers often don't give enough space.

5. Guessing Distance Without a Ruler

The paper also tried something new: estimating the distance without a ruler.

Using only the shape of the car's "box" in the video (how tall it looks and where the bottom of the box is), the system guessed the passing distance.
It was off by only about 13 to 14 centimeters (roughly 5 inches).
This is good enough to tell the difference between a "safe pass" and a "dangerous pass" without needing expensive laser sensors.

The Bottom Line

This paper solves a massive problem in cycling safety research. Before, researchers were stuck watching videos manually, which limited how much they could learn. Now, they have a tool that automatically finds these events from a simple camera, allowing them to study thousands of passes instead of just a few hundred.

It confirms that while we have bike lanes, drivers are still cutting it close. But more importantly, it shows that we can build a simple, low-cost system that could one day warn cyclists before a car gets too close, giving them that crucial extra second to stay safe.

Technical Summary: A Geometry-Informed Computer Vision Method for Detecting and Examining Overtaking Vehicles From A Bicycle

Problem Statement
Instrumented bicycle studies have historically provided critical field evidence regarding vehicle passing behavior and cyclist safety. However, a significant methodological bottleneck exists: extracting overtaking events from continuous rear-facing video requires manual, frame-by-frame annotation. This process constrains sample sizes, limits the scope of naturalistic cycling safety research, and introduces inter-rater variability. Existing computer vision approaches often rely on appearance-based classifiers that struggle with domain shifts (e.g., lighting, camera motion) or require multi-sensor fusion (LiDAR, radar) with complex calibration and synchronization, making them unsuitable for scalable, low-resource research deployments. Furthermore, distinguishing actual overtaking maneuvers from the background of non-overtaking vehicle trajectories (oncoming traffic, parallel lanes, decelerating vehicles) in rear-facing footage remains a challenge for standard detection and tracking algorithms.

Methodology
The authors propose a geometry-informed computer vision pipeline that automates overtaking event detection using a single bicycle-mounted camera without explicit camera calibration or multi-sensor configurations. The system processes 2D video sequences through three sequential stages:

Vehicle Detection: The system employs RT-DETR (Real-Time Detection Transformer), a transformer-based detector chosen for its robustness to scale variations and partial occlusions common in moving bicycle footage. The model identifies four vehicle classes (car, motorcycle, bus, truck) from the COCO dataset. A pre-tracking suppression step is implemented to remove duplicate bounding boxes for the same vehicle to prevent tracking errors.
Multi-Object Tracking: ByteTrack is utilized to link detections into continuous trajectories. The authors adjusted ByteTrack's parameters to suit the overtaking scenario, specifically raising the track activation threshold to reduce false starts from background objects and limiting the lost track buffer to prevent ID carryover between different vehicles. A same-frame duplicate filter is also applied to handle ID fragmentation during brief occlusions.
Geometric Validation: This is the core innovation. A three-stage validation module filters tracked trajectories to identify only completed overtaking maneuvers based on perspective projection principles:
- Bearing Angle Trend: The system calculates the bearing angle of the vehicle relative to a fixed reference point (bottom-left corner for the rear-left camera). Overtaking vehicles must exhibit a monotonically increasing angle (moving left-to-right in the frame). A linear regression on the angle history must show a positive slope; decreasing slopes (indicative of oncoming traffic) result in rejection.
- Apparent Size Growth: The system monitors the rate of change in the bounding box area. Valid overtaking tracks must show positive growth as the vehicle approaches, filtering out static objects or vehicles moving away.
- Spatial Confirmation: A track is only confirmed as an overtaking event when the right edge of the vehicle's bounding box crosses a configurable spatial threshold (set at 55% of frame width for the rear-left setup), ensuring the vehicle has moved sufficiently across the frame to constitute a pass.

Key Contributions

Automated Detection Pipeline: The authors present a three-stage system combining RT-DETR, ByteTrack, and geometric validation. Validated on 315 manually annotated real-world overtaking events from urban roads in Ann Arbor, Michigan, the system achieved 97.8% recall with zero false positives.
Quantitative Characterization of Passing Behavior: By automating event isolation, the authors analyzed lateral passing distances for 96 events using ultrasonic sensor ground truth. They found that 33.3% of passes occurred below the 5-foot (152.4 cm) safety threshold, a rate consistent with prior field and self-reported studies.
Advance Warning Capability: The system identified overtaking intentions a mean of 2.44 seconds before vehicle passage. 84.1% of events exceeded the 1.5-second human reaction time threshold, demonstrating feasibility for active cyclist warning applications.
Calibration-Free Distance Estimation: The paper demonstrates that bounding box geometric features (specifically the vertical position of the lower edge and apparent box area) can estimate lateral passing distance without explicit camera calibration. Under leave-one-out cross-validation on 96 events, the approach achieved a mean absolute error (MAE) of approximately 13 to 14 cm, sufficient to distinguish close passes from standard passes.

Results

Performance: The system successfully detected 308 of 315 ground truth events. The seven missed detections were attributed to occlusion by nearer-lane vehicles (57%) or intermittent RT-DETR detections preventing sufficient tracking history (43%).
Timing: The advance detection time ranged from 0.50 to 6.30 seconds, with a median of 2.40 seconds.
Distance Analysis: Measured lateral clearances ranged from 88.0 cm to 247.0 cm, with a mean of 169.0 cm. The distribution confirmed that typical passing behavior in these urban environments operates near recommended safety margins, with a significant portion falling below legal minimums.
Distance Estimation: The vertical position of the bounding box bottom edge was the dominant predictor of lateral distance (Spearman correlation $\rho = -0.774$ ), followed by bounding box area ( $\rho = -0.567$ ).

Significance and Claims
The paper claims that this work addresses the primary annotation bottleneck in instrumented bicycle research. By automating event isolation from consumer-grade camera footage, the system enables data collection campaigns of substantially greater scope without proportional increases in manual effort. The authors emphasize that grounding event validation in geometric invariants (angle trends, size growth) provides a robust detection method that does not depend on sensor-specific calibration or appearance-based learning, which are often vulnerable to domain shifts.

The findings regarding passing distances provide quantitative, site-specific evidence that bicycle lane markings alone do not reliably produce compliant driver behavior, supporting the need for enforcement or infrastructure modifications. Furthermore, the preliminary distance estimation capability offers a scalable foundation for safety categorization without the need for expensive, calibrated ranging hardware. The authors conclude that while the system is currently limited to specific camera configurations and daylight conditions, it provides a scalable, modular foundation for vehicle-bicycle interaction analysis across larger datasets and diverse urban environments.