HelixTrack: Event-Based Tracking and RPM Estimation of Propeller-like Objects

HelixTrack is a fully event-driven method that jointly tracks propeller-like objects and estimates their RPM with microsecond latency by back-warping events to a rotor plane and refining pose through phase-coupled geometry, validated on a newly introduced TQE dataset where it outperforms existing baselines.

The Gist

Imagine you are trying to take a video of a hummingbird's wings with a regular camera. The wings move so fast that in a standard photo, they just look like a blurry mess. Now, imagine that the camera is also shaking because you are running while holding it. Trying to figure out exactly how fast the wings are spinning or where the bird is going becomes nearly impossible.

This is the exact problem HelixTrack solves, but for drones and their propellers.

Here is a simple breakdown of how it works, using some everyday analogies.

The Problem: The "Blurry Mess"

Standard cameras take pictures (frames) 30 or 60 times a second. If a drone propeller spins 10,000 times a minute, it moves way too fast for these cameras. The result is a blur.

• Event Cameras: Instead of taking pictures, event cameras act like a swarm of hyper-sensitive fireflies. They only "blink" when they see something move. This gives them super-speed (microsecond precision), but it creates a chaotic stream of dots rather than a clear picture.
• The Challenge: When you try to track a spinning propeller with these dots, the dots form a confusing spiral pattern. Most computer programs get confused because they assume objects move in smooth, straight lines. A spinning propeller breaks all those rules.

The Solution: HelixTrack

The researchers built a system called HelixTrack that treats the spinning propeller not as a confusing blur, but as a predictable pattern.

1. The "Unwinding" Trick (The Homography)

Imagine you are watching a record player spin. If you look at it from the side, the edge looks like it's moving up and down. If you look from above, it's a circle.
HelixTrack uses a mathematical "lens" (called a homography) to virtually flatten the spinning propeller.

• Analogy: Imagine the propeller is a spiral staircase. HelixTrack takes the chaotic dots from the camera and "unwinds" the staircase so it looks like a flat, straight hallway. Suddenly, the chaotic motion looks like a simple, straight line. This makes it easy for the computer to follow.

2. The "Conductor" (The Kalman Filter)

Once the dots are "unwound," the system needs to know exactly where the propeller is at any split second.

• Analogy: Think of the propeller as a dancer spinning on a stage. The event camera is the audience throwing confetti (dots) at the dancer.
• HelixTrack has a Conductor (a Kalman Filter) who watches the confetti. Even if the dancer moves slightly off-beat or the audience throws confetti in the wrong spot, the Conductor predicts the next move based on the rhythm. It updates the dancer's position and speed instantly, every time a single dot arrives.

3. The "Group Hug" (Batched Updates)

Sometimes, the "Conductor" needs a reality check to make sure it hasn't drifted off course.

• Analogy: Every few seconds, the system gathers a small group of dots (a "batch") and asks them, "Does this still look like a flat hallway?" If the answer is no, it adjusts the "lens" (the mathematical flattening) to make the hallway straight again. This keeps the tracking accurate even if the camera shakes or the drone moves erratically.

Why This Matters

The researchers didn't just build a cool tracker; they also built a new dataset (a library of test videos) called TQE because no one else had tested this specific problem before.

• Speed: HelixTrack is incredibly fast. It processes data 11.8 times faster than real-time. If a drone is spinning its propellers, HelixTrack can tell you the exact speed (RPM) and position almost instantly.
• Safety: This is crucial for "leader-follower" drones. If a drone needs to land on another drone or avoid a collision in a noisy, radio-silent environment, it can "listen" to the propeller's spin to know where the other drone is and what it's doing.

The Bottom Line

HelixTrack turns a chaotic, high-speed spinning mess into a neat, predictable line. It's like taking a tangled ball of yarn and instantly straightening it out so you can see exactly how it's moving. This allows robots and drones to see and react to fast-moving objects in a way that was previously impossible.

Technical Summary

Here is a detailed technical summary of the paper "HelixTrack: Event-Based Tracking and RPM Estimation of Propeller-like Objects."

1. Problem Statement

The paper addresses a critical gap in safety-critical perception for Unmanned Aerial Vehicles (UAVs) and rotating machinery: the inability of existing trackers to handle fast, periodic motion under egomotion (camera movement) and strong visual distractors.

• The Challenge: Conventional frame-based and event-based trackers rely on assumptions of "smooth translational motion." Propellers violate this by generating dense, strictly periodic event patterns. This causes standard trackers to drift, fail, or lose track entirely.
• The Specific Gap: While event-based rate estimation exists for static scenes, no prior work has successfully formulated a method to jointly track a specific propeller-like object and estimate its instantaneous Rotations Per Minute (RPM) from raw asynchronous events in a dynamic environment.
• Data Scarcity: There was no public dataset available for this specific task, necessitating the creation of a new benchmark.

2. Methodology: HelixTrack

HelixTrack is a fully event-driven, low-latency framework that treats the periodic motion of a propeller not as noise, but as a structured feature to be modeled.

Core Concept: The Helical Signature

The key insight is that when events from a rotating blade are "lifted" from the image plane into a rotor plane (via a homography), they form a helical signature in spatiotemporal coordinates. The method estimates two coupled quantities:

1. Geometry: An image-to-rotor-plane homography ($H$).
2. Latent Phase State: A state vector $x(t) = [\phi(t), \omega(t), \alpha(t)]^\top$ representing phase, angular velocity, and angular acceleration.

Algorithmic Pipeline

The system operates in two concurrent loops:

A. Per-Event Phase Tracking (Extended Kalman Filter - EKF)

• Back-warping: Incoming events are back-warped from the image plane to the rotor plane using the current homography estimate.
• Phase Residual: Events are mapped to a signed azimuth angle. A "wrapped phase residual" is calculated between the event's implied phase and the predicted phase state.
• State Update: An EKF updates the phase, speed, and acceleration estimates for every single event.
• Model Dynamics: The system models angular acceleration with "white jerk" (Gaussian noise), allowing the state uncertainty to grow predictably between irregular event arrivals.
• Gating: Events are filtered using a von Mises distribution (angular selectivity) and a spatial ring gate to focus on the blade tips.

B. Batched Pose Refinement (Gauss-Newton Optimization)

• Objective Function: Periodically, the system accumulates events into a batch to refine the homography parameters ($q$). The objective function minimizes a weighted sum of:
  • Phase Residuals: Aligning events with the helical phase model.
  • Radial Residuals: Ensuring events cluster around the unit radius (blade tip).
  • Polarity Residuals: Aligning event polarity (ON/OFF) with expected edge transitions.
  • Soft Band Constraints: Keeping events within an informative annulus ($r_{in}$ to $r_{out}$).
  • Balanced Tip Occupancy (BTO): A regularizer preventing the tracker from collapsing to "only inside" or "only outside" the blade tip radius.
• Update Trigger: A pose update is triggered when the accumulated helical phase shift exceeds $\pi$ (half a rotation), ensuring the geometry adapts to camera motion without lag.

3. Key Contributions

1. Problem Formulation: Defined the task of joint tracking and RPM estimation for periodic objects, bridging the gap between general event trackers and rate-only estimators.
2. HelixTrack Model: Proposed a novel architecture combining:
   • Asynchronous EKF for microsecond-latency phase/speed tracking.
   • Batched Gauss-Newton optimization for robust pose estimation.
   • A physics-based helical phase model that exploits the periodicity of the target.
3. TQE Dataset (Timestamped Quadcopter with Egomotion):
   • Created a new dataset with 13 high-resolution event sequences.
   • Contains 52 rotating objects (4 propellers per quadcopter).
   • Captured at 2m and 4m distances with 7 levels of increasing egomotion.
   • Ground Truth: Provides microsecond-accurate RPM measurements via synchronized IR sensors, solving the annotation difficulty of fast-moving hubs.
4. Performance Benchmark: Demonstrated that HelixTrack significantly outperforms adapted asynchronous blob trackers (AEB) and deep learning-based trackers (DeepEv) in this specific domain.

4. Experimental Results

• Accuracy: On the TQE dataset, HelixTrack consistently achieved the lowest Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) across all distances and motion levels.
  • Baselines (AEB, DeepEv) suffered from catastrophic failures (errors in the thousands of RPM) as motion intensity increased.
  • HelixTrack maintained errors in the range of ~100 RPM even under high motion.
• Latency & Speed:
  • Processes events at microsecond latency.
  • Runs at approximately 11.8× real-time speed (processing full-rate events).
  • Even with event subsampling (processing every $k$-th event), it maintains high throughput with acceptable accuracy trade-offs.
• Robustness:
  • Initialization: Tolerates coarse initialization errors (up to ~30% deviation in position/scale/RPM).
  • Ablation Studies: Confirmed that the coupling of phase and geometry (via the homography) and the specific loss terms (polarity, band constraints) are critical for stability. Removing these leads to order-of-magnitude error increases.

5. Significance and Impact

• Safety-Critical Applications: Enables reliable monitoring of UAVs in radio-denied or silent environments. The instantaneous RPM and phase can serve as communication-free signals for leader-follower swarms, collision avoidance, and intent detection (e.g., spool-up/spool-down).
• Paradigm Shift: Demonstrates that periodic motion, often considered "adversarial" for trackers, can be transformed into a modeling advantage when using event cameras and appropriate geometric priors.
• Generalizability: While focused on propellers, the framework applies to any radially symmetric, periodic actuator (e.g., fans, rotors), offering a principled path for tracking-and-measuring behaviors that are intractable for conventional frame-based systems.
• Resource Efficiency: Achieves high precision without requiring heavy deep learning backbones or high-frame-rate cameras, making it suitable for edge devices with strict latency constraints.

In summary, HelixTrack represents a significant advancement in event-based vision by solving the specific, difficult problem of tracking fast-rotating objects, providing a robust, low-latency solution backed by a new high-quality dataset and rigorous benchmarking.