Event-based Star Tracking via Multiresolution Progressive Hough Transforms

Samya Bagchi, Tat-Jun ChinWACV 2020

TL;DR

This paper introduces a novel event-based star tracking system that leverages multiresolution Hough Transforms to efficiently and accurately estimate spacecraft attitude from high-speed, asynchronous event camera data. This approach promises lower power consumption and higher speeds compared to traditional star trackers, making it suitable for resource-constrained platforms like FPGAs.

What Problem Does This Paper Address?

Estimating a spacecraft's attitude (its orientation in space) is crucial for mission success. Traditional star trackers, while accurate, rely on conventional cameras that capture images at fixed intervals, which can be power-hungry and limited in speed. Event cameras offer a promising alternative: they detect changes in pixel intensity asynchronously and at very high temporal resolutions (microseconds). This allows for potentially faster and more energy-efficient attitude estimation. However, the asynchronous, high-speed nature of event data presents a significant algorithmic challenge. Existing methods often convert event streams into traditional images, negating the benefits of event sensors, or rely on computationally expensive algorithms. This paper aims to bridge this gap by developing an efficient and accurate attitude estimation algorithm specifically designed for event camera data.

Methods

The core of the proposed method is a multiresolution progressive Hough Transform (HT) bank. Instead of processing the entire event stream at once, the system uses multiple HT instances operating at different time resolutions (e.g., 400ms, 200ms, 100ms). This allows the system to track rotational motions over varying time scales.

Here's a breakdown of the key techniques:

  • Event Data Processing: The algorithm works directly with the stream of individual events (pixel coordinates, timestamp, polarity). Polarity information is initially ignored for estimating relative rotation.
  • 3D Line Representation: Events are backprojected into 3D rays. The algorithm then seeks to find linear structures in the spatio-temporal domain of these rays, which correspond to stars.
  • Hough Transform for Line Detection: A multi-resolution Hough Transform is employed to detect these linear structures. The novelty lies in its progressive and event-triggered update mechanism. Each incoming event incrementally updates the state of the Hough cells, including the least-squares fitted line parameters and covariance matrix. This allows for constant-time processing per event, making it suitable for hardware implementation (e.g., FPGAs).
  • Relative Rotation Estimation: Once linear structures (potential star tracks) are identified, their "end points" at the start and end of a time interval are used to form 2D-3D correspondences. These correspondences are then used to estimate the relative rotation between two time points using a least-squares SVD-based approach.
  • Rotation Averaging: To obtain the absolute attitude, a set of relative rotations are fused with a small number of absolute attitude measurements (obtained via a separate star identification process on motion-compensated event images). A rotation averaging technique is used to jointly refine these measurements, denoising the results and producing a consistent sequence of absolute attitude estimates.

Datasets

The paper evaluates its method using event data from [8, 2]. This dataset consists of 11 event streams recorded by an iniVation Davis 240C event camera while observing star fields. The dataset also provides ground truth absolute attitudes for the entire stream, which is crucial for evaluation.

Key Insights and Ideas

  • Direct Event Processing is Key: The paper demonstrates that it's possible and beneficial to process event data directly without converting it into traditional images, thus retaining the advantages of event cameras.
  • Progressive Hough Transform for Efficiency: The event-triggered, incremental update of the Hough Transform is a significant innovation. It allows for constant-time processing per event, enabling real-time performance on resource-constrained hardware.
  • Multiresolution Approach for Robustness: Using multiple time resolutions for the Hough Transforms allows the system to track rotations over different durations, making it more robust to varying motion dynamics.
  • Rotation Averaging for Absolute Accuracy: Combining relative rotations with a few absolute measurements via rotation averaging effectively denoises the attitude estimates and achieves high accuracy.

Key Results

  • Superior Relative Rotation Accuracy: The proposed HT method consistently outperforms Contrast Maximization (CM) in estimating relative rotations across various event stream durations. It also generally shows better accuracy than Trimmed ICP (TICP), although TICP is more accurate in 4 out of 6 sequences.
  • Competitive Absolute Attitude Accuracy: The final absolute attitude estimates achieved by the HT method are highly competitive, satisfying the requirements of a commercial star tracker (achieving RMS angular error below 1 degree).
  • Efficiency: The HT method is significantly faster than CM, especially for longer event chunks. The runtime for rotation averaging is reported to be around 2 seconds per sequence.
  • Hardware Feasibility: The event-triggered processing of the HT is designed to be feasible for parallel computation on FPGAs.

Limitations and Future Work

  • Polarity Ignored: The current method does not utilize the polarity information of events for relative rotation estimation. The authors suggest that incorporating polarity could be beneficial for future work.
  • Star Identification Cost: The frequency of absolute attitude estimation is limited by the computational cost of star identification, which is performed on motion-compensated event images.
  • Synchronization: While the event processing is asynchronous, the rotation averaging step requires a synchronized set of relative and absolute attitude measurements.
  • Future Work: The authors suggest exploring the use of polarity information and further optimizing the implementation for even greater efficiency.

Summary updated

BibTeX
@inproceedings{bagchi2020event,
  author = {Samya Bagchi and Tat-Jun Chin},
  title = {{Event-based Star Tracking via Multiresolution Progressive Hough Transforms}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2020},
}
Figure