Determining the Translational Speed of a Camera from Time-Varying Optical Flow

Anton van den Hengel, Wojciech Chojnacki, Michael J. BrooksIWCM 2004 · DOI

TL;DR

This paper presents a method to determine the relative translational speed of a camera over time, even when only weakly calibrated. By tracking a single point in the scene and using time-varying optical flow, the camera's trajectory can be recovered, overcoming the inherent scale ambiguity of instantaneous optical flow.

What Problem Does This Paper Address?

Self-calibrating a moving camera using optical flow is a well-studied problem. However, a fundamental limitation exists: instantaneous optical flow alone cannot determine the magnitude of the camera's translational velocity. This is akin to the "scale ambiguity" problem in stereo vision. While we can determine the direction of motion and relative motion between points, the actual speed remains unknown without external information. This paper tackles this by showing how to recover the relative translational speed over time, effectively allowing us to reconstruct the camera's trajectory in a consistent scale.

Methods

The core of the method relies on analyzing time-varying optical flow from a single tracked point in a static scene. The authors derive a closed-form solution in the continuous realm, which is then adapted for discrete image sequences.

Key steps include:

  • Weak Calibration: The camera is assumed to be "weakly calibrated," meaning its intrinsic parameters (like focal length and principal point) and the direction of its translational and rotational velocities are known, but the magnitude of the translational velocity is not.
  • Time-Varying Optical Flow: Instead of just instantaneous flow, the paper leverages the flow of a point over a period of time. This allows for the analysis of how the optical flow changes.
  • Differential Epipolar Geometry: The derivation involves differentiating the projection equations and relating them to the camera's motion parameters.
  • Closed-Form Solution: A key contribution is the derivation of an equation that relates the ratio of translational speeds at different time points to an integral of known quantities derived from the optical flow and camera parameters. This integral effectively "smooths out" noise and provides a robust estimate of the relative speed.

Datasets

Not discussed in the paper. The paper focuses on a theoretical derivation and experimental validation using synthetic data.

Key Insights and Ideas

  • Time-Varying Flow is Key: The paper demonstrates that while instantaneous flow has scale ambiguity, time-varying flow, even from a single point, can resolve this ambiguity relatively.
  • Single Point Suffices: Surprisingly, tracking just one point in the scene over time is sufficient to determine the relative translational speed. This significantly reduces the complexity compared to methods requiring many points.
  • Relative Speed, Not Absolute: The method recovers the ratio of the current translational speed to the initial speed. This means if the initial speed is known or assumed, the entire trajectory's scale is determined.
  • Weak Calibration is Enough: The approach works with a "weakly calibrated" camera, which is more achievable through self-calibration techniques than full "strong" calibration.

Key Results

The paper presents a theoretical result (Theorem) showing that the ratio of translational speeds ||v(t)|| / ||v(a)|| can be uniquely determined. This is expressed as an exponential of an integral involving known quantities derived from the optical flow and camera parameters.

The experimental evaluation uses synthetic data:

  • Synthetic Testing: The method was tested using simulated video data with 100 optical flow elements and added Gaussian noise.
  • Accuracy: The estimated velocity magnitude using 25 frames per second (fps) had a maximum error of 0.93%.
  • Frame Rate Impact: A lower frame rate of 2.5 fps resulted in a decrease in accuracy, with a maximum error of 1.5%, highlighting the importance of sufficient temporal information.

Limitations and Future Work

  • Weak Calibration Assumption: The method relies on the camera being weakly calibrated. While weak calibration is more accessible than strong calibration, it still requires some prior information or a self-calibration step.
  • Noise Sensitivity: While the method shows good performance with noise, extreme noise levels or very few tracked points could still pose challenges.
  • Future Work: The authors suggest exploring:
  • Kalman-like smoothing for improved robustness.
  • Using multiple optical flow trajectories to further reduce error accumulation.
  • Techniques to reduce error accumulation over longer time periods.

Summary updated

BibTeX
@inproceedings{hengel2004determining,
  author = {Anton van den Hengel and Wojciech Chojnacki and Michael J. Brooks},
  title = {{Determining the Translational Speed of a Camera from Time-Varying Optical Flow}},
  booktitle = {IWCM},
  year = {2004},
  doi = {10.1007/978-3-540-69866-1\_14},
}
Figure