Access Official Technical PDF

Share Summary:LinkedInEmail

Handheld antenna assessment is challenging when performance depends on how a device is positioned and moved. For 5G systems using millimeter-wave (mmWave) frequencies, antenna-array response can also be affected by nearby objects and body absorption. Fixed calibration in a controlled chamber can provide a reference, but it does not reproduce the varied motion and orientations of everyday use. A camera-based approach offers a way to measure device motion during testing and relate that motion to radio-frequency (RF) observations.

A University of Calgary thesis examined whether computer vision (CV) pose estimates could be aligned with antenna phase measurements. The work used a single antenna as a demonstration, rather than calibrating a complete 5G array. Its broader aim was to assess whether camera-derived motion information could eventually contribute to more realistic antenna characterization.

Why handheld antenna assessment needs motion data

Beamforming uses the phase and amplitude of signals at antenna elements to shape or steer a transmitted or received beam. In an array, the relative phase of signals is important: changing an element’s position or the direction of an incoming wave changes the path length and therefore the phase relationship. Device motion and orientation can consequently affect the observed antenna response.

This creates a practical testing problem. A robotic setup can control movement, but it may not reproduce the small, irregular movements associated with a person holding a device. Measuring the device’s pose while it moves can help put RF observations in context. The thesis considered a camera as an alternative to relying only on inertial measurement, noting that inertial sensors can accumulate position error through drift.

Computer vision pose estimation for antenna testing

Pose describes both an object’s position and its orientation. In three dimensions, that means estimating translation along the x, y and z axes, as well as rotation about those axes: six degrees of freedom in total.

The approach in the thesis used visible feature points on a planar template. Distinctive corners are useful because image intensity changes in two approximately perpendicular directions, making them easier to locate than points on a plain surface or along a single edge. Camera calibration is needed to account for imaging characteristics such as lens distortion and the relationship between image and camera coordinates.

With feature locations identified across image frames, perspective geometry relates points on the template to their image projections. The Direct Linear Transform (DLT) expresses this relationship as a set of linear equations and is used to estimate the camera’s pose relative to the template. Because the camera is attached to or moves with the test device, this estimate can represent the device’s motion when the camera-to-antenna relationship is known.

Using filtering to manage noisy measurements

Feature-point locations are not perfectly stable from frame to frame. Image noise, blur and imperfect detection can make calculated pose estimates fluctuate. The thesis therefore explored ways to combine observations over time rather than treating every frame as an isolated measurement.

A Kalman filter (KF) estimates a system’s state recursively by combining measurements with a model of how that state changes. In this application, smooth hand motion can be represented as a process model, while the camera observations provide measurements. The filter balances the two according to their estimated uncertainty. If the relationship between pose and image measurements is nonlinear, an Extended Kalman Filter (EKF) can apply a local linear approximation.

The work considered two broad processing paths:

  • DLT followed by refinement: estimate pose from image geometry, then use a least-squares method or filtering to refine the result.
  • Direct EKF estimation: model the perspective relationship within the filter and estimate pose without first solving each frame independently with DLT.

These approaches involve a trade-off. DLT-based processing requires geometric calculations before refinement, while an EKF can incorporate the nonlinear measurement model but relies on an approximation when linearizing it. Filter performance also depends on the suitability of the motion model and the assumed process and measurement uncertainties.

Comparing camera pose with RF phase

To examine whether the estimated motion was relevant to antenna measurements, the experimental setup paired a moving antenna with a stationary antenna and a network analyzer. As the test antenna moved, the analyzer recorded changes in signal phase. The camera independently tracked the motion of the device relative to a template.

The comparison used the two data streams to calculate the changing separation between the antennas: one estimate came from camera-derived pose, and another from RF phase measurements. The thesis examined controlled translations and a combined translation-and-rotation case. This provided a way to assess whether the visual motion estimate corresponded with the phase response, rather than treating camera accuracy as an isolated image-processing question.

The reported work demonstrated alignment between CV pose estimates and antenna phase measurements using a single antenna. That is a feasibility result, not evidence that the method fully calibrates a multi-element array or that it meets a particular accuracy requirement in every operating condition. The thesis identified millimeter- and degree-level pose accuracy as a goal for future antenna characterization, not as a universal performance guarantee.

Practical implications for test and calibration workflows

For engineers and system integrators, the key idea is to pair RF measurements with a record of the device’s physical motion. This can help distinguish changes associated with antenna behavior from changes associated with the test device’s pose. It also points toward testing that better represents handheld use than a fixed-orientation measurement alone.

Applying the method requires attention to the complete measurement chain: camera calibration, reliable feature tracking, the geometry between camera and antenna, time alignment between video and RF data, and a motion model appropriate to the test. The thesis itself examined timing offsets and different motion cases, underscoring that synchronization and setup geometry matter when comparing independent measurements.

Conclusion

Computer vision can provide a useful motion reference for handheld antenna assessment by estimating device position and orientation from image sequences. Combined with filtering and RF phase measurements, this approach offers a practical foundation for investigating antenna behavior under movement. The demonstrated single-antenna results support further study, while full array calibration and dependable accuracy across real-world conditions remain questions for subsequent validation.

Access Official Technical PDF

About The Author