Skip to Content
DEEPCRAFT™ Studio 5.14 has arrived. Read more →

Radar models

This page adds radar checks to the test stages. Refer to the Gesture Classification Ready Model Report for the range, confusables, and test method.

Radar setup

  • Set the radar configuration as described in the report.
  • Pass the raw radar data straight to the model. The Ready Model includes its own feature extraction.

Setup checks

  • The radar responds and reports the expected frame rate.
  • The mount is stable and does not vibrate during a test.
  • With no user in the scene, the model stays idle.
  • Record the enclosure material and thickness in front of the sensor. The enclosure wall sits in the radar beam and changes the signal.

Placement

  • The report describes how the model was set up during development. Use it as a guide for placement.
  • Test every mount height and orientation your product allows in its intended use. The report states the height and orientation it tested at. Gesture directions are relative to the sensor, so a rotated or face-up mount changes which gesture is which.
  • Test across the full distance and angle range stated in the report, including the limits.
  • Record the mounting surface. A surface that moves or vibrates changes the results.

Environment

  • People or objects moving in the field of view can trigger detections. Test with and without other people moving nearby.
  • Other radars in the same frequency band can interfere.
  • Sleeves, objects held in the hand, and jewelry near the gesturing hand can change results.

User familiarity

Give participants only the instructions your product gives its users, such as a quick-start guide or an on-screen hint. Do not coach participants during the test. Include people who are new to the product as well as people from your team.

Users learn with use. Record results from the first session separately from later sessions. This shows how quickly real users reach good performance.

Counting gestures

Apply the counting rule from Define how you count detections, with these additions for gestures:

  • Report recall (TP / (TP + FN)) for each gesture separately.
  • Record which gestures the model confuses, for example Push detected as Swipe Down.
  • Count triggers from non-gesture movements as a separate category.
  • Record follow-on movements. For example, lowering the hand after Swipe Up can trigger Swipe Down. Decide in your counting rule how your application handles these. The post-processing included in the generated code filters out some follow-on triggers. Refer to the Gesture settings on the Ready Models page.

Radar test matrix

Copy this table and adapt it to your test. The middle part of each ID is the stage. Priority is Core (recommended for every test plan) or Optional (add where it fits your product). Rows with a pass/fail result follow Sample size. Rows that record how the model behaves do not need a fixed number of test events.

ID                      PriorityPurposeMovementDistanceAngleExpected result
R-1-01CoreRadar check, empty sceneNoneN/AN/AFrame rate matches the report, model stays idle
R-2a-01CoreBaseline per gestureEach supported gestureNear0°Recall meets your criteria
R-2b-01CoreDistance sweepEach gestureNear, mid, and far0°Record recall at each distance
R-2b-02CoreAngle sweepEach gestureMid0° and off-axisRecord recall at each angle
R-2b-03CoreMount height and orientationEach gestureMid0°Record recall at each height and orientation your product allows
R-2b-04OptionalEnclosure effectEach gestureMid0°Compare with and without the enclosure
R-2c-01CoreNon-gesture movementHand waves, walking, sittingMid0°False positives within your criteria
R-2c-02CoreConfusable movementFast body movements that resemble gesturesMid0°Record any gestures detected
R-3a-01CoreDeployment environmentEach gestureDeployment distanceDeployment angleMeets your acceptance criteria
R-3a-02OptionalOther people nearbyGestures with someone moving nearbyMid0°Record the effect on recall and false positives
R-3a-03CoreFirst-time usersEach gesture, product instructions onlyDeployment distanceDeployment angleRecord first-session and later-session results separately
R-3b-01CoreDevice handlingDevice or mount touched or bumpedN/AN/ARecord how the model behaves
Last updated on