-
Notifications
You must be signed in to change notification settings - Fork 287
Add section on obtaining vision measurements for pose estimation #3137
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Changes from all commits
65c976b
746b088
0ef2dc2
1c86dc1
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,303 @@ | ||
| # Using AprilTags for Pose Estimation | ||
|
|
||
| AprilTag detection is most commonly used to help your robot determine its position on the field. This article explains how to obtain vision measurements from AprilTags and use them with WPILib's pose estimators. | ||
|
|
||
| ## Overview | ||
|
|
||
| The process of using AprilTags for pose estimation involves: | ||
|
|
||
| 1. **Detect AprilTags**: Use a vision solution to detect AprilTags on the field | ||
| 2. **Get Tag Pose**: Look up the known field position of the detected tag from the ``AprilTagFieldLayout`` | ||
| 3. **Calculate Robot Pose**: Use the camera-to-tag transform and camera-to-robot transform to calculate where the robot is on the field | ||
| 4. **Apply Measurement**: Pass the calculated pose to your pose estimator using ``AddVisionMeasurement()`` | ||
|
|
||
| ## Using Vision Libraries | ||
|
|
||
| Most teams use existing vision processing libraries that handle the complex mathematics for you: | ||
|
|
||
| ### PhotonVision | ||
|
|
||
| PhotonVision provides the ``PhotonPoseEstimator`` class which simplifies the entire process: | ||
|
|
||
| .. tab-set-code:: | ||
|
|
||
| ```java | ||
| import org.photonvision.PhotonPoseEstimator; | ||
| import org.photonvision.PhotonPoseEstimator.PoseStrategy; | ||
|
|
||
| // Create the pose estimator | ||
| PhotonPoseEstimator photonPoseEstimator = new PhotonPoseEstimator( | ||
| fieldLayout, // AprilTagFieldLayout | ||
| PoseStrategy.MULTI_TAG_PNP_ON_COPROCESSOR, | ||
| camera, // PhotonCamera | ||
| robotToCam // Transform3d from robot to camera | ||
| ); | ||
|
|
||
| // In your periodic method | ||
| var result = photonPoseEstimator.update(); | ||
| if (result.isPresent()) { | ||
| var estimatedPose = result.get(); | ||
| poseEstimator.addVisionMeasurement( | ||
| estimatedPose.estimatedPose.toPose2d(), | ||
| estimatedPose.timestampSeconds | ||
| ); | ||
| } | ||
| ``` | ||
|
|
||
| ```c++ | ||
| #include <photon/PhotonPoseEstimator.h> | ||
|
|
||
| // Create the pose estimator | ||
| photon::PhotonPoseEstimator photonPoseEstimator{ | ||
| fieldLayout, // frc::AprilTagFieldLayout | ||
| photon::PoseStrategy::MULTI_TAG_PNP_ON_COPROCESSOR, | ||
| camera, // photon::PhotonCamera | ||
| robotToCam // frc::Transform3d from robot to camera | ||
| }; | ||
|
|
||
| // In your periodic method | ||
| auto result = photonPoseEstimator.Update(); | ||
| if (result) { | ||
| poseEstimator.AddVisionMeasurement( | ||
| result->estimatedPose.ToPose2d(), | ||
| result->timestamp | ||
| ); | ||
| } | ||
| ``` | ||
|
|
||
| ```python | ||
| from photonlibpy.photonPoseEstimator import PhotonPoseEstimator, PoseStrategy | ||
|
|
||
| # Create the pose estimator | ||
| photon_pose_estimator = PhotonPoseEstimator( | ||
| field_layout, # AprilTagFieldLayout | ||
| PoseStrategy.MULTI_TAG_PNP_ON_COPROCESSOR, | ||
| camera, # PhotonCamera | ||
| robot_to_cam # Transform3d from robot to camera | ||
| ) | ||
|
|
||
| # In your periodic method | ||
| result = photon_pose_estimator.update() | ||
| if result is not None: | ||
| pose_estimator.addVisionMeasurement( | ||
| result.estimatedPose.toPose2d(), | ||
| result.timestamp | ||
| ) | ||
| ``` | ||
|
|
||
| See the `PhotonVision documentation <https://docs.photonvision.org/en/latest/docs/programming/photonlib/robot-pose-estimator.html>`_ for complete details. | ||
|
|
||
| ### Limelight | ||
|
|
||
| Limelight cameras with MegaTag provide robot poses directly through NetworkTables: | ||
|
|
||
| .. tab-set-code:: | ||
|
|
||
| ```java | ||
| import edu.wpi.first.networktables.NetworkTable; | ||
| import edu.wpi.first.networktables.NetworkTableInstance; | ||
|
|
||
| NetworkTable limelightTable = NetworkTableInstance.getDefault().getTable("limelight"); | ||
|
|
||
| // In your periodic method | ||
| double[] botpose = limelightTable.getEntry("botpose_wpiblue").getDoubleArray(new double[6]); | ||
| if (botpose.length > 0 && botpose[0] != 0.0) { | ||
| Pose2d visionPose = new Pose2d(botpose[0], botpose[1], Rotation2d.fromDegrees(botpose[5])); | ||
| double latency = limelightTable.getEntry("tl").getDouble(0) + limelightTable.getEntry("cl").getDouble(0); | ||
| double timestamp = Timer.getFPGATimestamp() - (latency / 1000.0); | ||
|
|
||
| poseEstimator.addVisionMeasurement(visionPose, timestamp); | ||
| } | ||
| ``` | ||
|
|
||
| ```c++ | ||
| #include <networktables/NetworkTable.h> | ||
| #include <networktables/NetworkTableInstance.h> | ||
|
|
||
| auto limelightTable = nt::NetworkTableInstance::GetDefault().GetTable("limelight"); | ||
|
|
||
| // In your periodic method | ||
| auto botpose = limelightTable->GetEntry("botpose_wpiblue").GetDoubleArray({}); | ||
| if (!botpose.empty() && botpose[0] != 0.0) { | ||
| frc::Pose2d visionPose{units::meter_t{botpose[0]}, units::meter_t{botpose[1]}, | ||
| frc::Rotation2d{units::degree_t{botpose[5]}}}; | ||
| auto latency = limelightTable->GetEntry("tl").GetDouble(0) + limelightTable->GetEntry("cl").GetDouble(0); | ||
| auto timestamp = frc::Timer::GetFPGATimestamp() - units::millisecond_t{latency}; | ||
|
|
||
| poseEstimator.AddVisionMeasurement(visionPose, timestamp); | ||
| } | ||
| ``` | ||
|
|
||
| ```python | ||
| from ntcore import NetworkTableInstance | ||
| from wpilib import Timer | ||
| from wpimath.geometry import Pose2d, Rotation2d | ||
|
|
||
| limelight_table = NetworkTableInstance.getDefault().getTable("limelight") | ||
|
|
||
| # In your periodic method | ||
| botpose = limelight_table.getEntry("botpose_wpiblue").getDoubleArray([]) | ||
| if len(botpose) > 0 and botpose[0] != 0.0: | ||
| vision_pose = Pose2d(botpose[0], botpose[1], Rotation2d.fromDegrees(botpose[5])) | ||
| latency = limelight_table.getEntry("tl").getDouble(0) + limelight_table.getEntry("cl").getDouble(0) | ||
| timestamp = Timer.getFPGATimestamp() - (latency / 1000.0) | ||
|
|
||
| pose_estimator.addVisionMeasurement(vision_pose, timestamp) | ||
| ``` | ||
|
|
||
| .. note:: Use ``botpose_wpiblue`` or ``botpose_wpired`` based on your alliance color. These provide poses in the correct field coordinate system. | ||
|
|
||
| See the `Limelight documentation <https://docs.limelightvision.io/docs/docs-limelight/apis/complete-networktables-api#apriltag-and-3d-data>`_ for the complete NetworkTables API. | ||
|
|
||
| ## Important Considerations | ||
|
|
||
| ### Timestamps | ||
|
|
||
| It's critical to use the **timestamp from when the image was captured**, not when it was processed. Vision processing introduces latency (typically 20-100ms), and the pose estimator needs the actual capture time to properly fuse the measurement with odometry data. | ||
|
|
||
| Most vision libraries provide this timestamp: | ||
|
Collaborator
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. formatting issues |
||
| - PhotonVision: ``result.timestampSeconds`` | ||
| - Limelight: Calculate from ``tl`` (targeting latency) + ``cl`` (capture latency) | ||
|
|
||
| ### Standard Deviations | ||
|
|
||
| The accuracy of vision measurements varies based on several factors: | ||
|
|
||
| - **Distance from tags**: Measurements are less accurate when far from tags | ||
| - **Number of tags**: Seeing multiple tags improves accuracy | ||
| - **Tag ambiguity**: Low-resolution or angled views reduce accuracy | ||
| - **Camera quality**: Higher resolution cameras provide better accuracy | ||
|
|
||
| You should scale the standard deviations passed to ``AddVisionMeasurement()`` based on these factors: | ||
|
|
||
| .. tab-set-code:: | ||
|
|
||
| ```java | ||
| // Example: Scale standard deviations based on distance and tag count | ||
| double distance = /* calculate distance to nearest tag */; | ||
| int tagCount = /* number of tags seen */; | ||
|
|
||
| // More tags = more trust, greater distance = less trust | ||
| double xyStdDev = 0.5 * Math.pow(distance, 2) / tagCount; | ||
|
Collaborator
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This formula seems suspect. If you're 1 meter away with 1 tag, it would give a standard deviation of half a meter. But if you're 10 meters away, it would give a standard deviation of 50 meters. |
||
| double thetaStdDev = 999999.9; // Don't trust rotation from single tag | ||
|
Collaborator
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This comment would imply that there should be a conditional at use a large tehetaStdDev if there's only a single tag and a different one if there's more then one tag visible |
||
|
|
||
| poseEstimator.addVisionMeasurement( | ||
| visionPose, | ||
| timestamp, | ||
| VecBuilder.fill(xyStdDev, xyStdDev, thetaStdDev) | ||
| ); | ||
| ``` | ||
|
|
||
| ```c++ | ||
| // Example: Scale standard deviations based on distance and tag count | ||
| double distance = /* calculate distance to nearest tag */; | ||
| int tagCount = /* number of tags seen */; | ||
|
|
||
| // More tags = more trust, greater distance = less trust | ||
| double xyStdDev = 0.5 * std::pow(distance, 2) / tagCount; | ||
| double thetaStdDev = 999999.9; // Don't trust rotation from single tag | ||
|
|
||
| poseEstimator.AddVisionMeasurement( | ||
| visionPose, | ||
| timestamp, | ||
| {xyStdDev, xyStdDev, thetaStdDev} | ||
| ); | ||
| ``` | ||
|
|
||
| ```python | ||
| # Example: Scale standard deviations based on distance and tag count | ||
| distance = # calculate distance to nearest tag | ||
| tag_count = # number of tags seen | ||
|
|
||
| # More tags = more trust, greater distance = less trust | ||
| xy_std_dev = 0.5 * (distance ** 2) / tag_count | ||
| theta_std_dev = 999999.9 # Don't trust rotation from single tag | ||
|
|
||
| pose_estimator.addVisionMeasurement( | ||
| vision_pose, | ||
| timestamp, | ||
| (xy_std_dev, xy_std_dev, theta_std_dev) | ||
| ) | ||
| ``` | ||
|
|
||
| See :doc:`/docs/software/advanced-controls/state-space/state-space-pose-estimators` for more information about tuning standard deviations. | ||
|
|
||
| ### Rejecting Bad Measurements | ||
|
|
||
| You should reject vision measurements in certain situations: | ||
|
|
||
| - **No tags detected**: Only use measurements when tags are visible | ||
| - **High ambiguity**: Reject measurements with low confidence (check tag ambiguity values) | ||
| - **Unrealistic poses**: Reject measurements that are outside the field boundaries or far from your current estimate | ||
| - **During rapid motion**: Vision measurements may be less reliable during fast turns or acceleration | ||
|
|
||
| Example rejection logic: | ||
|
Collaborator
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This should specify what checks the code below does and what are left as exercise for the readers. This looks like it only handle the no tags detected and outside field boundaries cases. |
||
|
|
||
| .. tab-set-code:: | ||
|
|
||
| ```java | ||
| var result = photonPoseEstimator.update(); | ||
| if (result.isPresent()) { | ||
| var estimatedPose = result.get(); | ||
|
|
||
| // Check if pose is reasonable (within field boundaries) | ||
| if (estimatedPose.estimatedPose.getX() >= 0 && | ||
|
Collaborator
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. It feels like you wouldn't want to reject a tag that's on the edge of the field if it was reported as just outside the field, as that would still be a reasonable measurement. |
||
| estimatedPose.estimatedPose.getX() <= fieldLayout.getFieldLength() && | ||
| estimatedPose.estimatedPose.getY() >= 0 && | ||
| estimatedPose.estimatedPose.getY() <= fieldLayout.getFieldWidth()) { | ||
|
|
||
| poseEstimator.addVisionMeasurement( | ||
| estimatedPose.estimatedPose.toPose2d(), | ||
| estimatedPose.timestampSeconds | ||
| ); | ||
| } | ||
| } | ||
| ``` | ||
|
|
||
| ```c++ | ||
| auto result = photonPoseEstimator.Update(); | ||
| if (result) { | ||
| // Check if pose is reasonable (within field boundaries) | ||
| if (result->estimatedPose.X() >= 0_m && | ||
| result->estimatedPose.X() <= fieldLayout.GetFieldLength() && | ||
| result->estimatedPose.Y() >= 0_m && | ||
| result->estimatedPose.Y() <= fieldLayout.GetFieldWidth()) { | ||
|
|
||
| poseEstimator.AddVisionMeasurement( | ||
| result->estimatedPose.ToPose2d(), | ||
| result->timestamp | ||
| ); | ||
| } | ||
| } | ||
| ``` | ||
|
|
||
| ```python | ||
| result = photon_pose_estimator.update() | ||
| if result is not None: | ||
| # Check if pose is reasonable (within field boundaries) | ||
| if (0 <= result.estimatedPose.X() <= field_layout.getFieldLength() and | ||
| 0 <= result.estimatedPose.Y() <= field_layout.getFieldWidth()): | ||
|
|
||
| pose_estimator.addVisionMeasurement( | ||
| result.estimatedPose.toPose2d(), | ||
| result.timestamp | ||
| ) | ||
| ``` | ||
|
|
||
| ## Custom Vision Solutions | ||
|
|
||
| If you're implementing your own vision processing, you'll need to: | ||
|
|
||
| 1. **Detect tags and get camera-to-tag transforms**: Use a library like OpenCV's ``solvePnP`` to calculate the transformation from your camera to each detected tag | ||
| 2. **Transform to robot pose**: Apply your camera-to-robot transform (determined by camera mounting position) | ||
| 3. **Transform to field pose**: Use the tag's known field position from ``AprilTagFieldLayout`` to calculate the robot's field position | ||
| 4. **Handle latency**: Capture and use the image timestamp, accounting for processing delay | ||
|
|
||
| This approach requires solid understanding of 3D geometry and coordinate transformations. Most teams are better served using existing vision libraries that handle these details. | ||
|
|
||
| ## See Also | ||
|
|
||
| - :doc:`/docs/software/advanced-controls/state-space/state-space-pose-estimators` - Using pose estimators with vision measurements | ||
| - :doc:`/docs/software/basic-programming/coordinate-system` - Understanding the FRC coordinate system | ||
| - `PhotonVision Documentation <https://docs.photonvision.org/>`__ - Complete PhotonVision documentation | ||
| - `Limelight Documentation <https://docs.limelightvision.io/>`__ - Complete Limelight documentation | ||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -4,3 +4,4 @@ | |
| :maxdepth: 2 | ||
|
|
||
| apriltag-intro | ||
| apriltag-pose-estimation | ||
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This jumps right into the robot code side, but there should probably be some mention about needing to set-up photonvision and limelight for April tag tracking.