Demo 2 Visualize key observations

Objective

• Visualize target and current speed/acceleration
• Visualize target and current steering angle/effort
• Visualize Position of car, signs, and dynamic obstacles on Lanelet map
• Visualize Localization result confidence
• Visualize heartbeat data

Challenge outline

• Current and target values for speed, acceleration, and steering angle are easily obtainable from existing topics and can be easily visualized in any chosen framework.
• Spacial position of elements are easily obtainable from existing topics, but would be difficult to visualize without existing support for point clouds or lanelet maps. in underlying framework.
• Localization confidence likely does not exist in current topics and likely is not calculated by existing Autoware localization nodes.
• Heartbeat visualization is dependent on how the node status heartbeats are implemented but will be easy to visualize. Only Localization Confidence has significant difficulty apparent at this time, and the next section is dedicated entirely to exploring that difficulty.

Localization Result Confidence

Through inspection of topics and nodes used in the localization stack, there does not appear to be an already-implemented confidence estimate topic, and a quick search of the Autoware docs does not reveal any indication that it is calculated inside the nodes either. A possible exception might be odometry localization. However, there are still a few options forward.

Sensor Confidence: Although the localization results may not calculate confidence estimates, it may be possible to obtain the sensor confidence level:

• The LIDAR localization confidence is is said to be around 10cm and the margin of error for the LIDAR results themselves are around +-3cm. The LIDAR does not publish dynamic confidence estimates so we will have to rely on these static estimates.
• The GPS confidence may be reported by the sensor itself. I have not found the model or documentation for the GPS module we use, so input is requested by someone familiar with our GPS implementation. Even if it does not, there are a number of ways to calculate it ourselves that may be adapted to our moving-target scenario.
• Sensor confidence of the odometry information can be obtained with more information about the car itself.

Comparing Localizations: Between GPS, NDT, and odometry, we have multiple different localization results. Even if we can't obtain a confidence estimate from any single one of these, by comparing their results we may be able to tell if one of these has drifted. The most reliable of these is NDT, which as mentioned has an accuracy of around 10cm. Odometry is likely the second most reliable, but is vulnerable to long-term drift due to accumulation of error. GPS is not reliable moment to moment, but is not subject to the same long-term issues as Odometry. We already compare all three localizations to get our final position estimate, so it isn't too large of a cognitive leap to compare them for our confidence as well.

Adding confidence estimates to the localization nodes: This requires in-depth knowledge of the implementation and theory of each of the localization algorithms and is likely much more than an undergraduate student can do in a semester at 5 hours/week.

Visualization Tools

• https://github.com/dheera/rosboard: Built for the purpose we need it for. Designed to be extendable if we need something it doesn't have.

Notes from Meeting 1

• The final cost from the NDT error function can be a stand-in for confidence
• Odometry localization has yet to be implemented
• The GPS module might even publish confidence