Section 3 Object Fusion

In this task, we will complete the Kalman Filter with the Object Fusion step.

Task 3: Implement Object Fusion

Your task is to fill gaps in the C++ code of the object fusion module: workshops/section_3/object_fusion/src/modules/fuser/StateFuser.cpp

Fusion code

As usual, we already implemented the less interesting parts of the fusion. The core functionality is still missing though. Implement the Kalman filter measurement update formulas for each dynamic object as described in the slides.

As in the object predicition you need to overwrite the global object list data_->object_list_fused.objects instead of using return statements. Note: when you use the utility functions like IkaUtilities::getEigenStateVec(myIkaObject), they will return an Eigen vector or matrix that points to the original memory, so by modifying this Eigen data structure, you will automatically overwrite the corresponding myIkaObject.

After completing the task, rebuild the workspace with catkin build. You should finally see a fully functional object fusion similar to the video below.

Bonus Task: Explore fusion parameters

Parameter tuning

We have finished the fusion, however sometimes there still occur some glitches. This indicates that the fusion parameters are not ideally chosen.

Play with the process noise matrix $\mathbf{Q}$ entries by manipulating the diagonal of time_variant_process_noise_matrix in the config file object_fusion_wrapper/param/kalman_filter.yaml. Try out the following values for each diagonal entry:

• 0.001
• 0.1
• 100

What effect does it have on the fused object? Can you explain why?

Also, play with the association parameters in object_fusion_wrapper/param/fusion.yaml. For the Mahalanobis distance, set the fusion_passat:constants:mahalanobis_threshold to

• 2.0 (standard deviations)
• 3.0 (standard deviations)
• 4.0 (standard deviations)
• 5.0 (standard deviations)

For the Intersection over Union set the fusion_passat:constants:iou_overlap_threshold to

• 0.1
• 0.2
• 0.5

What effect does it have on the fused object list? Can you explain why? Which association criterion do you see as more robust?

There are many more parameters in the mentioned files. Can you manage to perfectly tune the parameters and achieve a perfect object fusion?

Wrap-up

• You learned the fundamentals and challenges of Object Fusion and Tracking
• You learned what a Multi-Instance Kalman filter is
• You got a deeper understanding of the steps Object Prediction, Object Association and Object Fusion
• You explored the effects of fusion parameters
• You deepened your knowledge of ROS
• You learned how to inspect properties in RViz

References

• ROS
• Flatland
• Eigen