Relative Bearing Accuracy
This page shows results published as part of our work Toolbox Release: A WiFi-Based Relative Bearing Sensor for Robotics. The accuracy of estimated relative bearing (i.e AOA) returned by our toolbox is impacted by the error in displacement estimation from an on-board inertial sensor. Following results show the obtained AOA accuracy during our hardware experiments when using both groundtruth displacement and estimated displacement (using Intel T265 Tracking camera). For robot displacement in 2D only azimuth AOA is considered; the AOA profile is flattened along the elevation direction.
Note: Internal code logic for AOA error and some bug fixes were made to improve AOA accuracy and the updated results will be updated in the paper after the first review.
The results are for default configuration. For performance using other configurations, please refer here
As seen from the above results, the standard deviation for AOA accuracy is high for NLOS. This is also partly due to presence of more multipath peaks in the AOA profile making estimation of AOA difficult. Hence we use the profile variance to determine whether the obtained AOA is a potential outlier or not.
The profile on the left is typical example of a high variance profile while the one on the right resembles a low variance profile. The cumulative results for AOA accuracy vs profile variance are shown below.
Related stats as are follows:
- When using groundtruth displacement
LOS
Rejected samples = 0.6%
Mean error for rejected LOS samples 5.711851119995117 degrees
Median error for rejected LOS samples 5.711851119995117 degrees
NLOS
Rejected samples = 10%
Mean error for rejected NLOS samples 80.17779336869717 degrees
Median error for rejected NLOS samples 73.12239074707031 degrees
- When using estimated displacement
LOS:
Rejected sampled = 0.3%
Mean error for rejected LOS samples 2.526263952255249 degrees
Median error for rejected LOS samples 2.526263952255249 degrees
NLOS:
Rejected samples = 10%
Mean error for rejected NLOS samples 77.66558815132488 degrees
Median error for rejected NLOS samples 78.47801971435547 degrees
We use a profile variance threshold of 0.9; any sample with a variance greater than this is treated as an outlier. Indeed we see that the mean and standard deviation of rejected samples has very error.
Improvement is observed in AOA accuracy after applying outlier rejection for both groundtruth displacement (left) and estimated displacement (right).
Coming Soon!