PAVEsense - IRI Correlation: Difference between revisions
English source text for translation |
English source text for translation |
||
| (2 intermediate revisions by the same user not shown) | |||
| Line 1: | Line 1: | ||
<languages/> | <languages/> | ||
<translate> | <translate> | ||
<!--T:1--> | |||
[[File:Port of Antwerp Groundtruth vs eIRI measurements of ASAsense.jpg|thumb|eIRI ASAsense vs IRI reference]] | [[File:Port of Antwerp Groundtruth vs eIRI measurements of ASAsense.jpg|thumb|eIRI ASAsense vs IRI reference]] | ||
<!--T:2--> | |||
''Valid from PAVEsense version 8 (November 2024)'' | ''Valid from PAVEsense version 8 (November 2024)'' | ||
The ASAsense eIRI values are calculated from the road profile measured with the accelerometer in the sensor box. We combine the measurements of the different passes and take the '''average''' eIRI of all passes that meet the [[ | <!--T:3--> | ||
The ASAsense eIRI values are calculated from the road profile measured with the accelerometer in the sensor box. We combine the measurements of the different passes and take the '''average''' eIRI of all passes that meet the [[PAVEsense/en#Convergentie|convergence criterion]]. | |||
==== Comparison with reference measurements in the Port of Antwerp ==== | ==== Comparison with reference measurements in the Port of Antwerp ==== <!--T:4--> | ||
<!--T:5--> | |||
The correlation was made using reference data that we kindly received from the Port of Antwerp. After applying some corrections for GPS drift, we observe a correlation (R² factor) of 70%. Given that we measure after the damping of the vehicle and across several passes, whereas the reference data is recorded once and always along one side of the vehicle, we consider this a good correlation factor. More important, however, is the effect on the analysis. There is a very good similarity between the reference data and the ASAsense data: the worse roads in the reference data also come out as bad in the ASAsense data, and the good ones as good. | The correlation was made using reference data that we kindly received from the Port of Antwerp. After applying some corrections for GPS drift, we observe a correlation (R² factor) of 70%. Given that we measure after the damping of the vehicle and across several passes, whereas the reference data is recorded once and always along one side of the vehicle, we consider this a good correlation factor. More important, however, is the effect on the analysis. There is a very good similarity between the reference data and the ASAsense data: the worse roads in the reference data also come out as bad in the ASAsense data, and the good ones as good. | ||
<!--T:6--> | |||
[[File:Distributie eIRI vs referentie.png|thumb|eIRI ASAsense compared with the PoA distribution]] | [[File:Distributie eIRI vs referentie.png|thumb|eIRI ASAsense compared with the PoA distribution]] | ||
<!--T:7--> | |||
When we look at the overall distribution of the data, we also see a very strong agreement. | When we look at the overall distribution of the data, we also see a very strong agreement. | ||
<!--T:8--> | |||
Below we have also compared the "driving quality" indicator with the reference measurements. We analyse this with a "confusion matrix", a standard approach for classification problems. Quality is mainly predicted in neighbouring quality labels, as the confusion matrix shows. For example, of the roads whose true label is "very good", 96% are predicted as "good" or "very good". | Below we have also compared the "driving quality" indicator with the reference measurements. We analyse this with a "confusion matrix", a standard approach for classification problems. Quality is mainly predicted in neighbouring quality labels, as the confusion matrix shows. For example, of the roads whose true label is "very good", 96% are predicted as "good" or "very good". | ||
==== Development ==== | ==== Development ==== <!--T:9--> | ||
===== PAVEsense version 8 ===== | ===== PAVEsense version 8 ===== <!--T:10--> | ||
<!--T:11--> | |||
The alignment of the measured data was improved. As a result, peaks now coincide better at the correct location. This makes averaging over different passes a more effective way to obtain the final result. Moreover, it is a simpler method and also works when there are few passes over the same location. The following table shows the R² of the new version compared with the previous one. We also see that the prediction of ''Driving Quality'' has improved significantly compared with the previous version. The table was created with the same data for v7 and v8, which makes comparison possible. | The alignment of the measured data was improved. As a result, peaks now coincide better at the correct location. This makes averaging over different passes a more effective way to obtain the final result. Moreover, it is a simpler method and also works when there are few passes over the same location. The following table shows the R² of the new version compared with the previous one. We also see that the prediction of ''Driving Quality'' has improved significantly compared with the previous version. The table was created with the same data for v7 and v8, which makes comparison possible. | ||
<!--T:12--> | |||
{| class="wikitable" | {| class="wikitable" | ||
! eIRI version | ! eIRI version | ||
| Line 34: | Line 43: | ||
|} | |} | ||
<!--T:13--> | |||
[[File:Driving quality confusion matrix.png|thumb|eIRI driving quality confusion matrix, normalised on the true label]] | [[File:Driving quality confusion matrix.png|thumb|eIRI driving quality confusion matrix, normalised on the true label]] | ||
===== PAVEsense version 7 and earlier ===== | ===== PAVEsense version 7 and earlier ===== <!--T:14--> | ||
<!--T:15--> | |||
In PAVEsense v7 and earlier, the 85<sup>th</sup> percentile was used to combine different passes into one measurement. This version under-reported the measured eIRI, so many roads ended up in a better category than in reality. | In PAVEsense v7 and earlier, the 85<sup>th</sup> percentile was used to combine different passes into one measurement. This version under-reported the measured eIRI, so many roads ended up in a better category than in reality. | ||
<!--T:16--> | |||
[[File:PAVEsense v7 driving quality confusion matrix.png|thumb|PAVEsense v7 (older software) driving quality confusion matrix, normalised on the true label]] | [[File:PAVEsense v7 driving quality confusion matrix.png|thumb|PAVEsense v7 (older software) driving quality confusion matrix, normalised on the true label]] | ||
</translate> | </translate> | ||
Latest revision as of 12:07, 14 October 2026
Valid from PAVEsense version 8 (November 2024)
The ASAsense eIRI values are calculated from the road profile measured with the accelerometer in the sensor box. We combine the measurements of the different passes and take the average eIRI of all passes that meet the convergence criterion.
Comparison with reference measurements in the Port of Antwerp
The correlation was made using reference data that we kindly received from the Port of Antwerp. After applying some corrections for GPS drift, we observe a correlation (R² factor) of 70%. Given that we measure after the damping of the vehicle and across several passes, whereas the reference data is recorded once and always along one side of the vehicle, we consider this a good correlation factor. More important, however, is the effect on the analysis. There is a very good similarity between the reference data and the ASAsense data: the worse roads in the reference data also come out as bad in the ASAsense data, and the good ones as good.
When we look at the overall distribution of the data, we also see a very strong agreement.
Below we have also compared the "driving quality" indicator with the reference measurements. We analyse this with a "confusion matrix", a standard approach for classification problems. Quality is mainly predicted in neighbouring quality labels, as the confusion matrix shows. For example, of the roads whose true label is "very good", 96% are predicted as "good" or "very good".
Development
PAVEsense version 8
The alignment of the measured data was improved. As a result, peaks now coincide better at the correct location. This makes averaging over different passes a more effective way to obtain the final result. Moreover, it is a simpler method and also works when there are few passes over the same location. The following table shows the R² of the new version compared with the previous one. We also see that the prediction of Driving Quality has improved significantly compared with the previous version. The table was created with the same data for v7 and v8, which makes comparison possible.
| eIRI version | R² (dtw-aligned) |
|---|---|
| PAVEsense v7 and earlier | 0.61 |
| from PAVEsense v8 | 0.7 |
PAVEsense version 7 and earlier
In PAVEsense v7 and earlier, the 85th percentile was used to combine different passes into one measurement. This version under-reported the measured eIRI, so many roads ended up in a better category than in reality.
