Objective Tire Footprint Segmentation Assessment from High-Speed Videos
The tire establishes the contact between the vehicle and the road. It transmits all forces and moments to the road via its contact patch or footprint and vice versa. The visual inspection of this contact patch using modern optical equipment and image processing techniques is essential for evaluating tire performance. Quantitative image-based analysis can be useful for accurate determination of tire footprint under various operating conditions. Very frequently, methods used in tire footprint segmentation cannot be assessed quantitatively due to the lack of a reference contact area to which the different algorithms could be compared. In this work, we present a novel methodology to characterize the dynamic tire footprint and evaluate the quality of its segmentation from various video sequences in the absence of a ground truth.ABSTRACT
ATM scheme. A tire rolls from right to left while its movement is captured with a high-speed camera placed underneath the glass plate. Note that the darkest areas depict the tire footprint.
Set of points, pθ, on the border of the segmentation of the tire footprint (blue contour) are computed using RFD. These points are used to align the curves and obtain the ground truth.
(left) Set of points obtained for all the curves during a video sequence. (right) Aligned curves are used to compute the statistical model. Black marks represent the average point in every direction of θ. Blue and green contours represent the extreme deformations of the mean shape using the eigenvalues K−λ and K+λ, respectively.
Ground truth. (left) Curves KS, which wrap the tire footprints, are computed throughout the video sequence. (center) Curves are aligned (blue contours), and the mean shape, Kμ, and the extremal variations E+λ and E−λ (red and green contours, respectively) are calculated. (right) Mean shape, Kμ, is used as the ground truth of the video sequence.
Tire footprint segmentation comparison. Bold values represent the best scores. DI and RFE are the average errors of the sequence. Original image column shows the frame extracted at the middle of the sequence.
Sequence segmentation after contrast enhancement. ANOVA compares four unsupervised methods: Otsu's threshold, GMM, K-means++, and Fuzzy C-means. (left) RFE and (right) DI. In both cases, Otsu's method presents the smallest error and variance.
Comparison of tire footprint segmentation with contrast enhancement OTSU11 and without contrast enhancement OTSU00 using (left) RFE and (right) DI. Note that the influence of contrast enhancement reduces the number of outliers.
Contributor Notes