Tire-Wear Estimation System Using Semantic Segmentation and Its Deployment

Tire-wear significantly impacts vehicle operation and passenger safety, making tire-wear monitoring a critical task. Traditional manual groove-depth measurements are precise but impractical for general drivers. Existing sensor and intelligent tire-based methods also have limitations that require additional equipment. This paper presents a novel Tire-Wear Estimation (TWE) system using mobile phone cameras, which leverages close-up tire videos to estimate individual groove-depths. Due to the lack of an existing dataset that captures tire grooves, we have collected a large number of tire videos with the ground truths of groove-depths to build and evaluate our system. For the semantic segmentation model, we select one from the U-Net family by considering complexity as well as output quality, and develop a post-processing method to improve the quality of masks (segmentation results). After obtaining frame-wise tire-masks from the input video, we measure the width and depth of each <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">dent</i> (the indented parts in the masks). By tracking the dimensions of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">dents</i> over the video frames, we estimate the actual depth of grooves. Additionally, we implement a model lifecycle-based service to improve the performance of our TWE system. Since it is not feasible to inspect all user inputs and their results, we have also developed a mask quality pre-screening method based on mask generation to facilitate the data validation process. The proposed TWE system has shown an absolute error of 0.94 <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mm</i>, with an average latency of 2.44 seconds, to obtain results from tire videos of around 10 seconds.