Benchmarking Vision-based Analog Gauge Reading Methods for Vehicle Dashboard Validation

Automated vehicular dashboard validation and verification play a key role in the rapid integration of high-precision dashboards. Vision-based methods for accurately estimating tachometer and speedometer readings, validated against internal vehicle signals, are met with two major challenges: limited training data and the selection of the most effective gauge estimation method. To address these challenges, we leverage automated segmentation and tracking via the Segment Anything Model and DeAOT (SAM-track) to streamline the gauge labeling process required for training machine learning models, facilitating the rapid creation of a training dataset. This dataset is then used to train a Detectron 2 model for segmenting speedometer and tachometer gauges. For angle estimation, we train three different machine learning models on augmented data to predict gauge angles based on segmented masks. Additionally, our pipeline includes conventional rule-based angle estimation methods. Preliminary benchmarking results highlight ResNet as the most accurate method, achieving a Mean Absolute Error (MAE) of 0.86 km/h for speed estimation and 49.74 rounds per minute (RPM) for tachometer readings. The study further examines how different resolution levels and visual distortions impact performance, two common issue in real-world applications. When gauge masks are resized between ¼ and 4 times their original resolution, traditional rule-based methods, such as ellipse fitting, demonstrate greater robustness in maintaining cross-resolution accuracy compared to machine learning models, highlighting their reliability for real-world applications.