Modeling and estimating lidar intensity for automotive surfaces using gaussian process regression: An experimental and case study approach

Abstract

LIDAR technology is widely used in autonomous driving and environmental sensing, but its accuracy is significantly affected by variations in vehicle surface reflectivity. This study models and predicts the impact of different LIDAR sensor specifications and vehicle surface paints on laser intensity measurements. Laser intensity data from the experiments of Shung et al. were analyzed alongside vehicle color, angle, and distance. Multiple machine learning models were tested, with Gaussian Process Regression (GPR) performing best (RMSE = 0.87451, R2 = 0.99924). To enhance the model's physical interpretability, laser intensity values were correlated with LIDAR optical power equations, and curve fitting was applied to refine the relationship. The model was validated using the input parameters from Shung et al.'s experiments, comparing predicted intensity values with reference measurements. The results show that the model achieves an overall accuracy of 99% and is successful in laser intensity prediction. To assess real-world performance, the model was tested on the CUPAC dataset, which includes various traffic and weather conditions. Spatial filtering was applied to isolate laser intensities reflected only from the vehicle surface. The highest accuracy, 98.891%, was achieved for the SW-Gloss (White) surface, while the lowest accuracy, 98.195%, was recorded for the SB-Matte (Black) surface. The results confirm that the model effectively predicts laser intensity across different surface reflectivity conditions and remains robust across different channels LIDAR systems.