Grid-level Multiclass Soiling Classification Model for Driving Systems

This study proposes a soiling detection algorithm to identify and locate contamination in vehicle camera lenses. Research on AI applications that utilize cameras and distance sensors in driving systems is directly related to the advancement of autonomous driving systems., Detecting soiling, such as mud and water droplets, is a particularly critical issue. Traditional methods using piezoelectric, ultrasonic, and thermal sensors can introduce additional design and maintenance complexity. Therefore, this study aims to reduce this complexity by utilizing existing surround-view cameras installed on vehicles without additional sensors. The proposed algorithm employs image-processing techniques and a lightweight neural network architecture based on ResNet18 to detect lens contamination in real-time across various driving environments. Experiments were conducted using 5,000 images from the WoodScape Soiling Dataset. The input images were divided into a 16 × 16 grid and classified into four labels: opaque, semitransparent, transparent, and clean. The proposed grid-level multiclass soiling classification model demonstrated effective performance in detecting contamination, including mud, water droplets, and foggy dust. This study is expected to enhance the safety and convenience of driving systems.