University Archives

Advanced Driver Assistance Systems (ADAS) often rely on multiple sensors such as LiDAR, radar, and cameras to interpret the driving environment. While these systems perform well on standard highways, off-road environments introduce shadows which can significantly degrade image-based segmentation. In gravel or dirt road scenes, shadows can cause the same surface to appear as multiple colors or intensity levels. As a result, segmentation algorithms frequently misclassify shadowed portions of the road surface as surrounding vegetation or terrain.

This poster presents an image-based approach that incorporates a shadow-mitigation method to improve segmentation consistency in such scenes. Shadows are detected using the Hue-Saturation-Value (HSV) color model through binarization of dark regions, followed by geodesic dilation to capture both deep and lighter shadow areas. Shadowed pixels are then intensity-corrected using ratios derived from RGB statistics of shadow and non-shadow road regions. The correction is applied iteratively to accommodate varying shadow strengths and is followed by Gaussian smoothing to reduce boundary artifacts.

By normalizing shadowed road regions while preserving texture structure, the proposed method produces more consistent image features and improves the reliability of subsequent segmentation in off-road imagery.