Efficient and Robust Crosswalk Segmentation under Adverse Weather Using ConvNeXt-Enhanced DeepLabv3

Abstract

Reliable crosswalk perception is crucial for first-person vision (FPV) navigation in assistive guidance and intelligent transportation, but segmentation accuracy often decreases under glare, rain reflections, nighttime illumination, and worn low-contrast markings. This study proposes ConvNeXt-Enhanced DeepLabv3 (CEDL), a pixel-level segmentation architecture that integrates DeepLabv3 atrous multiscale encoding with the modern convolutional design of ConvNeXt-Tiny. Experiments were conducted on the FPVCrosswalk2025 dataset, containing synthetic and real FPV images captured under sunny, cloudy, rainy, and night conditions. The proposed model was compared with DeepLabv3 using ResNet50 and MobileNetV3-L backbones under the same training and evaluation protocol. CEDL achieved the best overall performance, with 0.946 mean IoU and 0.972 Dice, while maintaining strong percondition robustness and improved boundary preservation for thin crosswalk structures. It also achieved practical inference speed at 20.6 ms per frame, nearly five times faster than ResNet-50, despite having more parameters than MobileNetV3-L. Qualitative results show more continuous crosswalk stripes and fewer missed segments under adverse conditions. These findings indicate that CEDL provides a robust and computationally practical solution for FPV crosswalk segmentation on a mixed synthetic-real benchmark.