FlowSegModel: Advancing Perception in Autonomous Driving Through Weather-Resilient Segmentation

Abstract: This paper introduces a novel methodology for semantic segmentation in adverse weather conditions, employing a FlowSegModel architecture that integrates RGB images with optical flow inputs within a modified DeepLabV3 framework, rigorously evaluated on the synthetic VKITTI2 dataset. The FlowSegModel adapts the DeepLabV3 backbone, by reconfiguring its initial convolutional layer to accommodate a 5-channel input (3 RGB + 2 optical flow), enabling seamless fusion of spatial and motion features. This is complemented by a tailored Atrous Spatial Pyramid Pooling (ASPP) module, optimized with adjusted dilation rates to capture multi-scale context under degraded visibility. The proposed FlowSegModel achieves mean Intersection over Union (IoU) scores of 0.8406 under normal conditions and 0.8034 in adverse scenarios (fog, rain, overcast), demonstrating a robust 4.4% degradation. This resilience is attributed to optical flow’s complementary motion cues, enhancing visibility in degraded environments. Comparative analyses reveal that fine-tuned SEA-RAFT and RAFT models outperform classical methods (e.g., Farneback, Horn-Schunck) by over 96% in Average End-Point Error (EPE) reduction, with SEA-RAFT excelling by 56.6% over RAFT. An ablation study confirms the superiority of RGB-flow fusion, yielding a 46.1% mean IoU improvement over an optical-flow-only variant, which struggles with static scene discrimination absent color and texture data. These findings hold significant implications for autonomous driving and outdoor vision systems, offering a practical approach to weather-invariant perception that may reduce real-world failure rates. This work advances robust segmentation techniques, contributing to safer and more reliable environmental perception