A High-Performance Deep Learning Model for Autonomous Driving Segmentation Under Challenging Conditions

Abstract

The recent success of Semantic Segmentation techniques in highway driving datasets has generated significant interest in various related application fields. In particular, many of these applications require real-time prediction capabilities. However, the reliability of these systems remains suboptimal under challenging conditions, such as adverse weather, varying lighting, and complex situations characterized by numerous overlapping objects. This paper proposes a novel model, SceneSegAttentionNet, specifically designed to effectively enhance urban landscape segmentation. The model integrates Coordinate Attention to encode spatial information, Triplet Attention to capture interactions between spatial and channel dimensions of the input, and Outlook Attention to encode fine-level features and contexts into tokens. These components have been demonstrated to be critical in enhancing recognition performance. On the cityscape dataset, SceneSegAttentionNet achieves the highest MIoU of 80.2% while maintaining high inference performance at 75.3 FPS. Compared to STDC-2-Seg75, RegSeg, and SFNet, SceneSegAttentionNet achieves an MIoU improvement of 3.2%, 2.1%, and 1.2%, respectively, while offering faster inference speeds. On the Shandong Highway dataset, SceneSegAttentionNet establishes a new benchmark in the speed-accuracy trade-off for Semantic Segmentation, achieving an MIoU of 75.3% and surpassing state-of-the-art models such as STDC-2-Seg75, RegSeg, and SFNet, by significant margins of 5.1%, 4.1%, and 7.3%, respectively.

Keywords: Adverse weather conditions, Coordinate Attention, Outlook Attention, Real-time prediction, Semantic segmentation, Triplet Attention