Research on Structured Road Segmentation Algorithm Based on Improved Deeplabv3+

To address the demands of semantic segmentation for complex road markings in autonomous driving scenarios, this paper presents an optimized DeeplabV3+ model aimed at enhancing segmentation accuracy and real-time performance. First, MobileNetV3 replaces Xception as the backbone feature extraction network, leveraging its lightweight architecture and Squeeze-and-Excitation attention mechanism to significantly reduce computational complexity and boost efficiency on resource-constrained devices. Additionally, the ACMix module is incorporated to improve the model's capacity for capturing both local and global features through dynamic fusion of convolution and self-attention mechanisms. Experimental results demonstrate that the modified model achieves a 20.5% increase in mean Intersection over Union (mIoU) compared to the traditional DeeplabV3+. The model also attains a detection speed of 70 fps, outperforming both Xception and MobileNetV2 versions, and exhibits enhanced robustness under challenging conditions like uneven lighting and heavy occlusion. The proposed model not only achieves high segmentation accuracy but also supports realtime inference, providing reliable technical support for visual perception in autonomous driving within complex traffic environments.