融合自注意力机制和低层特征增强的交通场景语义 分割方法研究

Semantic segmentation technology is a significant research direction in the field of computer vi-sion, aiming to precisely classify each pixel in an image into its corresponding semantic category. However, existing methods have notable limitation in complex traffic scenarios such as severe loss of detail information and insufficient ability to capture object features, resulting in generally low accuracy in target contour segmentation. Therefore, this paper proposes an improved semantic segmentation model based on DeepLabv3+ and integrating the coordinate at-tention mechanism. By optimizing the structure of the original atrous spatial pyramid pooling (ASPP) module and in-troducing in a low-level feature branch, deep semantic information is further fused with shallow spatial information to achieve clearer segmentation results. The improved algorithm outperforms similar algorithms in terms of segmenta-tion accuracy on the Cityscapes traffic scene dataset. Additionally, this paper constructs a scene classification sys-tem based on two dimensions, dividing the Cityscapes dataset into four typical driving scenarios. The proposed mod-el shows significant performance improvement in all four scenarios compared to the original model. Finally, through real vehicle data collection and testing on an autonomous driving platform, the effectiveness of the algorithm optimi-zation in this paper is further verified.