Research on Semantic Segmentation Method Based on Improved BiSeNetV2 Model

Leveraging the continuous advancements in fully convolutional networks, deep learning approaches have progressively superseded traditional image processing methods for low-level feature segmentation, thereby improving processing efficiency and enhancing the generalization ability of image segmentation. Semantic segmentation, in particular, refines pixel-level semantic information in annotated images, enriching perceptual map information and effectively enhancing system environmental understanding. Point cloud semantic maps, based on high-precision geometric maps, integrate fine-grained information from image semantic segmentation to achieve region-specific annotations, thereby enhancing the expressiveness of maps. As embedded devices for autonomous driving advance, lightweight and high real-time performance algorithms have become a new trend. Common segmentation models, such as encoder-decoder networks and dilated convolution networks, offer good segmentation accuracy but often face challenges with high memory usage and computational costs, making them less suitable for lightweight requirements in autonomous driving devices. BiSeNetV2, as a real-time semantic segmentation model with a bilateral structure, effectively balances segmentation accuracy, speed, and network complexity. Therefore, this paper enhances the BiSeNetV2 model by integrating the CBAM and ECA attention mechanisms and replacing the GSConv convolutional module, thereby improving the segmentation accuracy and robustness of the semantic segmentation network.