融合语义流的门控自适应交通场景语义分割算法OA

Focusing on the existing traffic scene semantic segmentation algorithms being difficult to accurately segment the object body and object edges at the same time,resulting in small objects being easily integrated into the surrounding background and discontinuous segmentation of slender stripes,a gated adaptive traffic scene semantic segmentation algo-rithm with fused semantic streams is proposed.Firstly,a parallel complementary adaptive gating module based on local and global is constructed.The complementary feature information extracted by adaptive fusion of multi-head self-atten-tion mechanism and deep convolution is used to overcome the similar feature encoding problem between different catego-ries and enhance the discriminative representation between different objects.Secondly,taking advantage of the receptive field advantage and inductive bias characteristics of convolution,a lightweight local feature extraction module based on deep separable convolution is constructed to capture local detail features of feature maps at multiple scales to enhance the network's representation ability for small objects while avoiding network structural redundancy.Finally,a semantic flow-guided cross-layer aggregation module is designed to align and aggregate adjacent feature maps under downsampling operations to further model the global context and alleviate the problem of discontinuous segmentation of thin and long objects.Experi-mental results on the traffic scene datasets Cityscapes and CamVid demonstrate that the proposed algorithm can achieve 80.16％and 84.78％mIoU,and can improve the segmentation accuracy of small objects and thin strips while refining the object segmentation edges.