【摘要】 针对现有的自动驾驶道路场景的语义分割模型在特征提取阶段提取特征不充分以及忽略了不同级别特征图中存在的特征重要性差异进而导致分割结果不佳的问题，提出了一种改进的DeepLab V3+算法，该方法结合了不同的注意力机制；针对原始模型的主干网络结构较为复杂且需要更多的计算资源来训练的问题，将主干网络更换为ResNet50。通过实验验证，本方法平均交并比(MIoU)指数达到了0.73,相较于其他方法具有更高的分割准确度。更多还原

【Abstract】 In response to the problem of insufficient feature extraction in the feature extraction stage of existing semantic segmentation models for autonomous driving road scenes, as well as the neglect of feature importance differences in different levels of feature maps, resulting in poor segmentation results, this paper proposes an improved DeepLab V3+ algorithm that combines different attention mechanisms. In response to the complex backbone network structure of the original model and the need for more computing resources for training, this article replaces its backbone network with ResNet50. Through experimental verification, the MIoU(Mean Intersection over Union) index of this method reaches 0.73, which has higher segmentation accuracy compared to other methods.更多还原