【摘要】 相对于结构化道路来说，非结构化道路的道路特征更少。针对非结构化道路分割问题，本文在UNet模型的基础上，采用加深网络层数和引入ECA注意力机制提高模型的特征提取能力，并在编码阶段融入残差网络代替原模型中的VGG16，在瓶颈阶段加入ASPP扩大模型感受野，在解码阶段加入深度监督模块获得图像各层特征信息，使最后的分割结果更加精准。在非结构化道路数据集上的实验结果表明，本文方法的准确率和平均交并比分别为97.55%和83.22%，比原UNet模型和DeeplabV3模型的性能更好。更多还原

【Abstract】 Compared with structured road, unstructured road has fewer road features and more complex environment.Aiming at the segmentation of unstructured road, a deeper layer of network was adopted and ECA attention mechanism was introduced to improve the feature extraction ability of the model based on UNet. Residual network was integrated into the coding stage of the model to replace VGG16. ASPP was added to expand the receptive field at the bottleneck stage of the model and the depth supervision module was added in the decoding stage of the model to obtain the feature information of each layer of the image, to make the final segmentation results more accurate.The experimental results on the unstructured road data set showed that the accuracy and average intersection ratio of the proposed method were 97.55% and 83.22%,which were higher than the original UNet model and DeeplabV3 model.更多还原