一种轻量化非结构化道路语义分割神经网络

非结构化道路由于没有明显车道线且道路特征多、地域差异大，现有的结构化道路分割方法无法满足非结构化道路分割在实际应用中的实时性与准确性要求．为了解决上述难点，本文基于DeepLabv3+网络提出一种G-lite-DeepLabv3+网络结构，使用Mobilenetv2网络替换解码器中的Xception特征提取网络，并通过在Mobilenetv2网络与空洞空间金字塔池化模块中使用分组卷积替换普通卷积，且有选择地取舍批规范层来减少参数量，在不影响精度的同时提升分割效率．同时针对非结构化道路在图像里分布位置相对较固定的特点，引入注意力机制对高级语义特征进行处理，提升网络对有用特征的敏感度与准确性．选用与我国非结构化道路路况相似的印度道路驾驶IDD进行训练，并与其他经典语义分割网络进行实验对比，结果表明，相比于其他网络，本文提出的G-lite-DeepLabv3+准确率与实时性均表现较好、误分割与边缘清晰度均好于对照网络；与经典算法进行对比，平均交并比mIoU提升1.3%，平均像素精度提升6.2%，帧率提升22.1%.

A lightweight unstructured road semantic segmentation neural network

In unstructured roads, there is no obvious lane line, many road characteristics and large regional differences. As a result, the existing structured road segmentation methods can not meet the real-time and accuracy requirements of unstructured road segmentation in practical application.To solve these problems, a new neural network called G-lite-DeepLabv3+ is proposed based on the DeepLabv3+ network. Specifically, the Xception network is replaced by the Mbilenetv2 network in the decoder, the convolutions in Mobilenetv2 and ASPP are replaced by group convolution and the batchnorm layer is chosen selectively to reduce the amount of parameters, improving the segmentation efficiency without affecting the accuracy; at the same time, attention mechanism is introduced to deal with high-level semantic features to improve the sensitivity and accuracy of the network to useful features, considering relatively fixed distribution position of unstructured roads in the image. India driving dataset(IDD) is chosen to train the model taking into account that the roads included in the dataset are similar to the unstructured road in China. The established model is compared with other classical semantic segmentation networks, and the results show that the accuracy and real-time performance of G-lite-deeplabv3+ proposed in this paper are better than those of other networks. The proposed network also outperforms other networks on the indices of improper segmentation and edge clarity. Compared with the traditional network, the mean Intersection over Union(mIoU) is improved by 1.3%; the average pixel accuracy(mPA) is improved by 6.2% and the frame per second (FPS) is improved by 22.1%.