面向交通场景的图像分割网络

Image Segmentation Network for Traffic Scenes

摘要: 语义分割技术是自动驾驶的重要基础技术之一。在交通场景的图像语义分割中，图像语义分割希望对图像的每一个像素进行分类，并进行颜色标注，使车辆能够准确地检测道路上的交通参与者和可以行驶的路面区域。目前，典型的图像语义分割算法通常融合骨干网络生成的不同阶段的特征图，以提高分割性能。简单的融合方法不能充分利用这些特征信息。这将导致相似物体之间的分割错误和小物体边界分割粗糙，使车辆无法准确感知周围环境，从而影响上层的决策，甚至对其他交通参与者造成严重的安全隐患。针对这一问题，本文设计了一种密集特征融合与边界细化网络(DFBNet)，它包括两部分：特征融合网络(FFN)利用多个不同感受野大小的分支和卷积核提取特征图中的信息，并利用注意力机制为提取的特征分配权重；边界细化网络(BRN)利用空间注意力对每个像素位置赋予给予权值，使得目标对象的边界区域分割得更精细。我们在两个数据集上进行实验：Cityscapes和Camvid数据集。我们取得了良好的分割结果，在Cityscapes验证集上的平均交并比(mIoU)为79.47%。在Camvid测试集上的mIoU为75.13%。

Abstract: This Semantic segmentation technology is one of the important basic technologies of automatic driving. In automatic driving, image semantic segmentation hopes to classify every pixel of the image and make color labeling, so that the vehicle can accurately detect the traffic participants on the road and the pavement area that can be used. At present, typical image semantic segmentation algorithms usually fuse the feature maps of different stages generated by the backbone network to improve segmentation performance. The simple fusion method cannot fully utilize this feature information. This will result in segmentation errors between similar objects and rough boundary segmentation of small objects, making the vehicle unable to accurately perceive the surrounding environment, thus affecting the decision-making of the upper level, and even causing serious safety hazards to other traffic participants. To solve this problem, this paper designs a dense feature fusion and boundary refinement network (DFBNet), which includes two parts: Feature Fusion Network (FFN) using multiple branches and convolutional kernels of different receptive field sizes to extract the information in the feature map and using the attention mechanism to assign weights to the extracted features; Boundary Refinement Network (BRN) using spatial attention to give weights to each pixel position, making the boundary area of the target object more finely segmented. We experimented on two datasets: Cityscapes and Camvid. We achieved good segmentation results with an average intersection (mIoU) of 79.47% on the Cityscapes validation set and 75.13% on the Camvid test set.