多尺度特征提取的道路场景语义分割

SEMANTIC SEGMENTATION OF ROAD SCENE BASED ON MULTI-SCALE FEATURE EXTRACTION

道路场景语义分割是自动驾驶系统的重要组成部分。道路场景中环境复杂、物体种类繁多且尺寸差异较大,已有的全卷积神经网络(FCN)特征提取能力不足,导致语义分割精度较低。对此,提出一种多尺度特征提取网络(Multi-scale Feature Extraction Network, MFNet),该网络采用并行的特征提取模块提取不同尺度下的不变特征,增强特征多样性,通过逐层的反卷积操作,将特征上采样恢复至原始图像大小,设计分级训练方法并优化loss函数。在多个公开数据集上对该算法进行评估,取得了良好的分割效果。

Semantic segmentation of road scenes is an important aspect of autonomous driving systems. The existing fully convolutional networks(FCN) has insufficient feature extraction ability in the context of complex environment, various types of objects and large size differences in road scenes, resulting in low semantic segmentation accuracy. We presented a multi-scale feature extraction network(MFNet). MFNet used parallel feature extraction module to extract invariant features at different scales to enhance the feature diversity, and then restored the feature of upsampling to the original image size by layer-by-layer deconvolution operation to obtain the segmentation result. Finally, we designed a hierarchical training method and optimized the loss function. We evaluate the proposed method on a number of public datasets and achieve a satisfactory segmentation effect.