基于自注意和对抗学习的道路场景水体检测方法

Water Puddle Detection Method for Road Scene Based on Self-Attention and Adversarial Learning

无人驾驶车辆近年来一直是研究的热点.无人车运行环境复杂、不确定因素多,尤其当其意外驶入水坑、泥潭等地形时可能直接导致抛锚,造成不可估量的损失,因此水体检测对无人车的运行有着重要意义.借助深度网络的强大学习能力,本文首先将反射注意力单元和自注意力机制相结合,并在(U shape Network,U-Net)模型基础上添加残差卷积块和上采样卷积模块,得到了新的道路场景水体区域检测模型(U shape Network with Attention for Road,URA-Net),该模型能够更好地捕捉特征依赖关系,提高水体语义特征的表示能力.进一步,本文提出了一种基于双生成器对抗学习的训练模型(Redundant With Dual Generative Adversarial Network,RWD-GAN),它对URA-Net稍做修改,拓展成两个生成器,通过在对抗网络框架下让生成器与鉴别器、生成器与生成器之间实现对抗学习,促进不同网络模型之间的信息传递.在公开数据集上的大量实验表明URA-Net达到了87.18%的F1指标,而RWD-GAN模型能够进一步提高水体检测的精度,使提升到了88.54%,URA-Net和RWD-GAN均超出现有深度网络水体检测方法的性能表现.

There has been much interest in the research of self-driving cars.Yet the detection of potentially danger⁃ous obstacles on road makes this investigation more challenging.Water puddles,a typical obstacle of this kind,could trap a self-driving car or even cause serious accidents.Therefore,detecting water puddles is of great importance.To this end,this paper propose a novel water puddle detection model,URA-net(U shape Network with Attention for Road).Building its back⁃bone on U-net(U shape Network)with residual and upsampling blocks added,URA-net combines both the reflection atten⁃tion units and self-attention units,which can better characterize the dependence among image features so as to improve the representative capability to locate water puddles.Furthermore,a two-generator conditional adversarial network RWD-GAN(Redundant With Dual Generative Adversarial Network)is proposed,where two URA-Nets with a minor revision become the two generators to facilitate the information interaction in the adversarial learning process between the generators and the discriminator,as well as between the two generators themselves.Experiments on the public water puddle dataset demon⁃stration that URA-net achieves 87.18%measure,while RWD-GAN can further improve the accuracy of URA-net,pushing F1-score to 88.54%.Both URA-net and RWD-GAN outperforms the state-of-the-arts.