基于动态适应的交叉伪标签自动驾驶语义分割算法

Semi-supervised Semantic Segmentation Based on Cross Pseudo Supervision for Autonomous Driving

自动驾驶已经成为了现代社会热门的研究方向之一,其高度集成半导体产业和人工智能产业最新研究的特性使得它成为当下最前沿的应用领域,但是由于算法对于大量标注数据的需求和对算法识别精度的要求日益增长,依赖于人工标注的真实数据进行算法的优化和落地越来越困难。本文采用双分支的网络结构构建卷积神经网络,并进一步使用动态适应的优化方案对网络的鲁棒性进行增强。通过使用交叉指导学习的方法对网络进行更新,提高了测试指标和整体算法性能。一系列实验结果表明:提出的改进方法相比于现有的方法具有更好的性能,在由真实行驶场景构建的数据集Cityscape上验证了本文方法的有效性,与现有方法相比预测准确率更高。

Autonomous driving has became a trendy research field in the modern world.Assembling the latest productions of semi-conductor and artificial intelligence industry,it’s definitely one of the most cutting-edge technology.Due to the increasing demand for training data and precision a step further,it’s hard to deploy and optimize algorithms merely depending on real scenes dataset annotated by human.In this paper,a convolutional neural network based on teacher-student model is adopted,enhanced in robustness by using dynamic adaption.By using exponential moving average to update the parameter of the network,a series of experimental results show that the proposed improved method has better performance and generalization effect than the existing method,and the validity of the proposed method is verified on the real data set of autonomous driving,Cityscape,and the segmentation accuracy is higher than that of the existing network.