摘要
利用经典算法强大的特征提取能力以及多特征融合的优势,提出了一种新的神经网络模型,可帮助无人驾驶系统在行驶视觉场景中对周围环境做出更好的理解。具体选择了在ImageNet中预训练得到的ResNet-34和VGG-16两种经典算法作为神经网络的主要路径,分别提取具备不同特点的全局信息与局部信息,加入其他融合模块并引入注意力机制以更好地优化算法性能。经实验验证,与SegNet等其他先验模型相比,对道路场景的分割准确率由原来的65.9%提高至77.3%,在CamVid数据集中表现优秀。
Using the strong feature extraction ability of classical algorithm and the advantage of multi-feature fusion,a new neural network model is proposed. The model can help the driverless system to better understand the surrounding environment in the driving visual scene. Two classical algorithms,ResNet-34 and VGG-16,which are pre-trained in ImageNet,are chosen as the main paths of neural network. They extract global and local information with different characteristics,and then add other fusion modules and attention mechanism to optimize the performance of the algorithm. Compared with other priori models such as SegNet,the segmentation accuracy of road scene is improved from 65. 9% to 77. 3%,which shows excellent performance in CamVid dataset.
作者
叶绿
朱家懿
段婷
YE Lü;ZHU Jiayi;DUAN Ting(School of Information and Electronic Engineering,Zhejiang University of Science and Technology,Hangzhou 310023,China;School of Mechanical and Energy Engineering,Zhejiang University of Science and Technology,Hangzhou 310023,China)
出处
《实验室研究与探索》
CAS
北大核心
2020年第10期88-92,共5页
Research and Exploration In Laboratory
关键词
深度学习
行驶视觉
语义分割
多特征融合
deep learning
driving vision
semantic segmentation
multi-feature fusion