• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文会议>International Conference on Electronics, Information, and Communication >Life Long Across All Four Seasons Scene Understanding
【24h】

Life Long Across All Four Seasons Scene Understanding

机译：整个四个季节的人生经历

获取原文
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

Recently, automatous driving is grabbing everyone's attention. The technical partition of the automatic vehicle driving is scene understanding, driving policy-making and motion control. First and foremost, the scene understanding plays an imperative role. Meanwhile, the scene understanding breaks down into four branches, which are free space scene semantic segmentation, lane detection, moving objects detection and moving objects tracking. Main challenges of all four challenging branches are unrestricted open vocabulary and diverse scenes. This paper has proposed a new theory which combines the improved fully convolutional network (iFCN) and the transfer learning across all four seasons to alleviate all the challenges mentioned above. The proposed iFCN changes the traditional structure of FCN in the pooling layer and replaces the learning rate with a self-adapted and much more reasonable one, which generates a significant improvement on the scene understanding assignment. The global prior representation of transfer learning across all four scenes is to avoid training entire network from scratch with random initialization, which makes the portability of the network improved quite a lot. Experiment results indicate that the proposed theory achieves the state-of-the-art mean IoU on the challenging GAC R&D environment perception benchmark with the value of 86.2%, which gains 34.67% improvement compared with conventional work.
机译：最近，自动驾驶正吸引着所有人的注意力。自动驾驶汽车的技术领域是场景理解，驾驶政策制定和运动控制。首先，对场景的理解起着至关重要的作用。同时，场景理解分为自由空间场景语义分割，车道检测，运动物体检测和运动物体跟踪四个分支。这四个具有挑战性的分支的主要挑战是不受限制的开放词汇和多样化的场景。本文提出了一种新理论，将改进的全卷积网络（iFCN）与跨四个季节的迁移学习相结合，以缓解上述所有挑战。提出的iFCN改变了FCN在池化层中的传统结构，并以一种自适应的，更合理的方式代替了学习率，从而极大地改善了场景理解任务。跨所有场景的转移学习的全局优先表示是避免使用随机初始化从头开始训练整个网络，这使网络的可移植性得到了很大的改善。实验结果表明，该理论在具有挑战性的GAC研发环境感知基准上达到了最先进的平均IoU，值为86.2％，与常规工作相比，提高了34.67％。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号