|
| 基于深度卷积神经网络的视觉里程计研究 |
| Research on visual odometry using deep convolution neural network |
| 投稿时间：2019-10-24 |
| DOI：10.3969/j.issn.1005-5630.2020.04.006 |
| 中文关键词: 视觉里程计 自主定位 深度学习 卷积神经网络 |
| 英文关键词:visual odometry self-localization and navigation deep-learning convolutional neural network |
| 基金项目:国家自然科学基金（61374197） |
|
| 摘要点击次数: 2885 |
| 全文下载次数: 1884 |
| 中文摘要: |
| 视觉里程计利用视频信息来估计相机运动的位姿参数，实现对智能体的定位。传统视觉里程计方法需要特征提取、特征匹配/跟踪、外点剔除、运动估计、优化等流程，解算非常复杂，因此，提出了基于卷积神经网络的方法来实现端对端的单目视觉里程计。借助卷积神经网络对彩色图片自动学习提取图像帧间变化的全局特征，将用于分类的卷积神经网络转化为帧间时序特征网络，通过三层全连接层输出相机的帧间相对位姿参数。在KITTI数据集上的实验结果表明，提出的Deep-CNN-VO模型可以较准确地估计车辆的运动轨迹，证明了方法的可行性。在简化了复杂模型的基础上，与传统的视觉里程计系统相比，该模型的精度也有所提高。 |
| 英文摘要: |
| The visual odometry uses visual cues to estimate the pose parameters of the camera motion and localize an agent. Existing visual odometry employs a complex process including feature extraction, feature matching/tracking, and motion estimation. The processing is complicated. This paper presents an end-to-end monocular visual odometry by using convolution neural network (CNN). The method modifies a classification CNN into a sequential inter-frame variation CNN. In virtue of the deep learning technique, the method extracts the global inter-frame variation feature of video images, and outputs pose parameters through three full-connection convolution layers. It has been tested in the public KITTI database. The experimental results show the proposed Deep-CNN-VO model can estimate the motion trajectory of the camera and the feasibility of the method is proved. On the basis of simplifying the complex model, the accuracy is improved compared with the traditional visual odometry system. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
