# ekf-imu-depth **Repository Path**: whitbyli/ekf-imu-depth ## Basic Information - **Project Name**: ekf-imu-depth - **Description**: 单目相机 + IMU 深度估计神经网络 - **Primary Language**: Python - **License**: MulanPSL-2.0 - **Default Branch**: master - **Homepage**: None - **GVP Project**: No ## Statistics - **Stars**: 3 - **Forks**: 1 - **Created**: 2022-10-03 - **Last Updated**: 2023-10-06 ## Categories & Tags **Categories**: Uncategorized **Tags**: 单目深度估计, 深度学习 ## README # DynaDepth #### 介绍 DynaDepth 是 2022 年悉尼大学团队提出的基于卡尔曼滤波器将IMU与单目相机融合估计图像深度的神经网络架构,官方源码地址:https://github.com/SenZHANG-GitHub/ekf-imu-depth.git 这里,根据论文介绍的 IMU 运动估计原理,以 KITTI 数据集为例,以 Jupyter Notebook 形式,梳理了 IMU 运动估计的过程,详细了解请见 [imu_odometry.ipynb](imu_odometry.ipynb) ### 模型训练 模型训练使用 KITTI 官方原始数据集,训练及验证数据集选取了大约 4.4 万帧相机图像,具体选取的子数据集及对应图像帧见文件 `splits/eigen_zhou/train_files.txt` 和 `splits/eigen_zhou/val_files.txt`,对应于官方数据集序列为: ```bash 2011_09_26/2011_09_26_drive_0001_extract 2011_09_26/2011_09_26_drive_0005_extract 2011_09_26/2011_09_26_drive_0011_extract 2011_09_26/2011_09_26_drive_0014_extract 2011_09_26/2011_09_26_drive_0015_extract 2011_09_26/2011_09_26_drive_0017_extract 2011_09_26/2011_09_26_drive_0018_extract 2011_09_26/2011_09_26_drive_0019_extract 2011_09_26/2011_09_26_drive_0022_extract 2011_09_26/2011_09_26_drive_0028_extract 2011_09_26/2011_09_26_drive_0032_extract 2011_09_26/2011_09_26_drive_0035_extract 2011_09_26/2011_09_26_drive_0039_extract 2011_09_26/2011_09_26_drive_0051_extract 2011_09_26/2011_09_26_drive_0057_extract 2011_09_26/2011_09_26_drive_0061_extract 2011_09_26/2011_09_26_drive_0070_extract 2011_09_26/2011_09_26_drive_0079_extract 2011_09_26/2011_09_26_drive_0087_extract 2011_09_26/2011_09_26_drive_0091_extract 2011_09_26/2011_09_26_drive_0095_extract 2011_09_26/2011_09_26_drive_0104_extract 2011_09_26/2011_09_26_drive_0113_extract 2011_09_28/2011_09_28_drive_0001_extract 2011_09_29/2011_09_29_drive_0004_extract 2011_09_29/2011_09_29_drive_0026_extract 2011_09_30/2011_09_30_drive_0020_extract 2011_09_30/2011_09_30_drive_0028_extract 2011_09_30/2011_09_30_drive_0033_extract 2011_09_30/2011_09_30_drive_0034_extract 2011_10_03/2011_10_03_drive_0034_extract 2011_10_03/2011_10_03_drive_0042_extract ``` 训练完成生成的模型保存至 `models` ### KITTI 数据集深度信息生成 KITTI 数据集深度生成代码为 `generate_kitti_depth.py`,详细过程见 [imu_depth](imu_depth.ipynb),这里生成 `2011_09_26/2011_09_26_drive_0009_extract` 数据集深度展示结果如下所示: ![depth_vedio](imgs/depth_video.gif)