Leveraging Deep Learning for Visual Odometry Using Optical Flow.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Traditional visual odometry (VO) relies on complex, hand-engineered components.
• Deep learning (DL) offers a promising alternative for automating VO tasks.
• Existing DL methods for VO have limitations in handling scale ambiguity and sequence dynamics.

Purpose of the Study:

• To develop a novel deep learning architecture for monocular visual odometry (VO).
• To improve ego-motion estimation and trajectory generation using sequence-based learning.
• To implicitly learn absolute scale without camera intrinsics or post-calibration.

Main Methods:

• Proposed a hybrid deep neural network architecture combining Convolutional Neural Networks (CNNs) for optical flow-based motion estimation and Recurrent Neural Networks (RNNs) for sequence modeling.
• Network directly outputs relative 6-DOF camera poses.
• Integrated trajectory generation without post-calibration, learning absolute scale implicitly.

Main Results:

• The proposed DL-based VO method demonstrates competitive performance on the KITTI dataset.
• Achieved accurate camera pose estimation and trajectory integration.
• Outperforms or matches traditional and other DL-based VO approaches.

Conclusions:

• Deep learning, particularly CNNs and RNNs, provides an effective framework for monocular visual odometry.
• The proposed architecture successfully addresses scale ambiguity and sequence dynamics.
• This approach offers a robust and automated solution for camera motion estimation.