Visual odometry (VO) is the real-time estimation of a camera or robot trajectory from sequential visual observations - it computes incremental motion between frames to track pose as the agent moves through an environment.

What Is Visual Odometry?

• Definition: Estimate relative translation and rotation over time from camera input.
• Input Types: Monocular, stereo, or RGB-D image streams.
• Output: Incremental and integrated trajectory in 3D space.
• Difference from SfM: VO prioritizes online incremental updates for real-time operation.

Why Visual Odometry Matters

• Navigation Core: Provides motion estimate for autonomous platforms.
• Low Infrastructure: Works without external localization beacons.
• Sensor Flexibility: Runs on camera-only hardware for lightweight systems.
• Foundation for SLAM: Supplies front-end motion estimates before global map correction.
• Deployment Utility: Used in drones, AR devices, and mobile robots.

VO Approaches

Feature-Based VO:

• Track keypoints and solve geometric motion from correspondences.
• Robust under moderate texture and lighting.

Direct VO:

• Optimize photometric consistency over pixels.
• Uses more image information but sensitive to illumination shifts.

Learned VO:

• Neural models infer pose changes directly from frame sequences.
• Often fused with geometric constraints for stability.

How It Works

Step 1:

• Estimate frame-to-frame correspondences and solve relative camera transform.

Step 2:

• Integrate transforms over time to build trajectory and optionally refine with local optimization.

Visual odometry is the real-time motion estimation engine that keeps an agent oriented as it moves through unknown space - robust VO is a prerequisite for reliable autonomous navigation.