Visual Odometry

Deep Dive

Visual Odometry (VO) is the process of estimating a robot’s egomotion (position and orientation) using only camera images. Unlike full SLAM, VO focuses solely on motion estimation without building a persistent map—it outputs incremental 6-DoF pose updates, similar to wheel encoders but using vision.

Note

VO answers “how much did I move?” rather than “where am I on the map?” Typical relative position error is 0.1–2% of distance traveled.

VO vs SLAM vs VIO

System	Map Building	IMU Fusion	Drift Correction
Visual Odometry	No	No	No
Visual SLAM	Yes	Optional	Yes (loop closure)
Visual-Inertial Odometry (VIO)	No	Yes	Partial (IMU helps)

Visual Odometry:     Sensors → [VO] → Pose only
Visual SLAM:         Sensors → [SLAM] → Pose + Map + Loop Closure

Camera Configurations

Monocular

Single camera setup.

Pros: Minimal hardware, low cost, lightweight Cons: Scale ambiguity (cannot determine absolute distances), requires motion for initialization

Use case: Drones where weight matters, simple robots

Caution

Monocular VO cannot determine if the robot moved 1 meter or 10 meters without external scale reference.

Stereo

Two cameras with known baseline.

Pros: Metric scale from triangulation, more robust Cons: Higher cost, calibration required, degrades to monocular at long distances

Use case: Ground robots, autonomous vehicles

RGB-D

RGB camera + depth sensor.

Pros: Direct depth measurement, simplified algorithm Cons: Limited range (~4m for structured light), affected by sunlight

Use case: Indoor robots, manipulation

Algorithmic Approaches

Feature-Based

Extract and track distinctive keypoints across frames.

Pipeline:

1. Detect keypoints (ORB, FAST, SIFT)
2. Compute descriptors
3. Match features between frames
4. Estimate motion from correspondences

Examples: ORB-SLAM (odometry component), RTAB-Map

Pros: Robust to moderate lighting changes, efficient Cons: Fails in textureless environments (blank walls)

Direct

Use raw pixel intensities instead of features.

Pipeline:

1. Select pixels with high gradient
2. Minimize photometric error between frames
3. Joint optimization of pose and depth

Examples: LSD-SLAM, DSO (Direct Sparse Odometry)

Pros: Works in low-texture environments, can produce dense depth Cons: Sensitive to lighting changes, exposure variations

Feature Extractors Comparison

Extractor	Speed	Accuracy	Notes
FAST	Fastest	Lower	No scale invariance
ORB	Fast	Good	Best speed/accuracy tradeoff
SURF	Medium	High	Scale/rotation invariant
SIFT	Slow	Highest	Most robust, computationally expensive

Processing Pipeline

┌─────────────────────────────────────────────────────────────────┐
│                    Visual Odometry Pipeline                     │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  1. IMAGE ACQUISITION                                           │
│  ┌──────────────┐   ┌──────────────┐                           │
│  │   Camera     │──►│  Undistort   │                           │
│  │   Frame(s)   │   │  + Rectify   │                           │
│  └──────────────┘   └──────────────┘                           │
│                            │                                    │
│  2. FEATURE PROCESSING     ▼                                    │
│  ┌──────────────┐   ┌──────────────┐   ┌──────────────┐        │
│  │   Feature    │──►│   Feature    │──►│   Outlier    │        │
│  │  Detection   │   │   Matching   │   │  Rejection   │        │
│  │ (ORB, FAST)  │   │  (BF/FLANN)  │   │  (RANSAC)    │        │
│  └──────────────┘   └──────────────┘   └──────────────┘        │
│                                               │                 │
│  3. MOTION ESTIMATION                         ▼                 │
│  ┌──────────────┐   ┌──────────────┐   ┌──────────────┐        │
│  │    Pose      │◄──│   Bundle     │◄──│   Essential/ │        │
│  │   Output     │   │  Adjustment  │   │  Fundamental │        │
│  │   (6-DoF)    │   │  (optional)  │   │    Matrix    │        │
│  └──────────────┘   └──────────────┘   └──────────────┘        │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Visual-Inertial Odometry (VIO)

VIO fuses camera images with IMU measurements for more robust tracking.

Coupling Approaches

Approach	Description	Examples
Loosely-coupled	Independent visual and inertial estimates, fused via Kalman filter	ROVIO
Tightly-coupled	Joint optimization over all states (state-of-the-art)	VINS-Mono, OpenVINS, MSCKF

IMU Preintegration

IMU measurements are summarized into single constraints between keyframes, reducing computational cost while maintaining accuracy. Essential for real-time operation.

┌────────────────┐
│ Stereo Camera  │────┐
└────────────────┘    │    ┌─────────────┐
                      ├───►│  Tightly-   │───► Fused Pose
┌────────────────┐    │    │  Coupled    │     (6-DoF)
│      IMU       │────┘    │  Optimizer  │
│  (200+ Hz)     │         └─────────────┘
└────────────────┘

Tip

VIO is particularly useful for drones and handheld devices where fast motion can cause motion blur, and IMU data helps bridge visual tracking gaps.

Challenges

Common Failure Modes

• Low texture: Blank walls, uniform surfaces → insufficient features
• Dynamic objects: Moving people/vehicles corrupt motion estimate
• Lighting changes: Day/night transitions, flickering lights
• Motion blur: Fast rotation or translation during exposure
• Pure rotation: Scale ambiguity in monocular setups
• Rolling shutter: Image distortion during fast motion

Isaac ROS Visual SLAM (cuVSLAM)

NVIDIA’s GPU-accelerated solution supports both VO and full SLAM modes.

Key Features

• Stereo visual-inertial odometry (SVIO)
• Multi-camera support (up to 32 cameras / 16 stereo pairs)
• Automatic IMU fallback when visual tracking fails
• Sub-1% trajectory error on KITTI benchmark

Performance (Jetson Orin AGX)

• Stereo mode: 2.7 ms per frame
• Stereo-Inertial: 30 FPS camera, 200 Hz IMU

Launch Example

# Launch cuVSLAM with RealSense camera
ros2 launch isaac_ros_visual_slam isaac_ros_visual_slam_realsense.launch.py

# Check odometry output
ros2 topic echo /visual_slam/tracking/odometry

Output topics:

• /visual_slam/tracking/odometry — nav_msgs/Odometry (6-DoF pose)
• /visual_slam/vis/observations_cloud — Feature point cloud
• /visual_slam/status — Tracking status

ROS 2 Integration

RTAB-Map Visual Odometry

# Stereo odometry
ros2 launch rtabmap_ros stereo_odometry.launch.py \
    left_image_topic:=/stereo/left/image_rect \
    right_image_topic:=/stereo/right/image_rect \
    left_camera_info_topic:=/stereo/left/camera_info \
    right_camera_info_topic:=/stereo/right/camera_info

Sensor Fusion with robot_localization

Combine VO with wheel odometry and IMU for robust state estimation:

┌────────────────┐
│ Wheel Encoders │────┐
└────────────────┘    │    ┌─────────────┐
                      ├───►│   EKF /     │───► Fused Odometry
┌────────────────┐    │    │   UKF       │
│ Visual Odometry│────┤    └─────────────┘
└────────────────┘    │
                      │
┌────────────────┐    │
│      IMU       │────┘
└────────────────┘

Tip

When fusing VO with other odometry sources, use velocity (not position) from VO to avoid double-counting position information.

Applications

Domain	Use Case	Why VO
Drones/UAVs	Autonomous flight	GPS-denied environments
Ground robots	Navigation, AGVs	Wheel slip compensation
Autonomous vehicles	Self-driving	GPS complement, tunnels
AR/VR	Head tracking	Low latency, 6-DoF
Underwater robots	Inspection	GPS unavailable

Prerequisites

Cameras Camera types and calibration fundamentals

IMU Inertial measurement for VIO fusion

Coordinate Frames Understanding spatial transforms

Related Terms

SLAM Full SLAM with mapping and loop closure

Isaac ROS GPU-accelerated perception on Jetson

Nav2 ROS 2 navigation using odometry

TF2 Transform broadcasting for robot frames

Sources

• Isaac ROS Visual SLAM — cuVSLAM documentation, multi-camera support, KITTI benchmarks
• cuVSLAM Concepts — GPU-accelerated VO/SLAM architecture
• Visual Odometry Tutorial (Scaramuzza) — Foundational VO theory and algorithms
• RTAB-Map ROS 2 — Visual odometry strategies and ROS 2 integration
• OpenVINS — Open-source VIO with active ROS 2 support
• robot_localization — EKF/UKF sensor fusion for ROS 2