Autonomous Mower Navigation with ORB-SLAM3
High-precision stereo vision SLAM system for autonomous mower navigation in complex outdoor environments using NVIDIA Jetson Xavier NX.
The Challenge
A high-tech robotics startup building autonomous mowers needed a visual SLAM system that could handle the unique challenges of outdoor landscaping: rapid movements, changing lighting conditions, uneven terrain, and visual occlusions from vegetation. GPS alone wasn't precise enough for the centimeter-level accuracy required.
Our Approach
We implemented a stereo vision SLAM system based on the ORB-SLAM3 library, optimized for real-time performance on embedded hardware.
Stereo Camera Calibration
Using IMX219 stereo cameras, we performed advanced calibration with OpenCV to achieve precise depth perception. Sub-pixel rectification ensured accurate disparity maps critical for outdoor depth estimation.
Real-Time SLAM on Edge
The entire pipeline runs on an NVIDIA Jetson Xavier NX, processing stereo frames in real-time. ORB-SLAM3's visual-inertial odometry provides robust tracking even during rapid turns and temporary feature loss.
Loop Closure & Relocalization
The system incorporates loop closure detection to correct accumulated drift during long mowing sessions. Relocalization mechanisms allow the mower to recover its position after temporary tracking failures caused by occlusions.
Technical Stack
- Hardware: NVIDIA Jetson Xavier NX, IMX219 stereo cameras
- SLAM: ORB-SLAM3
- Libraries: OpenCV, Boost, Qt
- Languages: C++
- OS: Linux (JetPack)
Results
- Centimeter-level localization accuracy in outdoor environments
- Robust tracking through lighting changes and occlusions
- Real-time performance on embedded hardware
- Efficient path planning with obstacle avoidance