Task3D object detection, 3D pose estimation
IndustryAgriculture
DeviceTexas Instruments TDA4VM
Context and Challenges
As agricultural machinery technology evolves, perception enables machinery like tractors to operate more autonomously and to increase efficiency. Specifically, in this project, the perception goal was to localize trailers from a sensor mounted on a tractor.
In developing a robust perception solution, several critical challenges emerged:
- Robustness: ensuring resistance to mechanical stress, extreme temperatures, unfavorable lightning and dust.
- Precision: localization had to be accurate, metric and low-latency due to the moving nature of the scene.
- Environmental unpredictability: handling unpredictable variables like different trailer shapes or different driving practice from all around the world.
Solution
To address these challenges, our solution leveraged advanced technologies and methodologies:
- 3D Sensing: employed a robust 3D sensor offering enhanced resilience to visual variability compared to traditional cameras.
- Point cloud algorithms: developed specialized algorithms that process point clouds rather than pixel-based imagery, through a mix of traditional computer vision, computational geometry, robust fitting algorithms and deep learning models to accurately detect object position and orientation.
- Advanced Integration: optimized all the processing for deployment on the Texas Instruments TDA4VM embedded platform, an industrial and automotive-grade solution known for its high performance, low cost, and low power consumption.
Results
Following extensive testing conducted over multiple seasonal harvesting cycles, the system consistently demonstrated high accuracy and robustness, and moved towards productization
