Outdoor Characterization and Geometry-Aware Error Modelling of an RGB-D Stereo Camera for Safety-Related Obstacle...

Area of Science:

• Robotics and Automation
• Computer Vision
• Agricultural Engineering

Background:

• Stereo cameras (depth or RGB-D cameras) are vital for obstacle detection and navigation in machinery.
• Their outdoor performance, especially at medium to long ranges under varying light, is not well-documented.
• This is critical for agricultural machinery to detect workers and prevent collisions.

Purpose of the Study:

• To quantify the outdoor performance of stereo cameras in realistic farm settings.
• To develop a model for predicting and correcting depth errors.
• To benchmark the Intel RealSense D455 for agricultural safety applications.

Main Methods:

• A field protocol was established to test stereo cameras at distances from 4m to 16m.
• A 1 square meter planar target was used in outdoor environments with diverse lighting.
• Tests included varying the target's position in the camera's field of view (FoV) and adjusting camera presets.
• Disparity surfaces were fitted to model depth bias as a function of distance, light, and FoV position.

Main Results:

• A model was developed to predict depth errors with good precision (RMSE: 0.46-0.64 m, MAE: 0.40-0.51 m).
• The model effectively corrects systematic bias based on distance, lighting, and FoV.
• Results demonstrate the potential for benchmarking sensors and improving safety systems.

Conclusions:

• The developed model provides a reliable method for assessing stereo camera performance in agricultural contexts.
• Accurate depth error prediction supports the development of safety-critical perception systems.
• This work enables replication and benchmarking for various sensors and field conditions.