A Classical Vision 3-D Registration Methodology for a Successful Pose Estimation of Semi-deformable Objects

This paper presents a robust computer vision method for detecting and estimating the 3D pose of semi- deformable objects—a key challenge in robotics due to occlusions, viewpoint changes, and non-standard shapes. We propose an 11-step pipeline using classic computer vision techniques to estimate the pose of a bell pepper from multiple angles, enabling potential applications like robotic harvesting. The method was developed on Windows, ported to Ubuntu, and deployed on a Jetson Nano 2GB. It was also integrated as a ROS2 node and tested in Gazebo simulations. The system showed consistent performance across platforms and sensors, including Azure Kinect and a simulated Intel RealSense D435i. The framework is adaptable to various vision-based robotics tasks and is available in this repository.