Rajitha de Silva: Keypoint-semantic integration for improved feature matching in outdoor agricultural environments​

Bionote

Rajitha de Silva is a Postdoctoral Research Associate in Robotic Perception and Navigation at the University of Lincoln, UK. He earned his Ph.D. in Computer Science from the University of Lincoln, with a thesis on vision-based autonomous navigation for agricultural robots. He holds a First Class Honours BSc in Electrical and Electronic Engineering from the Sri Lanka Institute of Information Technology (SLIIT). Before moving to the UK, he worked as an Assistant Lecturer at SLIIT. His research interests include robotic perception, AI in agri-tech, and deploying robotics in real-world environments.

Reliable navigation for outdoor robots depends on perception systems that can effectively manage visual complexities like repetitive patterns and changing scene appearances. Visual feature matching plays a vital role in vision-based systems but becomes especially difficult in natural outdoor environments due to perceptual aliasing. This problem is pronounced in vineyards, where repetitive vine trunks and similar natural features produce ambiguous visual descriptors, leading to unreliable matching. We propose that incorporating semantic context at keypoint locations can reduce perceptual aliasing by making descriptors more distinctive. To support this idea, we present a method for integrating semantic cues into keypoint descriptors, refining them within semantically important areas of the image. This enhances the ability to distinguish between visually similar features. We demonstrate the utility of our approach on two key vineyard tasks: relative pose estimation and visual localization. Our results show consistent improvements in matching accuracy across various keypoint types, with robustness maintained over several months in challenging vineyard environments.

Co-Authors: Jonathan Cox, Marija Popovic, Cesar Cadena, Cyrill Stachniss, Riccardo Polvara