EPSRC Centre for Doctoral Training in Agri-Food Robotics: AgriFoRwArdS Logo. EPSRC Centre for Doctoral Training in Agri-Food Robotics: AgriFoRwArdS site title logo image.

EPSRC Centre for Doctoral Training in Agri-Food Robotics: AgriFoRwArdS - Calvin John

Calvin John

  • University of East Anglia in collaboration with CHC Tech Limited

Research Interests

Computer vision, artificial intelligence, robotics, machine learning, agricultural automation

Presentations

“Serving a Full English Breakfast” (oral) – AgriFoRwArdS CDT Summer School 2023 [March 2023] – Lincoln, UK.

“Robotic Waypoint Coverage Under Soft Parallel Path Constraints” (poster) – Towards Autonomous Robotic Systems (TAROS) 2023 / AgriFoRwArdS CDT Annual Conference 2023 / Joint Robotics CDT Annual Conference 2023 [September 2023] – Cambridge, UK.

“A GNSS/RTK Localization-based Agri-Robot Digital Twin for Precision Agri-tasks in Crop Fields” (oral) – CHC Tech industry visit [October 2023] – Norwich, UK.

“Towards Resilient Agricultural PNT: An Improved Open Keyframe Visual Inertial SLAM Sensor Fusion Model for a GNSS/RTK Localization-Based Agri-Robot in Shaded Environments” (oral) – AgriFoRwArdS CDT Quarterly PhD Research Progress Meeting [January 2024] – Online.

“A visual Geo-location Method For Improved GNSS Accuracy in Agricultural Robotic Applications” (oral) – University of East Anglia Summer School 2024 [June 2024] – Norwich,

“Visual Geo-localisation in Agriculture” (poster) – REAP 2024 [November 2024] – Newmarket, UK.

Other activities

  • Winner of the ‘Best Presentation’ award at the AgriFoRwArdS CDT Summer School 2023 [March 2023].
  • Winner of the ‘Best Teamwork’ award at the AgriFoRwArdS CDT Summer School 2023 [March 2023].
  • Member of the AgriFoRwArdS CDT Annual Conference 2024 Programme Committee [May to July 2024].

About me

My name is Calvin, I am originally from Manchester. I am interested in researching the use of Robotics and AI in agriculture in order to create more environmentally sustainable farming practices.  I joined the Agri-forwards CDT in October 2022. Before joining the CDT, I studied an MSc in Computer Science with Data Analytics at the University of York and a BA in Philosophy, Politics, and Economics from the University of Essex. I am excited to be joining the University of East Anglia for my Ph.D. project in Agricultural Automation. I chose UEA because of its world-leading research in computer vision and its strong industrial links. 

MSc Project

Comparative Study of Graph-Based Methods for Coverage Path Planning: An Analysis of the Minimum Spanning Tree vs. the Nearest Neighbours Algorithm

Coverage path planning (CPP) is a crucial field of research in autonomous mobile robotics. The fundamental objective of CPP is to define a path that allows an autonomous agent, such as a robot, to pass over all points in a defined space while considering a specific set of constraints. This is of paramount importance in a multitude of applications including automated floor cleaning, aerial photography, agricultural monitoring, mine sweeping, search and rescue, and many more. Effective CPP significantly reduces the cost of operations in these fields by optimizing the robot’s movement, thus minimizing the time taken and energy expended for task completion.

This project seeks to contribute to the field by undertaking a comparative analysis of two graph-based methods for CPP; the Minimum Spanning Tree (MST) and the Nearest Neighbours (NN) algorithm. Both algorithms have been widely utilized in CPP but have yet to be compared under a uniquely specified set of constraints.

The hard constraints imposed in this study prevent diagonal movement between nodes, therefore obliging the robot to move in four directions: up, down, left, and right. This condition closely simulates real-world scenarios where certain movement patterns may not be viable due to physical restrictions or safety considerations. The soft constraints, on the other hand, specify that the robot must strive to travel in parallel horizontal lines. While it is not mandatory, these constraints reflect operational efficiency in many real-world applications like agricultural field monitoring or automated vacuum cleaning where following a systematic, parallel line pattern optimizes coverage and conserves energy.

The effectiveness of both the Minimum Spanning Tree (MST) and Nearest Neighbours (NN) algorithms will be gauged using two specific performance metrics: computation time and the frequency of node revisitation during path traversal. By comparing MST and NN in this context, the study aims to provide insights and potential recommendations for researchers and practitioners in the field, ultimately driving the advancement of efficient, cost-saving autonomous robotics applications.

PhD Project

Geometric Deep Learning-Enhanced Visual–Inertial SLAM with Geospatial Methods for Robust Agricultural Localization in GNSS-Degraded Environments

Global Navigation Satellite System Real-Time Kinematic (GNSS RTK) systems are widely used to localize agricultural robots for precision farming tasks. However, these systems often fail in environments where signal obstruction from materials and structures causes meter-level localization errors. While Simultaneous Localization and Mapping (SLAM) approaches, such as LiDAR SLAM and visual SLAM, can provide alternatives in certain scenarios, they perform best in structured indoor environments with abundant distinguishable features. In contrast, large open fields present significant challenges due to sparse unique structures, resulting in poor feature detection and difficulties with loop closure.
To address these limitations, this project develops a visual–inertial SLAM framework that incorporates specialized feature detection methods to enhance feature matching during mapping. In regions with insufficient features, geospatial statistical techniques are employed to infer 3D points, improving map consistency. Importantly, most conventional SLAM methods assume Euclidean surfaces, yet many agricultural environments are hilly or irregular, requiring higher-dimensional, non-Euclidean analysis. To investigate this gap, this research experimentally explores the use of geometric deep learning techniques to model data on non-Euclidean surfaces, improving performance in such environments. Furthermore, although RANSAC is employed to detect outliers during image matching, deep learning methods are also used to remove dynamic objects (e.g., animals) from the frame, further enhancing feature correspondence. In addition, generative AI is applied to mitigate seasonal effects in imagery, enabling more consistent feature matching and robust pose estimation across different times of year. The refined point cloud estimation is then used to recover real-world GNSS RTK coordinates, achieving centimeter-level accuracy in environments where conventional methods struggle.

Calvin’s PhD project is being carried out in collaboration with CHC Tech Ltd and under the primary supervision of Dr YingLiang Ma.