Lesion Change Analysis (iToBoS Internship)

A pipeline for longitudinal analysis of dermoscopic lesions combining vision-language models with custom segmentation and temporal feature extraction. The approach borrows concepts from biomedical signal processing to quantify lesion evolution across timepoints. Multiple models were explored, including LLaVa-Med, GPT-based reasoning, MedGemma feature extraction, and the iToBoS change-detection framework.

• Vision-language models applied to lesion change assessment
• Custom segmentation and temporal feature extraction for longitudinal analysis
• Multi-model integration with LLaVa-Med, GPT-4o-mini, and MedGemma
• Dermatologist collaboration for verification and annotation of results
• Results refined in collaboration with dermatologists, ensuring clinical consistency

iToBoS 2024 – Skin Lesion Detection with 3D-TBP

Kaggle International Competition — 1st place among 16 teams (0.6731 mAP, IoU 0.5–0.75). Built a two-stage pipeline for detecting multiple skin lesions from clinical images produced by 3D total-body photography. The system combined YOLOv8 for bounding-box localization with MedViT for lesion classification, balancing accuracy, speed, and robustness to imaging noise.

• YOLOv8 used for lesion detection, leveraging decoupled head for improved localization
• MedViT integrated for classification, tuned for medical image features
• Pipeline optimized for speed, detection accuracy, and robustness to real-world artifacts
• Full cycle from training, validation, and post-processing to leaderboard submission
• Solution approach presented at the competition workshop

Integrated Mobile Manipulation with Exploration and Task Execution

Built an integrated pipeline that unified exploration, planning, perception, and manipulation on the Turtlebot–SwiftPro platform. Exploration was driven by Next Best View selection in OctoMap, with RRT–Dubins motion planning ensuring feasible navigation for the non-holonomic base. Manipulation was handled through a recursive task-priority controller on the SwiftPro arm, allowing pick-and-place tasks to run in parallel with navigation.

• Next Best View exploration integrated with RRT–Dubins trajectory planning
• Task-priority control enabled simultaneous navigation and arm manipulation
• ArUco detection triggered manipulation tasks from visual perception
• Behavior trees provided modular coordination of system tasks
• Validated on the real Turtlebot–SwiftPro platform with ROS

Sampling-Based Motion Planning with Dubins Paths

Developed a planning framework for a Turtlebot platform under non-holonomic constraints. Implemented an RRT-like planner that grows the tree in continuous space while enforcing forward motion and bounded turning radius. Dubins path primitives ensured trajectory feasibility by connecting sampled nodes with smooth curves instead of straight lines. The planner was integrated with ROS, consuming occupancy grid maps and generating feasible paths to navigation goals.

• Tree expansion using Dubins-compliant propagation in (x, y, ψ)
• Collision checking against occupancy grid for safe navigation
• Path reconstruction with concatenated Dubins maneuvers
• Online execution in ROS with continuous feedback from odometry

Task-Priority Based Kinematic Control for a Mobile Manipulator

Implemented a recursive task-priority controller for a Turtlebot base with a 4-DOF Swift Pro arm in ROS 1. The solver projected lower-priority commands into the null space of higher-priority ones, handling equality and inequality tasks. Behavior trees structured pick-and-place pipelines, integrating perception, navigation, and manipulation into a single flow. ArUco detection enabled online target acquisition, while joint limit enforcement and base suppression ensured safe actuation.

• Recursive redundancy resolution with weighted damped least squares
• Joint limit and Cartesian tracking tasks enforced in strict priority
• Behavior tree execution of pick-and-place routines
• Vision-driven adaptation with ArUco-based goals

Pose-Based EKF SLAM using ICP Scan-Matching on a Kobuki Turtlebot

Designed a localization pipeline on the Kobuki Turtlebot using a pose-based Extended Kalman Filter. The system fused encoder odometry, IMU yaw, and point clouds from a Realsense depth camera converted into 2D slices. Dead reckoning predicted motion through differential drive kinematics, while ICP alignment corrected accumulated drift by matching consecutive scans. IMU yaw was integrated as a pseudo-measurement to stabilize heading and guide ICP initialization. A Mahalanobis-based filter managed the state vector, keeping only statistically significant poses.

• Pose-based EKF formulation tracking a chain of robot poses
• Dead reckoning prior corrected by IMU yaw fusion
• ICP scan-matching as the main observation for trajectory refinement
• Mahalanobis distance test to suppress redundant updates

Weakly Supervised Seafloor Segmentation

Segmented side-scan sonar into sand, mud, maerl, and rock with weak labels. Trained with noisy masks, refined pseudo-labels using CRF, and improved class balance with Lovász-Softmax loss. The GIF cycles through qualitative samples and training curves.

• Noise-tolerant training with warm-up and attention dropout
• Pseudo-mask refinement using dense CRF post-processing
• mIoU progression tracked with cross-validation runs

Multipurpose Drone and Machine Vision System for Farmland Applications

TETFund 2020 NRF Project – Mechatronics Research Group, UNN. A two-year national research project on UAV-based precision agriculture. My master's thesis work focused on the design and control of the drone, from CAD modeling of the airframe and embedded systems in SolidWorks to full kinematic, dynamic, and state-space implementation in Python. I proposed a hybrid controller combining model predictive control, feedback linearization, and backstepping for trajectory tracking. The platform was co-developed for farmland mapping, crop monitoring, and selective weed management, integrating onboard vision with a targeted spraying mechanism.

• UAV body and embedded systems designed in SolidWorks
• Kinematics, dynamics, and state-space models implemented in Python
• Hybrid control (MPC + feedback linearization + backstepping) for tracking
• Machine vision for mapping, weed detection, and selective spraying

Lion Ozumba 551 Electric Vehicle

Nigeria's first electric campus shuttle, built at UNN. I led the CAD design team, developing the vehicle body and chassis models that supported fabrication and assembly.

• CAD design leadership for body and chassis modeling
• Prototyped for campus shuttle operations
• Enabled fabrication-ready assemblies and integration

Machine Learning Labs

Autonomous Systems Labs

Multiview Geometry Labs

Probabilistic Robotics Labs