Publicação
6D Pose Tracking of Objects with Very Partial View
| Resumo: | This dissertation addresses the problem of estimating the 6D pose of knee arthroscopic surgical instruments using only RGB video. Arthroscopy is a minimally invasive procedure where the instruments appear with strong symmetries, specular reflections, and frequent occlusions, making pose estimation highly challenging. The work integrates a markerless pipeline that combines a object detector, a dense correspondence method, and a geometric pose solver, with distortion-aware processing throughout the pipeline to handle the radial distortion of arthroscopic cameras. The pipeline was trained and tested on datasets of three instruments, with evaluation based on standard pose metrics and clinically relevant measures such as tip position and axis orientation errors. The results show reliable detection performance in most test sets, and competitive pose accuracy across instruments. Mean tip error remains within millimetric precision, and mean axis orientation error is generally below 5°. Despite these promising results, runtime analysis indicates that the current throughput of about 7.5 FPS is below the requirements for real-time surgical navigation. Overall, the study demonstrates that deep learning methods are paving the way to provide accurate and robust pose estimation, competing with physical markers approaches under occlusion and symmetry conditions, supporting their potential for integration into next-generation navigation systems in knee arthroscopy. Future improvements in dataset quality, processing speed, and method selection are expected to further enhance the practical deployment of this approach. |
|---|---|
| Autores principais: | Silva, João Filipe Cardoso |
| Assunto: | 6D pose estimation arthroscopy surgical navigation computer vision deep learning estimação de pose 6D artroscopia navegação cirúrgica visão por computador aprendizagem progunda |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso embargado |
| Instituição associada: | Universidade de Coimbra |
| Idioma: | inglês |
| Origem: | Estudo Geral - Universidade de Coimbra |
| Resumo: | This dissertation addresses the problem of estimating the 6D pose of knee arthroscopic surgical instruments using only RGB video. Arthroscopy is a minimally invasive procedure where the instruments appear with strong symmetries, specular reflections, and frequent occlusions, making pose estimation highly challenging. The work integrates a markerless pipeline that combines a object detector, a dense correspondence method, and a geometric pose solver, with distortion-aware processing throughout the pipeline to handle the radial distortion of arthroscopic cameras. The pipeline was trained and tested on datasets of three instruments, with evaluation based on standard pose metrics and clinically relevant measures such as tip position and axis orientation errors. The results show reliable detection performance in most test sets, and competitive pose accuracy across instruments. Mean tip error remains within millimetric precision, and mean axis orientation error is generally below 5°. Despite these promising results, runtime analysis indicates that the current throughput of about 7.5 FPS is below the requirements for real-time surgical navigation. Overall, the study demonstrates that deep learning methods are paving the way to provide accurate and robust pose estimation, competing with physical markers approaches under occlusion and symmetry conditions, supporting their potential for integration into next-generation navigation systems in knee arthroscopy. Future improvements in dataset quality, processing speed, and method selection are expected to further enhance the practical deployment of this approach. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.