Overview contents
Research question
How can images and video be turned into reliable information useful to a real process, even when lighting, viewpoint and capture conditions change?
Technical context
Computer vision includes methods for processing and interpreting images. Depending on the task, a system can classify content, locate objects, distinguish regions of an image or track elements over time.
An application pipeline combines capture, data preparation, models and output validation. The starting point is the problem to observe: which elements to recognise, which errors are acceptable and how to use the result within the process.
What we are studying
- 01
Capture, annotation and quality of datasets representative of the real setting.
- 02
Object recognition and localisation, segmentation and reading visual information.
- 03
Video sequence analysis and continuity of recognition across frames.
- 04
Integration with operational software, data management and output supervision.
Applications to explore
- Visual inspection of products, components and environments.
- Image cataloguing and search using visual features.
- Extracting text and information from photographed documents.
- Contextual recognition for augmented reality, visual localisation and automation.
How it can combine with other technologies
Computer vision can provide the recognition layer of a Visual Positioning System or add information to a 3D reconstruction. Results can also feed archives, management systems and AI assistants, preserving the link to the original images and checks performed.
What a prototype needs to establish
- V01
Accuracy, errors and cases where the system must request human verification.
- V02
Robustness across different lighting, backgrounds, perspectives and devices.
- V03
Annotation quality and separation between development and validation data.
- V04
Latency, resource consumption and usefulness in the chosen operational workflow.
Known limitations and risks
Occlusion, blurred images and conditions that differ from the development data can reduce reliability.
Recognising an element does not guarantee its identity or meaning within a process. Appropriate data, error handling and explicit verification criteria are needed.
Technical references