Final demo of the Project Course in Machine Vision

You are warmly invited to attend the final demo session to check it out and support our students.

During the session, the students will present the projects they have been working on throughout the course (in groups of 8–10), possibly have some live/recorded demos of their solutions, and will be available to answer questions and have a discussion.

Projects

Pointcloud reconstruction from 360 imagery for forestry mapping (collaboration with SLU)

This project explores whether low-cost 360° camera imagery and open-source reconstruction workflows can complement or replace traditional forestry measurement methods based on terrestrial laser scanning (TLS) for capturing forest geometry. 

The project focuses on generating point clouds from 360° video data using the photogrammetry tool, COLMAP and comparing the resulting geometry with TLS-based reference data. Practical challenges related to forest data acquisition are investigated, including image overlap, vegetation density, GPS integration, reconstruction quality. Forestry-related variables such as tree position and diameter at breast height (DBH) are compared against TLS-derived results and field measurements using CloudCompare and 3DFin. For an interactive scene representation of the filmed plots the project also explores Gaussian Splatting. 

From the videos captured with Insta360 X4 camera of a forest plot, pointclouds were created using COLMAP, from which forest variables were measured in CloudCompare with 3DFin, and the plots are interactively viewable from Gaussian Splatting using Brush.

Monocular depth estimation for marine applications (collaborationwith Knightech Group)

Monocular range estimation is the peocess of estimating range, only using a two-dimensional image. This requires the use of trained models to detect an object in an image and then estimate the range to it. The goal of this project is to research what is needed for an implementation of monocular range estimation in the real world on boats by creating a small scale implementation.

The proposed solution is to have to separate models, one based on YOLO26 for image detection, and a model based on DepthAnythingV2 for range estimation. These models have been trained on datsets containing two types of objects. The team has created two solutions, one based on the Basler Blaze 101 ToF camera, and one using the ToF sensor and camera sensor on an iPhone. Through these combined solutions we achieved many answers regarding the future developments of monocular range estimation in real world implementations.

Quality assurance of a seeding machine using RGB and IR cameras (collaboration with Killamån plantskola)

This project, carried out in collaboration with Kilåmon Plantskola, develops a machine-vision system for evaluating seed planting accuracy in seedling trays within a production-line environment. In the current process, incorrect seed placement, such as empty cells or cells containing multiple seeds, only becomes visible after germination, resulting in delayed feedback and potential losses. The proposed system detects and counts seeds in each tray cell immediately after planting, providing earlier quality-control information and supporting the identification of faults in the planting process. A central challenge is that trays move continuously on a conveyor belt, requiring image capture, tray localisation, pre-processing, seed detection, and result visualisation to be performed without interrupting the production flow. The project therefore addresses real constraints, including structured tray layouts, small seed size, variable soil backgrounds, changing tray formats, motion-related imaging challenges, and the need for a practical operator interface.

The proposed solution integrates image acquisition, tray preprocessing, cropping, grid detection, YOLO-based seed detection, and cell-level counting within a lightweight web-based user interface designed for live visualisation and reporting. The interface presents the live camera feed, displays the detected tray as a colour-coded grid, maintains a history of previously processed trays, supports seed-type selection, and generates PDF reports with summary statistics over user-defined time periods. A classical colour- and blob-based machine-vision method was additionally investigated as a deterministic baseline for birch seeds; however, its sensitivity to illumination, substrate appearance, and seed overlap prompted the adoption of a learning-based detector. The YOLO based approach is better suited to such setting, as it can learn visual features from annotated tray images and be retrained or fine-tuned for different seed types and imaging conditions. The project further explores improved imaging setups, including controlled lighting and infrared illumination, as prospective directions for enhancing robustness.