Bart Aalbers

Machine Learning Engineer

I design, develop and optimize machine learning solutions for real-world applications, with a strong focus on computer vision and deep learning. My background in mathematics and engineering allows me to translate complex technical challenges into working, scalable solutions.

Areas of expertise

• Computer visionObject detection, image segmentation, image classification and multi-object tracking.
• Deep learningDesign of neural network architectures, loss functions and training strategies.
• Video analyticsReal-time video processing, camera calibration and image-to-world coordinate transformations.
• Machine learningClassification, regression, feature engineering and model evaluation.
• Model optimizationImproving accuracy, computational efficiency and inference speed.
• Production-ready pipelinesWorking workflows for data preparation, training, validation and evaluation, delivered to your team.

Technologies

• Python
• PyTorch
• TensorFlow & Keras
• TensorRT
• ONNX / ONNX Runtime
• OpenCV
• Scikit Learn
• XGBoost & LightGBM
• Git

I work closely with clients and multidisciplinary teams and take ownership of the entire process, from problem analysis to a production-ready pipeline. My focus is on bridging the gap between machine learning research and real-world applications, through solid engineering and a deep understanding of the underlying algorithms.

Football Analytics - Real-Time 2D Radar

For a client, we are developing, with a small team, a computer vision system that converts live football broadcasts into a real-time 2D radar view. The positions of all players, goalkeepers, referees and the ball are projected onto a top-down view of the pitch. This produces a complete, up-to-date picture of the flow of the game, based solely on the camera footage of the live broadcast.

The system consists of several modular components that run in parallel, each performing a specific task. First, the camera viewpoint is classified in order to select suitable footage. Next, players, goalkeepers, referees and the ball are detected and tracked across consecutive frames using multi-object tracking. Through jersey number recognition and analysis of kit colors, players are identified and assigned to the correct team.

To project the positions accurately onto the pitch, the camera geometry is continuously calibrated based on the visible pitch lines. Using a homography, the pixel coordinates are then transformed into coordinates in meters on the football pitch. Finally, an agent-based system reconstructs the complete game situation. It uses not only the observed positions, but also smooths movement patterns and estimates the positions of players who are outside the camera view.

My work on this project consists of designing the system architecture and developing large parts of the pipeline end to end. I am also involved in designing model architectures and in training and evaluating the associated deep learning models. A key technical challenge is synchronizing the various parallel components at frame level while at the same time meeting the real-time processing requirements of live video. In particular, consistently tracking and identifying players in crowded situations, such as set pieces where players overlap, requires robust tracking and re-identification methods.

Mowhawk - Litter Detection

Litter is a major problem, both worldwide and in the Netherlands. This is one of the reasons why Datacadabra developed the Mowhawk following a tender by the Province of Overijssel. The Mowhawk is a camera system equipped with AI that is mounted on the mowers that cut the verges along roads. Within the collected footage, the system detects in real time whether litter or invasive exotic species are present, and the route of the mower is recorded. This information is then stored in a cloud environment and exported to map layers that can be loaded into the client's Geographic Information System (GIS).

In this development project, I worked on, among other things, the development of the litter detection system. Deep learning was used for this, with a pipeline set up for training the model. In preparation, a model architecture was designed and a suitable loss function was defined. Within this pipeline, the data is prepared and split into a training and validation set, the model is compiled and trained, and the final results are evaluated.

Surface Segmentation

Mobile mapping is an advanced technology used to collect data about the environment while a vehicle moves through an area. The vehicle is equipped with advanced sensors, such as LiDAR (Light Detection and Ranging), cameras and GPS, which capture 3D data and images of the surrounding world. As part of a team, I worked on a project in which a pipeline was developed to detect road surfaces in camera images. These detections are then converted into 2D geographic coordinates.

Within this project, my focus was mainly on the part of the pipeline that processes images and segments surfaces. I designed and elaborated a model architecture and developed a custom loss function. These elements were integrated into a pipeline for training the model, in which images are prepared and split into a training and validation set. The model is compiled, trained and evaluated.

In addition, I set up a pipeline for processing new images with the trained model, and this pipeline was integrated into the graphical user interface (GUI) to make it user-friendly. The end result is an effective and efficient method for automatically detecting and locating surfaces in mobile mapping images, which provides valuable information for various applications, such as urban planning, infrastructure development and traffic analysis.

Mowhawk - Invasive Species Detection

Because of globalization and climate change, the Netherlands has a growing problem with the spread of invasive exotic plants and animals. Besides displacing native species, some of these invasive species cause damage to infrastructure, roads and waterways. To tackle this challenge, the Mowhawk was developed as an advanced camera system, equipped with artificial intelligence, that is installed on the mowers that maintain the verges along roads.

A key goal of the Mowhawk is to detect, in real time, the presence of invasive plant species, such as Japanese knotweed and/or giant hogweed, within the collected footage. This valuable information is then stored in a cloud environment and exported to map layers that can be integrated into the client's Geographic Information System (GIS).

Within this development project, I worked on the development of the detection system for invasive species. For this task, I used deep learning techniques, with which we set up a solid pipeline for training the model. I was responsible for designing the model architecture and defining a suitable loss function to optimize the learning process. To prepare the model, we carefully preprocessed the relevant data and then split it into a training and validation set. The model was compiled and trained, and finally we evaluated the end results extensively to ensure that the system is accurate and reliable in its detection capabilities.