Urban Automation Lab

Having grown from the Smart City and privacy projects Connected Urban Twins (CUT), DiGuRal and DE4L, our research focuses on the analysis of geospatial and visual data, such as point clouds, aerial and street-view imagery, as well as 3D city models, and its application for urban planning and smart city solutions.

We combine expertise in Machine Learning with GIS data analysis to address complex tasks in urban environments and to create automated workflows to derive actionable knowledge for municipal databases while adhering to defined legal regulations.

In close collaboration with municipal administrations and external partners, we address current challenges in city planning, such as building efficiency, solar or greenery potential, and fair distribution and safety in the road space.

To implement these applications, we handle full software development, ranging from training custom models and implementing backend logic to frontend visualizations and interfaces as well as mobile apps.

For more updates, visit us on LinkedIn.

Glimpse Into Our Research

Road Space Data Collection & Visualization

The focus of the DiGuRal project was the development of a mobile data collection app and visualization platform for road space data. Building on the project's concept, our team has redeveloped the app and the visualization platform. Captured images undergo several anonymisation steps to obscure individuals and number plates in accordance with the GDPR (DSGVO). The images can then be used for various analyses of the road environment, such as assessing its condition. We visualise the results using open-source software such as Apache Superset. We are looking for local municipalities open to test this, so please do get in touch with us.

Building Analysis

The group also focuses on analysing various aspects of buildings. This involves identifying missing information on characteristics such as the number of floors or building type and age. We further conduct in-depth analyses of the potential for solar energy and vertical greenery, with a aim to address the current challenges posed by climate change.

Parking Space

In many major cities, the road space is shared by pedestrians, cyclists, public transport and private cars. As the number of private cars increases, stationary traffic is becoming a problem in an already limited road space. To enable cities to develop solutions and strategies for this problem, we have estimated the current parking capacities city wide. This makes it possible to identify problem areas and the use of parking spaces by long-term parkers.

3D Road Space Analysis

The use of LiDAR point clouds offers many advantages in spatial terms. Unlike images, objects can be precisely located within a space. We have capitalised on this advantage to demonstrate how LiDAR data can be used to monitor vegetation in the road space. To ensure that clearance heights are maintained, vegetation is only permitted to grow into the road space up to a certain point. Using LiDAR data analysis, we can, for example, identify areas where vegetation has grown too far into the road.

Language Models for City Data

Another focus of our group is applying Large Language Models (LLMs) to process and prepare natural language inputs from city council documents or participation platform posts. With city council documents accumulating quickly, it can be difficult to find specific information within individual documents. To address this, a chatbot has been developed that has been specifically trained on ALLRIS data and can search through documents to find information relevant to user enquiries. In future, it will also be able to contextualise these enquiries and visualise temporal relatitionships between documents.

Team

Miriam Louise Carnot

Computer Vision, 3D LiDAR point clouds, Spatial Data

Aruscha Kramm

Computer Vision, GIS data, application

Dr. Florens Rohde

Data integration, Computer Vision, Privacy

Julia Friske
(on maternity leave)

Large Language Models

Publications

Loading publications…

Finished theses

Loading theses...

Openings

Positions

No open positions at the moment.

Thesis Topics

Master

Inference optimization on Nvidia Jetson platform

Supervisor: Florens

Description:

The goal of this work is to optimize deep learning inference on NVIDIA Jetson edge computing platforms. The focus is on improving latency and throughput through model optimization and different deployment strategies (docker vs native). The project offers hands-on experience with embedded AI systems and performance benchmarking.

Master

Implementierung und Evaluation einer flächendeckenden Analyse von Solar- und Begrünungspotenzialen

Supervisor: Aruscha

Description:

Buildings offer significant potential to counteract heat islands through green facades or to harness solar energy with PV systems. Which approach makes sense where depends largely on shading from neighboring buildings and trees. Existing solar cadasters primarily map roofs by slope and orientation (see [3]); however, a radiation-based suitability assessment for green facades is largely lacking. The goal of this thesis is to develop and apply a Python-based method that calculates the time-resolved shading of roof and facade areas for the entire city of Leipzig and derives from this a comparable, use-specific suitability assessment for photovoltaics and green facades. The thesis may be written in English or German.

More details

Tasks:

Data preparation: LoD model, building cadastre (see [2–3]), extract sub-areas
Calculate the shading model (separately for a summer day and a winter day)
Calculate available roof area using images of the roof surface (includes training a computer vision model to recognize roof elements (windows, chimneys, etc.)
Determine an appropriate suitability assessment
Validation: PV potential can be compared with the solar cadastre ([3]); no such comparison exists for green roofs

Data sets:

Images of roof surfaces + facades
[1] https://hub.arcgis.com/datasets/2448f6a44d4c495bb772b0b7a39865e9_0/about
[2] https://opendata.leipzig.de/dataset/3d-stadtmodell
[3] https://www.leipzig.de/bauen-und-wohnen/bauen/geodaten-und-karten/3d-stadtmodell-energie-umwelt-klima

Profile:

Degree in computer science or a related field, Python, experience training CV models; experience working with geodata is a plus.