In this project, we detect and segment noise insulating walls near rail ways from aerial imagery to enable a fully automated noise mapping across Germany.
View the Project on GitHub merantix-momentum/dzsf-open-source
This project is a collaboration between Merantix Momentum and The German Center for Rail Traffic Research at the Federal Railway Authority (DZSF) to enable a fully automated noise mapping process across the whole Germany. The goal of the project is to segment noise insulating wall from aerial imagery near railways. This repository contains the exported ONNX model file and documentation for our machine learning solution tailored for the detection of noise insulating walls from aerial imagery near railways. The project is funded by the DZSF and the project page can be found on their website.
Our project focuses on developing a machine learning solution to detect noise insulating walls from aerial imagery. We employ state-of-the-art techniques in deep learning to accurately identify these walls, which are crucial for various urban planning and environmental management applications.
Our model architecture is based on a Vision Transformer backbone with register tokens with a segmentation head on top. We use the check point from vit_base_patch14_dinov2.lvd142m
from the timm package for the DINOv2-trained ViT. The segmentation head is a sequence of 3x3
convolutional layers with ReLU
activations. The frozen backbone encodes the input image which is then processed by the segmentation head and upsampled to generate pixel-wise predictions. The following diagram illustrates our model architecture:
Our model was trained on a dataset consisting of Orthophotos containing noise insulating walls covering the whole Germany. The Orthophotos had a ground resolution of 20 cm per pixel
. The annotations used for fine-tuning the segmentation head came in the form of linestrings which were then joined onto the raster layers and tiled togeter into 1000x1000 px
annotated images.
To get started with using our model in QGIS, follow these steps:
model.onnx
file into the Deepness plugin.Load default parameters
option.The expected inference time for a 5000x5000 TiF file is 3-10 minutes depending on the hardware. For a more extensive user guide, please refer to user instructions.
© GeoBasis-DE / BKG (2023)
If you have any questions, suggestions, or feedback, feel free to reach out to us through our contact form.
Our project is licensed under the MIT License.