Project Description

The automated detection of different trucks in or-thophotos with ArcGIS Pro

The automated detection of different trucks in orthophotos with ArcGIS Pro

Project Goal

This project aimed to be able to recognise different trucks based on orthophotos (raster images). For this purpose, the trucks had to be classified into the following classes according to their length. Three vehicle types were distinguished: single truck, semi-trailer truck and multi-trailer truck.

Dataset used

Digital orthophotos of North Rhine-Westphalia from OpenGeodata.NRW.de were used for this project. These images were 4-band images (red, green, blue, infrared) with a resolution of 10cm and an image size of 1000m × 1000m (10000pixel × 10000pixel). The NRW orthophotos were divided into 100m × 100m images and 2400 were selected from them. Of this number, 2148 were labelled for training and validation of the truck detection model.

Challenges

Model training requires a large amount of training data to achieve optimal results. For better results and analysis, more training data and better data resolution would have been necessary. For example, the model occasionally recognised cars with trailers (and in rare cases house walls) as trucks. Furthermore, it was difficult to differentiate between two vehicles parked close to each other, as they can be perceived as one vehicle.

Applied Methods

A deep learning model was trained for automated recognition and differentiation of the different vehicle types. A pre-trained TernausNet-16 model was used for this purpose. In the case of vehicle detection and classification, the output of the model consists of pixels belonging to the classes background, trailer and cabin. The detected trucks were then divided into 3 classes based on the length of the truck. Subsequently, the trained model and truck classification was implemented in ArcGIS Pro through an inference function and a specific esri model definition file.

Project outcome

The model resulted overall with good and reliable accuracy, precision and recall. Over 99% of the trucks were recognised by the model. Less than 10% of the detected trucks were classified in the wrong class. Further fine-tuning of the model, with additional training data from different conditions, can make the model even more robust.

Category

Technologies

Python
ArcGIS Pro

Download

Social Sharing

Other Use Cases in this category

All Use Cases at a Glance

Contact

Stefanie Supper
CEO

Contact

Stefanie Supper
CEO