Thesis Work: Automated damage classification in road crashes using deep learning

About this opportunity - Background

At Volvo Cars, our vision is that no one should be seriously injured or killed in a new Volvo. To achieve this, we need to understand real-world accidents and continually improve our methods of data collection and data analysis.

Since the 1970s, Volvo Cars has collected road traffic crashes occurring in Sweden involving at least one Volvo car. Road traffic crashes are an important asset to understanding the performance of our cars. The data collected is the basis of identifying focus areas for possible future improvements, to mitigate injuries and reduce crashes. Case coding, i.e. enriching Volvo Cars' road traffic data collection with additional information, is an essential part of the process. During case coding, the case analysts collect all relevant information about injuries of occupants and damages inflicted on the car, classify the conflict scenario according to Volvo standards, and perform several other steps to enable further analyses.

  
Currently, case analysts manually inspect images of vehicles involved in collisions and classify damage to the vehicle based on SAE standards. Specifically, Collision Deformation Classification (CDC) codes are specified for each crashed vehicle, indicating the location and extent of the damage on the vehicle, as well as the damage type and the principal direction of force applied to the vehicle. These variables can then be analyzed to relate deformation to the occurrence of specific injuries and guide the development of safety systems that can prevent such injuries. Determining CDC codes is a time-consuming process; moreover, it is not easy to achieve consistency in coding these variables across different analysts. 

  
The challenge, therefore, is to develop and implement a method for automating the CDC coding process. The topic of the planned thesis is the development and evaluation of a method, presumably based on machine learning models such as neural networks, that can take an image of a crashed car as input and return the CDC codes as output. In the development of the model, data from manually coded crashes as well as test data with precise deformation measurements can be used for training and evaluating the model against a human baseline. The ground truth dataset needs to be developed from the existing crash test parameters, and measurements may need to be made on the existing crashed objects in the laboratory. Having CDC codes as output can be used for categorizing the crashes by damaged side, as well as matching crashes to test cases with similar deformation locations.

Scope of the thesis work

The thesis will start with a literature review to understand the general principles for image analysis in an automotive context and explore available models. In parallel with this activity, an inventory of available data sources will be prepared. Next, an analysis plan is outlined with an indication of at least the following aspects, with an appropriate timeline:

•    Data to be used for model development and evaluation;
•    Selection of models to consider and an indication of the model architecture;
•    Preprocessing steps for the data, including annotation and potential data augmentation;
•    Data splitting with appropriate motivation (e.g., how is data leakage from the training set to the test set avoided?);
•    Evaluation of the model against a human baseline;
•    Sensitivity analysis;
•    Implementation of the model for test case matching (e.g., visualization of test cases together with real-world crashes in a low-dimensional space) and assessing its reliability.
    

The analysis plan is then followed up in weekly meetings with the supervisors. The thesis workers will also interact with other members of the Data Analytics & Artificial Intelligence group to obtain a better understanding of the underlying vehicle safety concepts, coding challenges, and potential uses of the coded data. These interactions will expose the students to a real-world working environment and ensure that the model selection and implementation deliver maximum utility and support safety-related work at Volvo Cars.   
  

Research questions that the thesis work should investigate and answer:  
  
A.    What types of models and architectures can be used for the automation of Collision Deformation Classification coding?  
B.    What input data are available for automated CDC coding, and which data preparation steps should be used (preprocessing of images, data splitting methods, data augmentation) to improve performance and generalizability of the model?   
C.    How does the performance of the developed model relate to a human CDC coding baseline and the ground truth deduced from test lab measurements? How sensitive is the method to different angles, light conditions, etc.?
D.    What methods are suitable for visualizing the real-world CDC classification together with the ground truth in a low-dimensional space?   

What you'll bring

The thesis work is performed in the Data Analytics & Artificial Intelligence group at the Safety Centre at Volvo Cars in Gothenburg. The project is appropriate for a pair of students with strong analytical and programming skills (preferably in Python), a good understanding of probability theory/statistics, and preferably experience with image analysis, machine learning / deep learning, or computer vision. Familiarity with vehicle safety concepts and data is a benefit. Appropriate university programs may include Computer Science, Data Science, Mathematics, and/or similar engineering programs.  

Duration

•    The work will start in January 2026
•    30 academic credits in agreement with your Thesis Advisor at the University
•    This thesis is to be conducted by 2 Students working in pair