Thesis Worker - Road condition prediction by camera-based deep learning models

About this opportunity - Background

Volvo Car AB is transforming from a traditional to a software-defined, data-driven automotive brand. Safety is still and will remain the central focus for the company. The success of autonomous drive (AD) and next-generation advanced driver assistance systems (ADAS) largely depends on active safety sensors for observing the drivable road conditions to enable accurate decision-making and optimal control of braking, acceleration, and steering functionalities. One primary component of drivable environment conditions is the road surface condition that varies largely with inclement weather, ground surface texture-based irregularities, and road debris. A key parameter associated with the tire grip is the road friction coefficient, which drastically reduces when snow, ice, water, or even wet leaves are present on the road surface compared to dry conditions. Recent advances now open a new opportunity: perception sensors such as cameras and lidar, combined with data-driven AI models, could potentially enable reliable classification of road surface and weather conditions ahead of the vehicle.

Scope of the thesis work

In this thesis, the students will focus on designing a deep-learning model to predict the road surface condition ahead of the vehicle using camera images. Surface conditions might include characteristics on the road such as snow, ice, water puddles, bumps, potholes, and roughness. Students are expected to aggregate the dataset for surface conditions. Different neural networks will be studied on the given dataset from both manual and automated annotations. Performance index, incl. accuracy and uncertainty matrix, will be generated in connection with the safety requirements from potential future friction-aware AD/ADAS functions. The industrial vision-based deep learning architecture will be investigated to integrate surface condition classification outputs. If time permits, the prediction model trained in the thesis can be downloaded and run in real time on Nvidia automotive-grade hardware. 

The thesis work will include the following parts:
•    Literature review of state-of-the-art vision-based road weather classification and anomaly detection methods.
•    Literature review on camera-based supervised deep learning algorithms.
•     Extract relevant road surface condition ground-truth data from Volvo Cars testing dataset on public roads.
•    Investigate suitable deep learning algorithms, ranging from support vector machines to deep neural networks, convolutional neural networks, or vision-transformers.
•    Accuracy and uncertainty estimation of suitable learning algorithms.
•    Selection of appropriate performance metrics for evaluating and comparing proposed learning algorithms.
•    Closed-loop simulation with certain AD/ADAS vehicle motion control functions and verification of the proposed algorithms. 
•    Stretched target: Stretched target: Verification of the designed algorithm on Nvidia hardware connected to sensors in a test vehicle.

What you'll bring

MSc programs within Mechatronics, Data science/AI, Embedded systems, or related fields. Knowledge in Vehicle Dynamics, Computer Vision is a merit. Experience with Python or C++ is a must. 

Duration

•    The work will start in January 2026
•    The duration for this thesis work is 20 weeks
•    30 academic credits in agreement with your Thesis Advisor at the University
•    This thesis is to be conducted by 2 Students working in pair