Description

Problem:

For vehicle safety, ride comfort and noise control, road surface properties such as coefficient of friction, sound absorption and roughness are critical. Traditional measurement techniques are often expensive and intrusive. They are also limited in location. As more heterogeneous vehicles, sensors and environmental data become available, such as noise, heat, weather and vehicle CAN-Bus, intelligent data-driven systems that can estimate multiple road surface properties simultaneously are required. However, challenges remain in the effective integration of multiscale and multimodal data, the handling of spatio-temporal dependencies, and the generalization to different road environments.

Aim:

To develop a deep learning-based technique that uses multimodal sensor data through fusion, attention mechanisms, and transfer learning techniques to accurately estimate multiple road surface properties (e.g., friction coefficient, sound absorption, roughness).

Goal:

1. To handle temporal and spatial variations in the data by designing multi-scale neural network architecture.

2. To use late and early fusion strategies to fuse multi-modal data (e.g. noise, images, weather, vehicle dynamics).

3. To improve feature relevance and model interpretability by implementing attention mechanisms.

4. To use transfer learning to generalize to different types of roads and environments with a limited amount of labelled data.

5. To evaluate the system on multiple datasets and comparing with traditional and single modal approaches.

Required Skills and Knowledge:

• Programming skills (preferably in Phyton or C++) • Basic knowledge of data analysis, image processing and computer vision libraries (PyTorch, OpenCv, Cuda, Pandas, NumPy, ScikitLearn, TensorFlow, Keras). • To have basic knowledge about signal processing, or the willingness to learn these things. Also, basic knowledge of image and video processing and patient annotation of images.