Deep learning approach to driving behavior prediction with data fusion

Context

Reliable driving behavior prediction is of great importance in autonomous driving, as the interaction between vehicles is complex and non-deterministic. To address this problem, multimodal data from different sources must be used in real time to ensure a reliable prediction. In the last decades, different approaches such as parametric models, Markov models, reinforcement learning (RL) and inverse RL have been developed.

Within the scope of this work, new methods for driving behavior prediction will be investigated depending on the chosen focus:

• Comprehensive literature review ("Review of SOTA")
• Development of a data fusion framework, hardware (e.g. sensors, GPS, cameras)
• Development of a data fusion framework, software (data simulation/generation, cloud data, data fusion)
• Development, training and validation of (RL-based) deep learning approaches for driving behavior prediction
• Optimization of neural networks, e.g. feature selection, network pruning & quantization, knowledge distillation
• Use and optimization of NN on the test bench

Tasks

• Literature research on the SOTA technique in the relevant field
• Conception of the work in terms of workload and publication potential
• Targeted development and implementation of the minimum viable product (MVP) for driving behavior prediction
• Evaluation and analysis of the results and writing of the final thesis

Prerequisites

• Motivation and interest in independently solving technical problems
• Knowledge of machine learning, ideally reinforcement learning and data fusion
• Experience with programming (Python, C++, Java...)
• Analytical, problem-solving and communication skills