Neural Architecture Search for Bird-Eye-View Perception Models for Automotive Applications

Context

Neural Architecture Search (NAS) has been a powerful tool for automating the design of DL models by optimizing the architectures of the NNs achieving better efficiency and performance. Meanwhile, Bird's-Eye-View (BEV) perception models [1] have become essential in automotive applications, enabling robust scene understanding for autonomous driving and Advanced Driver-Assistance Systems (ADAS). This thesis will utilize NAS to BEV perception models aiming to find architectures that can be used in embedded systems.

Targets

• In this work state-of-the-art (SOTA) in NAS and BEV approaches will be researched.

• Based on these results a NAS framework tailored for searching optimized BEV architectures will be designed and implemented.

• The performance of discovered models will be evaluated using datasets for autonomous driving on the embedded GPU platforms and compared to SOTA models.

Requirements

• Experience with Deep Learning (DL) and one of the DL Frameworks like PyTorch or Keras/Tensorflow

• Experience with Python and Computer Vision