Examenspresentation av Yingjie Huang och Robert Füllemann

Examiner: Lennart Svensson

Abstract:
4D radar systems attract much attention nowadays, but little research has been done to explore the abilities of deep learning methods on perception tasks in 4D radar point clouds. Therefore, this master's thesis studies semantic segmentation (SemSeg) and classification capabilities of neural networks applied to 4D radar data of static applications and compares these networks against themselves and classical techniques. First, eight SemSeg neural networks are evaluated in accuracy and speed, a topic not explored extensively before. Then the SemSeg task is separated into two subtasks: object masking and classification, both completed by PointNet. The performance of the new approach is compared against the same network directly for the SemSeg task and classical counterparts. As a continuation, a sequential classifier is also introduced to handle the sparsity inherent in 4D radar data, and specialized metrics and corresponding loss functions aimed at shifting false positives toward ground truth in SemSeg are proposed. The results reveal divergent performances between point- and convolution-based networks. It is found that splitting the SemSeg task improves overall accuracy, and sequential data use further enhances classification performance. Deep learning methods surpass the tested classical techniques in accuracy but are computationally heavier, highlighting the continued relevance of classical methods. This research provides meaningful insights into the applications of deep learning to 4D radar data, thus benefiting both academia and industry.

Welcome!
Yingjie, Robert and Lennart