This project explores a Shifted-Window Transformer modification for CenterNet/CenterTrack-type Object Detection and Tracking from FMCW radar point-clouds for autonomous driving.
The model is trained and tested on the nuScenes dataset.
📄 Full Thesis: 360° Perception with a Network of Radars (PDF)
-
Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
https://arxiv.org/abs/2103.14030 -
Objects as Points (CenterNet) https://arxiv.org/abs/1904.07850
-
Tracking Objects as Points (CenterTrack) https://arxiv.org/abs/2004.01177
-
nuScenes: a multimodal dataset for autonomous driving
https://arxiv.org/abs/1903.11027