July 2019
tl;dr: Use F-pointnet for car detection with sparse 4D radar data (x, y, $\tilde {v}_r$, $\sigma$).
Overall impression
From U of Ulm. Only one target per per car, in a controlled environment. A high precision GPS is used to create the dataset GT.
This is an extension to the radar point cloud segmentation.
Key ideas
- Three steps:
- Patch Proposal around each point –> this proposal is quite like point rcnn.
- Classify patch
- Segment patch (point cloud segmentation)
- Bbox estimation
Technical details
- Radar data often contain reflections of object parts not directly visible, like the wheel house (fender) on the opposite side.
- No accumulation of data across frames like the radar point segmentation work.