radar_detection_pointnet.md

2D Car Detection in Radar Data with PointNets

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.