6 Point Clouds from Radar and Lidar

6.1 Start With the Story

This section: Start With the Story — Radio Remi sees two maps of the same loading area.

Radio Remi sees two maps of the same loading area. Lidar draws sharp walls with many points. Radar shows fewer points, but some carry radial speed and radar cross-section. A moving trolley becomes several returns in each frame. Remi must align the sensor pose, remove implausible points, group neighbours, and track a cluster without pretending every dot is an object.

6.2 Overview

This section: Overview — A point cloud is a set of detections in a coordinate frame.

A point cloud is a set of detections in a coordinate frame. Lidar points commonly carry 3D position, intensity, and time. Radar detections may carry range, azimuth, Cartesian position, RCS, radial velocity, sensor identity, and time. Vehicle or robot pose places sensor points into a shared frame. Filtering removes only points that fail stated rules; clustering proposes objects; tracking links proposals over time with uncertainty.

6.3 Learning Objectives

This section: Learning Objectives

By the end of this chapter, you will be able to:

• Read radar and lidar points with position, intensity or RCS, velocity, timestamp, and pose fields.
• Apply coordinate transforms, filtering, clustering, and simple tracking.
• Report missed objects, false clusters, and dataset limits before making an IoT claim.

6.4 Represent every point with context

This section: Represent every point with context — Coordinates are meaningless without a frame and unit.

Coordinates are meaningless without a frame and unit. Sensor-frame points move with the device; world-frame points require a pose transform. Timestamps matter because the platform and targets can move during collection. RCS and intensity are sensor-dependent return measures, not universal material labels. Radial velocity measures motion along the radar line of sight, not full object speed.

6.5 Read the processing chain

This section: Read the processing chain — The diagram starts with radar and lidar records that keep their different fields.

The diagram starts with radar and lidar records that keep their different fields. The next card aligns time and pose, then applies bounded range, quality, and motion filters. The final card clusters neighbouring points and associates them over time. Read the uncertainty strip last: every transform, threshold, and association can create missed points, false groups, or identity swaps.

Figure Figure 6.1 keeps the stages in reading order. The diagram starts with radar and lidar records that keep their different fields.

+

In Figure 6.1, read left to right, then use the final boundary note to limit the conclusion. The arrows show processing order, not proof that the final output is correct in every setting.

6.6 Filter and cluster sparse returns

This section: Filter and cluster sparse returns — A range gate removes detections outside the application envelope.

A range gate removes detections outside the application envelope. Quality or RCS thresholds can suppress noise but may also remove weak people or surfaces. A connected-neighbour cluster joins points within a chosen distance and minimum count. Radar usually needs different thresholds from dense lidar. Tune on training scenes, then preserve a held-out scene so the result is not judged on the same points that set the rules.

6.7 Track and score

This section: Track and score — A simple tracker predicts a cluster position, finds a compatible new cluster, and updates the track.

A simple tracker predicts a cluster position, finds a compatible new cluster, and updates the track. Gating limits unlikely matches; track age and missed-update counts support deletion. Score point or cluster precision and recall, missed objects, false clusters, and identity switches. A small sample cannot represent all roads, rooms, weather, mounts, or people.

6.8 Dataset and synthetic boundaries

This section: Dataset and synthetic boundaries — The lab includes a small derived sample from RadarScenes sequence 158.

The lab includes a small derived sample from RadarScenes sequence 158. RadarScenes is licensed CC BY-NC-SA 4.0 and supplies automotive radar detections with point-wise labels and track IDs. The lidar room scan is synthetic because no compatible small public scan was selected for this wave. Both files are for teaching the pipeline; neither validates a product or a new environment.

6.9 How a track inherits measurement error

This section: How a track inherits measurement error — Transform each detection with its timestamp and sensor pose before comparing neighbours.

Transform each detection with its timestamp and sensor pose before comparing neighbours. If the platform moved between a lidar sweep and a radar return, applying one pose to both can manufacture a cluster even when their points were never co-located. The radar radial-velocity field also describes motion along its line of sight; it cannot be copied into a full planar velocity vector. Keep the original sensor label and quality fields after transformation so a reviewer can trace a false track back to its input.

A clusterer’s distance and minimum-count settings set two opposing failure modes: a tight radius fragments one object, while a loose radius merges nearby objects. On the chapter’s RadarScenes-derived sample, tune on a named training portion, then score missed objects, false clusters, and identity switches on the held-out portion. Repeat the same record for the synthetic lidar room, labelled as synthetic. Those two exercises test processing logic under different density; they do not establish a shared field accuracy or justify discarding the weaker modality’s contradictory returns.

6.10 Decision and Trade-offs

This section: Decision and Trade-offs — Keep radar when radial velocity and weather tolerance matter, but plan for sparse, ambiguous returns.

Keep radar when radial velocity and weather tolerance matter, but plan for sparse, ambiguous returns. Keep lidar when dense geometry matters and optical conditions, cleaning, power, and cost fit. Fuse them only after timestamps, coordinate frames, coverage, and failure independence are tested. A merged picture is not automatically more truthful than either input.

Use the Sensing Privacy and Consent checklist before retaining raw point clouds or sharing tracked paths.

6.11 Practice the Method

This section: Practice the Method — The linked lab uses a prepared online tool and retains its result evidence.

The linked lab uses a prepared online tool and retains its result evidence. Complete the steps, compare the result with the chapter boundary, and name one condition that the exercise did not test.

6.12 Check Your Reasoning

This section: Check Your Reasoning

6.13 Summary

This section: Summary — The useful output is an auditable chain from timestamped sensor records to scored associations.

The useful output is an auditable chain from timestamped sensor records to scored associations. Preserve the source frame, filtering rules, clustering settings, track state, and rejected points beside the final objects. That chain shows whether an apparent improvement came from better sensing, a tuned threshold, or a scoring rule that hid difficult cases.

• A point is useful only with its frame, unit, time, sensor, and quality fields.
• Filtering, clustering, and tracking introduce thresholds whose errors must be scored.
• The RadarScenes subset and synthetic lidar scan teach processing, not site or product validation.