AUTOMOTIVE CASE STUDY
Automotive case study
One road scene.Connected sensor data.
A road-perception project shows how scene planning, object associations and issue review become an inspectable, traceable project record.
All case studies
01 / PROJECT BRIEF
Define the task before the batch.
A perception team wants to examine scene coverage and object identities across sensors in one road-data batch. Sequences provide the structure; rules and review records follow every revision.
- Client profile
- automotive perception team
- Task
- 2D / 3D object and track associations
- setup
- 6 cameras + 1 LiDAR
- Scope
- 12 sequences, each with 100 synchronized keyframe groups
02 / SCENE COVERAGE
Four road scenarios. One association rule set.
Keep each scene’s scope and review focus visible. Select a group to explore the breakdown: 12 sequences and 1,200 synchronized keyframe groups in total.
JUNCTIONS
Junctions
Turning vehicles, pedestrians and roadside occlusion occur together. Review identity consistency across multiple views.
- Sequences
- 4
- Keyframe groups
- 400
Review focus
- Object references across cameras and LiDAR
- Track continuity and termination around occlusion
- Near-range boundaries and orientation
Example recordSEQ-J01 / KF-042 · Check the association ID for a turning vehicle across adjacent views.
NIGHT & BACKLIGHT
Night & backlight
Low illumination, reflections and backlight make boundaries uncertain. Keep uncertainty and review evidence in the record.
- Sequences
- 3
- Keyframe groups
- 300
Review focus
- Boundaries for objects with low visibility
- Joint review of weak image texture and point-cloud evidence
- Keep an unresolved status when evidence is insufficient
Example recordSEQ-N02 / KF-067 · One occluded object association remains unresolved and is excluded from passed groups.
ROADWORKS
Roadworks
Temporary barriers, cones and changing drivable space require consistent categories, attributes and scene tags.
- Sequences
- 2
- Keyframe groups
- 200
Review focus
- Classes and attributes for temporary infrastructure
- Visibility changes across consecutive frames
- Links between rule versions and issue records
Example recordSEQ-R01 / KF-031 · Check a cone category against the rule version and retain the revision record.
MIXED TRAFFIC
Mixed traffic
Vehicles, cyclists and pedestrians interact in nearby areas. Review identity switches and short-range occlusion.
- Sequences
- 3
- Keyframe groups
- 300
Review focus
- Class boundaries between road users
- Identity continuity after brief occlusion
- Consistency of motion, orientation and tracks
Example recordSEQ-M03 / KF-088 · Review a cyclist’s identity in adjacent frames and correct the association.
Four scenes cover 12 sequences and 1,200 synchronized keyframe groups, organized by sequence and rule version.
03 / WORKING METHOD
Turn difficult scenes into explicit review tasks.
Align rules with representative examples, then annotate, review and revise by sequence. Every issue leads back to a scene and a rule version.
- 01
Objects can be mismatched across sensors
ApproachVerify timing and coordinate references, then link images, point clouds and tracks with object IDs. Keep ambiguous links pending.
Record retainedTimestamps, object IDs, calibration versions and association tables
- 02
Occlusion and low light obscure boundaries
ApproachUse edge cases to align visibility and occlusion rules. Record uncertainty according to the rules instead of filling gaps by guesswork.
Record retainedRule versions, example notes and issue locations
- 03
Revisions need a fresh review
ApproachReturn each issue to its sequence and keyframe group. Record the initial check, rework and recheck separately.
Record retainedIssue IDs, revision references, recheck decisions and unresolved items
04 / REVIEW RECORD
Record the corrections. Keep open issues visible.
This stratified review covers 120 groups: 30 from each scene, so sampling fractions differ with scene size. The other 1,080 groups are uninspected and excluded from every pass rate.
| Scene | Sampled | Initial pass | Recheck pass / rework | Currently passed |
|---|---|---|---|---|
| Junctions | 30 | 29 | 1 / 1 | 30 |
| Night & backlight | 30 | 28 | 1 / 2 | 29 |
| Roadworks | 30 | 29 | 1 / 1 | 30 |
| Mixed traffic | 30 | 29 | 1 / 1 | 30 |
| Total | 120 | 115 | 4 / 5 | 119 |
Rates apply to the stated review sample. Batch acceptance follows agreed criteria; the 1,080 uninspected groups are tracked separately.
Explore delivery and quality05 / HANDOVER STRUCTURE
Align the data and the documentation.
Organize indexes, labels, rule versions and review records into one manifest for receipt, verification and iteration.
- 01
Sequence and synchronization indexes
Sequence IDs, timestamps, six image references and one point-cloud reference, grouped by scenario.
- 02
Labels and cross-modal associations
2D / 3D objects, track IDs and association tables organized against an agreed rule version.
- 03
Calibration and rule references
Sensor setup, coordinate and calibration versions, example notes and change records.
- 04
Review and unresolved issues
Sample denominators, initial and recheck decisions, revision references and open issues.
PROJECT TAKEAWAYS
Three principles for connected delivery.
- 01
Agree on counting units
Separate sequences, synchronized groups and sensor files so the scope stays clear.
- 02
Keep issues in context
Trace every revision to its scenario, sequence, rule and review decision.
- 03
Keep unresolved items visible
Show what has not passed and what has not been inspected, with a clear next step.
LET’S DEFINE YOUR DATA TASK
Start with your scenarios and rules.
Describe your sensor setup, scene scope and expected outputs to frame a suitable pilot task.