Honda closure review

Nine scenes comparing camera-only detection with one-sweep and temporal LiDAR. Each case says what worked, what failed, and why.

9reviewed scenes
2/9single-sweep LiDAR pairs published
3 / 4 / 2temporal useful / safe blank / missed lane
9.80 mslargest LiDAR-camera time difference
CameraFirst output from a fresh run of the generic three-camera model
One-step LiDAROne independently processed LiDAR sweep
Shared scope0.5–35 m ahead and 4 m to each side
Ground truthNone; this is a visual review, not an accuracy score
workingframe 900

Straight urban road

Working after LiDAR aggregation

Fair control: fresh generic camera step vs one LiDAR sweep. Click the image to inspect it.
Final temporal LiDAR result. Click the image to inspect it.
Camera-only

The camera finds a useful left and right road corridor, although some parts are broken into short pieces.

LiDAR: one sweep

One sweep finds two lines, but the 1.61 m pair is too narrow and does not surround the vehicle. Nothing is published.

LiDAR: final temporal result

Two boundaries published after ten-sweep aggregation

Direct comparison

The camera works in one step. LiDAR needs several sweeps, then gives a similar road corridor. The left LiDAR line may be a road edge rather than painted lane marking.

Main LiDAR limitOne-sweep line geometry
LiDAR lines found / published2 / 0
Camera marked cells694
Frame timing difference2.67 ms
LiDAR runtime1194 ms
Why this result happened

The useful LiDAR result appears only after aggregation. This is a recovery from sparse one-sweep evidence, not a one-step success.

Raw reasons LiDAR did not publish a pair
  • ego-centre offset 2.72m exceeds 1.00m
  • pair does not bracket the ego origin
  • separation 1.61m is below 2.50m
workingframe 1000

Clear downhill urban lane

Strongest working case

Fair control: fresh generic camera step vs one LiDAR sweep. Click the image to inspect it.
Final temporal LiDAR result. Click the image to inspect it.
Camera-only

The camera sees the road boundaries, but its output is broken into several short and noisy pieces.

LiDAR: one sweep

One sweep publishes a clean 3.07 m pair that surrounds the vehicle for 15.52 m.

LiDAR: final temporal result

Two boundaries published from one and ten sweeps

Direct comparison

LiDAR is clearer than the camera here. It produces a simple lane pair while the camera result is more fragmented.

Main LiDAR limitNo main LiDAR failure
LiDAR lines found / published3 / 2
Camera marked cells623
Frame timing difference7.84 ms
LiDAR runtime504 ms
Why this result happened

Both pipelines see useful road structure. LiDAR gives the cleaner output for this frame.

Raw reasons LiDAR did not publish a pair
  • pair does not bracket the ego origin
  • separation 1.37m is below 2.50m
  • separation 1.70m is below 2.50m
workingframe 2600

Narrow straight urban road

Working after LiDAR aggregation

Fair control: fresh generic camera step vs one LiDAR sweep. Click the image to inspect it.
Final temporal LiDAR result. Click the image to inspect it.
Camera-only

The camera produces two clear, long boundaries and works well in one step.

LiDAR: one sweep

One sweep proposes a short 1.55 m pair. It is too narrow and has only 4.00 m overlap, so nothing is published.

LiDAR: final temporal result

Ten sweeps recover two short aligned boundaries

Direct comparison

The camera is stronger for one-step detection. Temporal LiDAR becomes usable after aggregation, but its result is shorter.

Main LiDAR limitSparse and short single-sweep evidence
LiDAR lines found / published2 / 0
Camera marked cells1447
Frame timing difference6.66 ms
LiDAR runtime314 ms
Why this result happened

This is the clearest example of aggregation helping LiDAR. The camera does not need that recovery step.

Raw reasons LiDAR did not publish a pair
  • ego-centre offset 2.19m exceeds 1.00m
  • overlap 4.00m is below 5.00m
  • pair does not bracket the ego origin
  • separation 1.55m is below 2.50m
averageframe 800

Turning interval

Average: camera works better; LiDAR safely abstains

Fair control: fresh generic camera step vs one LiDAR sweep. Click the image to inspect it.
Final temporal LiDAR result. Click the image to inspect it.
Camera-only

The camera follows the curved road edges reasonably well, with some broken pieces in the bird's-eye view.

LiDAR: one sweep

One sweep produces one line, so it cannot form a lane pair.

LiDAR: final temporal result

Temporal window rejected for 21.51° yaw

Direct comparison

The camera is more useful in this curved scene. LiDAR gives no lane pair, but its final blank result is safer than merging strongly rotated scans.

Main LiDAR limitVehicle-motion check
LiDAR lines found / published1 / 0
Camera marked cells1488
Frame timing difference5.91 ms
LiDAR runtime449 ms
Why this result happened

This scene is outside the LiDAR method's straight or slowly-curving road assumption.

Raw reasons LiDAR did not publish a pair
  • None for the accepted pair.
averageframe 1400

Intersection and crosswalk

Average: conservative result at an intersection

Fair control: fresh generic camera step vs one LiDAR sweep. Click the image to inspect it.
Final temporal LiDAR result. Click the image to inspect it.
Camera-only

The camera detects crosswalk and road-edge structure, but it does not give a simple ego-lane pair.

LiDAR: one sweep

LiDAR fits two lines, but their slope and short overlap do not look like a valid longitudinal lane pair.

LiDAR: final temporal result

No lane family published

Direct comparison

The camera gives richer road-marking context. LiDAR gives less information, but correctly avoids calling the crosswalk a lane.

Main LiDAR limitFinal geometry checks
LiDAR lines found / published2 / 0
Camera marked cells425
Frame timing difference3.37 ms
LiDAR runtime406 ms
Why this result happened

The LiDAR pair has 4.00 m overlap, 0.302 maximum slope, and 0.138 slope difference. All three fail the fixed limits.

Raw reasons LiDAR did not publish a pair
  • individual slope 0.302 exceeds 0.250
  • overlap 4.00m is below 5.00m
  • slope difference 0.138 exceeds 0.080
averageframe 3000

Intersection approach

Average: safe blank result

Fair control: fresh generic camera step vs one LiDAR sweep. Click the image to inspect it.
Final temporal LiDAR result. Click the image to inspect it.
Camera-only

The camera marks several road edges and intersection features, but the result is not a clean lane pair.

LiDAR: one sweep

One LiDAR line is fitted. A second matching line is missing, so no pair can be published.

LiDAR: final temporal result

No lane family published

Direct comparison

The camera shows more scene structure. LiDAR is less informative but avoids publishing an unsupported lane pair.

Main LiDAR limitNot enough matching evidence on both sides
LiDAR lines found / published1 / 0
Camera marked cells826
Frame timing difference2.20 ms
LiDAR runtime351 ms
Why this result happened

After surface filtering, 312 candidates and 70 cleaned cells remain. Only one line is fitted.

Raw reasons LiDAR did not publish a pair
  • None for the accepted pair.
failureframe 348

Side-looking, nearly stationary road scene

Failure: one-sweep LiDAR false positive

Fair control: fresh generic camera step vs one LiDAR sweep. Click the image to inspect it.
Final temporal LiDAR result. Click the image to inspect it.
Camera-only

The camera mainly sees a road edge and sidewalk. It does not show a forward ego-lane pair.

LiDAR: one sweep

LiDAR wrongly publishes two long lines across kerb and roadside structure as a 4.28 m pair.

LiDAR: final temporal result

Temporal window rejected as redundant

Direct comparison

The camera is safer in this frame. One-sweep LiDAR has a clear false positive; the temporal motion check protects the final system.

Main LiDAR limitScene is outside the expected use
LiDAR lines found / published3 / 2
Camera marked cells1151
Frame timing difference2.32 ms
LiDAR runtime2125 ms
Why this result happened

The geometry checks alone cannot tell a kerb pair from a lane pair in this side-looking scene. The 0.03 m motion check stops it from becoming the final result.

Raw reasons LiDAR did not publish a pair
  • ego-centre offset 1.05m exceeds 1.00m
  • ego-centre offset 1.10m exceeds 1.00m
  • pair does not bracket the ego origin
  • separation 1.61m is below 2.50m
failureframe 1200

Straight road with visible paint

Failure: LiDAR false negative

Fair control: fresh generic camera step vs one LiDAR sweep. Click the image to inspect it.
Final temporal LiDAR result. Click the image to inspect it.
Camera-only

The camera produces strong left and right boundary traces and is clearly useful.

LiDAR: one sweep

LiDAR finds a plausible 3.08 m pair with 10.08 m overlap, but rejects it because the line slopes differ by 0.082.

LiDAR: final temporal result

No lane family published

Direct comparison

The camera clearly works better. LiDAR misses a visible lane because 0.082 is just above the fixed 0.080 slope-difference limit.

Main LiDAR limitFinal geometry threshold
LiDAR lines found / published2 / 0
Camera marked cells1326
Frame timing difference9.80 ms
LiDAR runtime455 ms
Why this result happened

The final gate is too strict for this borderline but visually useful pair.

Raw reasons LiDAR did not publish a pair
  • slope difference 0.082 exceeds 0.080
failureframe 2800

Straight road with asymmetric marking

Failure: LiDAR false negative

Fair control: fresh generic camera step vs one LiDAR sweep. Click the image to inspect it.
Final temporal LiDAR result. Click the image to inspect it.
Camera-only

The camera produces long, roughly parallel road boundaries and works well.

LiDAR: one sweep

LiDAR finds three line pieces, but no two pieces form an acceptable lane pair.

LiDAR: final temporal result

No lane family published

Direct comparison

The camera works better. LiDAR misses the lane because evidence is stronger on one side and the system requires a valid pair.

Main LiDAR limitUneven evidence and two-line requirement
LiDAR lines found / published3 / 0
Camera marked cells1672
Frame timing difference8.62 ms
LiDAR runtime1009 ms
Why this result happened

Candidate pairs have only 3.84–4.00 m overlap or implausible 1.73–7.65 m separation.

Raw reasons LiDAR did not publish a pair
  • ego-centre offset 2.97m exceeds 1.00m
  • overlap 3.84m is below 5.00m
  • overlap 4.00m is below 5.00m
  • pair does not bracket the ego origin
  • separation 1.73m is below 2.50m
  • separation 5.92m exceeds 4.50m
  • separation 7.65m exceeds 4.50m
Failure characterization

What pattern do these nine scenes show?

Short answer: LiDAR works best when the road is straight and both sides give long, roughly parallel evidence. Most misses come from weak or uneven evidence, a strict final check, or a scene outside that assumption.

Works bestStraight road + two clear sides

The cleanest one-sweep result is frame 1000.

More scans helpSparse evidence can be recovered

Temporal aggregation recovers useful boundaries in frame 900 and frame 2600.

Main weak conditionsTurns, intersections, side views and uneven markings

See 800, 1400, 3000, 348 and 2800.

Main failure types

False positiveRoadside structure looks like a lane

In a side-looking scene, kerbs and roadside structure pass the geometric checks. The temporal motion check prevents this one-sweep error from becoming the final result.

Frame 348 →
False negativeA fixed threshold is slightly too strict

A visually useful pair is rejected because its slope difference is 0.082, just above the 0.080 limit.

Frame 1200 →
False negativeOne side is weaker than the other

The method requires two matching lines. Uneven markings produce short pieces, poor overlap or an implausible pair width.

Frame 2800 →
Safe blankNo lane is better than a wrong lane

On a turn or at an intersection, the system has too little support for a simple lane pair and correctly publishes nothing. These are limitations, but not false detections.

800 · 1400 · 3000
Bottom line

The failures are not random. They mostly follow the method's assumptions: it expects two long, straight, parallel boundaries. Temporal aggregation helps when a single sweep is sparse, but it does not fix a hard threshold, a missing side, or a confusing roadside structure.

What to improve next: replace borderline hard cut-offs with confidence scoring, support a temporary one-boundary estimate when one side is weak, and add a stronger viewpoint or roadside-structure check.

Scope: nine selected scenes with no lane ground truth. These are observed failure patterns, not failure rates or an accuracy score.