Classical LiDAR Lane Detection

Checkpoint	Milestone	State
01, 03–05	Input, road surface, reflectivity, candidates	Complete
5.1–09	Candidate audit, road audit, BEV cleanup	Complete
11	Adaptive lane geometry	Complete
13–14.1	Better RANSAC and lane-family geometry	Complete

CHECKPOINT 01

Input & coordinate gate

Complete

Does	Reads K-Lane point clouds and labels. Matches each point cloud to its label and metadata. Verifies the LiDAR-to-BEV coordinate convention without using the ML detector.
Outcome	The input contract and top-forward, left-positive orientation were approved, establishing a deterministic CPU-only foundation. Source report

First three artifacts

CHECKPOINT 03

Dominant road plane

Complete

Does	Samples only points inside the fixed ROI. Fits one deterministic near-horizontal RANSAC plane. Retains complete point rows within a 0.12 m distance band.
Outcome	Road points, aligned masks, plane parameters, and support statistics were produced for every inspected frame. Source report

First three artifacts

CHECKPOINT 04

Raw reflectivity signal

Complete

Does	Selects reflectivity from the Stage-3 road-point rows. Preserves one-to-one alignment with those rows. Applies no threshold, normalization, clipping, or smoothing.
Outcome	A finite raw marking signal plus spatial and histogram diagnostics was exposed for each frame. Source report

First three artifacts

CHECKPOINT 05

Global Otsu candidates

Complete

Does	Computes one exact-value Otsu threshold per frame. Marks road returns above that threshold as candidates. Uses labels only for evaluation and visualization.
Outcome	The candidate mask is deterministic and label-independent, but still contains scan structures and clutter. Source report

First three artifacts

STAGE 5.1 · EXPLORATORY

Candidate quality audit

Complete

Does	Expands the candidate audit to 39 readable frames. Measures strict and one-cell-tolerant candidate quality. Tests a top-decile, row-budget proposal against raw Otsu.
Outcome	Tolerant F1 rose from 17.16% to 21.02%, but the proposal remained an ablation and was not adopted for Stage 6. Source report

Raw Otsu vs Stage 5.1
Raw Otsu	Stage 5.1
What it is: Otsu is an automatic thresholding method. For each frame, it chooses the reflectivity cutoff that best separates the points into a darker group and a brighter group; the brighter group becomes the lane-candidate set.	What it is: Stage 5.1 replaces that adaptive split with a fixed percentile rule: only the brightest 10% of road-point returns are eligible.
BEV selection: Every above-threshold road point remains a candidate, with no per-row limit.	BEV selection: Keeps at most the six strongest occupied cells in each BEV row.
Behavior: Can select too few points or broad transverse scan arcs.	Behavior: Raises recall while limiting how much of one row a scan arc can fill.

Candidate ablation · 39 development frames
Output & metric	Precision	Recall	F1
Raw Otsu · strict	6.36%	3.60%	4.60%
Stage 5.1 · strict	7.25%	4.72%	5.72%
Raw Otsu · tolerant	22.59%	13.84%	17.16%
Stage 5.1 · tolerant	25.24%	18.01%	21.02%

First three artifacts

CHECKPOINT 08

Road-plane coverage audit

Complete

Does	Keeps the existing Stage-3 fit unchanged. Measures retention across eight near-to-far bands. Checks whether available lane-label cells keep road-point support.
Outcome	Median spatial retention was 72.25%, but three frames had near-total far-field collapse; Stage 3.1 was deferred. Source report

Spatial coverage audit by sequence
Scope	Frames	Overall retention	Far retention	Label retention
Overall	39	72.25%	50.19%	77.17%
Sequence 01	10	78.37%	66.16%	84.35%
Sequence 02	10	83.18%	68.65%	82.24%
Sequence 03	10	72.54%	51.79%	78.00%
Sequence 04	9	57.24%	44.69%	46.68%

First three artifacts

CHECKPOINT 09 · STAGE 6

BEV candidate cleanup

Complete

Does	Rasterizes raw Otsu candidates at 288 × 288. Closes short forward gaps and measures connected components. Conservatively removes weak, long, or wide clutter.
Outcome	Precision rose from 22.59% to 25.75% and F1 from 17.16% to 17.83%, retaining 98.48% of baseline true positives. Source report

Operation-by-operation ablation · tolerant metric
Output	Precision	Recall	F1
Raw Stage 5	22.59%	13.84%	17.16%
Short forward gap close	22.45%	13.84%	17.12%
Direct component filter	25.78%	13.63%	17.83%
Final + dash chains	25.75%	13.63%	17.83%

First three artifacts

CHECKPOINT 11 · STAGE 7

Adaptive lane geometry

Complete

Does	Groups cleaned lane-like dots with RANSAC. Fits straight, quadratic, cubic, and adaptive curves. Applies spacing, crossing, slope, bend, and support limits.
Outcome	Adaptive fitting reached 19.44% held-out tolerant F1 and raised exact-cell F1 from 5.01% to 12.22%. Source report

Held-out model comparison · tolerant metric
Output	Precision	Recall	F1
Stage-6 input	26.34%	15.14%	19.23%
Straight	26.84%	14.30%	18.66%
Quadratic	27.94%	14.82%	19.37%
Cubic	28.77%	14.86%	19.60%
Adaptive · recommended	27.93%	14.91%	19.44%

Cubic has the highest mask F1; adaptive remains recommended because its stability evidence is stronger.

First three artifacts

CHECKPOINT 13

Better RANSAC evidence

Complete

Does	Compares four explicit RANSAC input routes. Aggregates raw returns into equal-vote BEV cells. Tests bounded reflectivity and road-height guidance.
Outcome	The Stage-6 baseline stayed best at 19.44% held-out tolerant F1; guided Stage 4 reached 13.62%. Source report

Frozen held-out input ablation · tolerant metric
Input mode	Precision	Recall	F1
Stage-6 mask	27.93%	14.91%	19.44%
Stage-5 points	24.53%	13.34%	17.28%
Stage-4 uniform	8.64%	11.51%	9.87%
Stage-4 guided	12.66%	14.73%	13.62%

Best ranked samples

Average ranked samples

Worst ranked samples

CHECKPOINT 14 · HISTORICAL

Guarded lane-family geometry

Superseded

Does	Preserves independent Stage-7 fitting as a baseline. Tests translated anchor copies and a shared road shape. Applies ordering, support, and physical safeguards.
Outcome	Independent fitting scored 17.43% held-out tolerant F1 versus 17.31% for guarded exact copies; raw claims were corrected in 14.1. Source report

Frozen geometry comparison · tolerant metric
Geometry mode	Precision	Recall	F1
Independent	24.73%	13.46%	17.43%
Exact copies · guarded	22.86%	13.92%	17.31%
Shared shape · guarded	23.19%	13.62%	17.16%

Exact-copy input sweep · held-out tolerant F1
Input	Precision	Recall	F1
Stage-6 mask	22.86%	13.92%	17.31%
Stage-5 points	24.84%	14.63%	18.42%
Stage-4 uniform	8.78%	11.94%	10.12%
Stage-4 guided	10.24%	12.63%	11.31%

The Stage-5 held-out reversal was reported but not adopted because it did not appear during development.

Best ranked samples

Average ranked samples

Worst ranked samples

CHECKPOINT 14.1 · FINAL

Corrected exact-copy family

Complete

Does	Selects one accepted Stage-7 lane as the anchor. Renders six literal lateral copies of its shape and gaps. Measures raw family masks without independent-fit fallback.
Outcome	Held-out tolerant F1 improved from 17.43% to 21.21%; exact copying won, while independent Stage 7 stayed the safer default. Source report

Corrected raw-family comparison · tolerant metric
Mode	Precision	Recall	F1
Independent Stage 7	24.73%	13.46%	17.43%
Raw exact copies	20.40%	22.09%	21.21%
Raw shared shape	27.72%	13.28%	17.96%
Guarded family policy	25.08%	13.25%	17.34%

Raw-exact input-route ablation
Input	Development F1	Held-out F1	Held-out strict F1
Stage-6 mask	23.37%	21.21%	11.91%
Stage-5 points	20.41%	21.03%	12.07%
Stage-4 uniform	13.13%	12.03%	6.13%
Stage-4 guided	15.11%	13.61%	6.60%

Best ranked samples

Average ranked samples

Worst ranked samples