Improve precision of lane attribution

[vcharraut]

What

Improve lane attribution at intersections and between opposing lanes, while reducing frame-to-frame lane flicker.

Replace the linear distance/heading score and topology-dependent switching costs with a vehicle-axis fit:

fit = distance_to_segment²
    + (vehicle_length / 2)² × sin²(heading_difference)
    + (vehicle_width / 2)² × (1 − cos(heading_difference)) / 2

The axis term distinguishes crossing lanes while treating forward and reverse headings as parallel to the same lane. The smaller direction term resolves nearby opposing lanes without dominating physical proximity.

• Preserve finite-segment projection, the three-vehicle-width search cutoff, and interpolated height filtering.
• Add 0.35 × (vehicle_width / 2)² when switching away from an existing lane. This stabilizes near-equal candidates; an ineligible previous lane cannot remain selected.
• Keep one best candidate in one segment scan. Remove topology lookups and four obsolete selection-weight/switching constants; no candidate groups, timers, or new agent state.

How

This PR has been totally made via LLM. I created a benchmark to quantify the current score of the lane selection algorithm and asked a LLM to improve it.

Benchmark and methodology

• Dataset: 330 deterministic poses on CARLA Town10HD, selected across 17 categories from a 2,040-case generator. Coverage includes aligned/ambiguous intersections, opposite and perpendicular headings, lateral offsets, parallel/opposing lane boundaries, lane history/transitions, off-lane poses, and vertical separation.
• Challenging boundary cases: 25 poses at a 40/60 split between parallel lane centers, 25 at 30/70, and 30 positioned 70% toward an opposing lane while facing along the farther lane.
• Reproducibility: x/y perturbations are uniform in [-0.25, 0.25] m with seed 20260808. Vehicles are 4.5 m long, 2 m wide, and evaluated at 10 m/s.
• Ground truth: human annotations specify acceptable lane IDs and a wrong-way flag. Multiple lane IDs may be accepted for an ambiguous pose. Annotation images display lane IDs and direction arrows without the selector output; the generator does not assign expected lanes.

Before / after results

Metric	Previous	Latest	Change
Lane accuracy	80.30% (265/330)	93.64% (309/330)	+13.33 pp
Wrong-way accuracy	84.24% (278/330)	97.27% (321/330)	+13.03 pp
Combined accuracy	78.48% (259/330)	91.52% (302/330)	+13.03 pp
Weighted global score	78.33%	91.77%	+13.44 pp

Lane accuracy accepts any human-listed lane ID. Wrong-way accuracy compares the annotated flag with the selected lane's heading (cos(heading_difference) < 0 above 2.5 m/s). Combined accuracy requires both to be correct.

The weighted score is sum(case_weight × combined_correct) / sum(case_weight): all between_* cases have weight 1.2, ambiguous intersections 0.95, and other cases 1.0. Percentage-point changes use unrounded values.

Largest lane-accuracy gains:

Category	Previous	Latest
Between opposing lanes, 70/30	8/30	27/30
Ambiguous intersections	5/25	18/25
Opposite heading	33/40	40/40
Between parallel lanes, 30/70	23/25	25/25

On combined correctness, 46 cases improve and 3 regress: intersection_000_ambiguous, perpendicular_lane_97, and offset_left_lane_167.

Intersection stability

A separate temporal ablation replays 50 intersection poses for 20 frames each, alternates position by ±2 cm and heading by ±0.5°, and feeds the selected lane back into the next frame.

• Vehicle-axis fit without hysteresis: 114 lane switches across 6 flickering cases.
• Vehicle-axis fit with the latest switching margin: 0 switches.

This comparison isolates the switching margin within the new algorithm; it is not the old-versus-new static benchmark. Increasing the coefficient from 0.25 to 0.35 leaves all 330 static result rows unchanged.