File size: 22,908 Bytes
1f47b06
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
from __future__ import annotations

import os

# macOS: PyTorch + NumPy/OpenCV may load libomp twice; avoid abort on import.
os.environ.setdefault("KMP_DUPLICATE_LIB_OK", "TRUE")

import argparse
import sys
from datetime import datetime

import torch
import cv2
import numpy as np
from pathlib import Path
from typing import Optional, Tuple

from ultralytics import YOLO

_orig_load = torch.load


def _patched_load(*args, **kwargs):
    kwargs["weights_only"] = False
    return _orig_load(*args, **kwargs)


torch.load = _patched_load

_DIR = Path(__file__).resolve().parent
_ROOT = _DIR.parent
MODELS_DIR = _ROOT / "models"
RESULTS_DIR = _ROOT / "results"

MODEL_PATH = MODELS_DIR / "segmentor" / "tram-dynamic-roi-tracker-yolo11s.pt"
VIDEO_PATH = str(_ROOT / "/Users/monadayzek/Desktop/Ayzek/1-spbstu/diploma/objects-recognation/yolo/a.mp4")
OUTPUT_PATH = str(RESULTS_DIR / "videos" / "roi_zones_stable___3.mp4")

CONF = 0.325


def _detect_device() -> str:
    if torch.cuda.is_available():
        return "cuda"
    if torch.backends.mps.is_available():
        return "mps"
    return "cpu"


DEVICE = os.environ.get("TRAM_DEVICE") or _detect_device()

LOW_CONF_DIR = RESULTS_DIR / "low-conf"
LOW_CONF_NORM = 0.5

RAIL_CLS = {0, 1, 3, 4}

RAIL_BRANCH = 0
_MAX_CENTER_JUMP_PX = 120

GAUGE_MM = 1524
DEFAULT_FOV_DEG = 70.0
ZONE_DISTS_MM = {
    "red": 500,
    "orange": 1000,
    "yellow": 1500,
}
GRID_STEP_MM = 2000

ZONE_ALPHA = 0.30
ZONE_COLORS = {
    "red": (0, 0, 255),
    "orange": (0, 140, 255),
    "yellow": (0, 220, 255),
}
GRID_COLOR = (255, 255, 255)
TEXT_COLOR = (255, 255, 255)

MORPH_KERNEL = (5, 5)
MORPH_CLOSE_ITER = 1
MORPH_OPEN_ITER = 1
USE_MORPH = False

# Mask extraction: soft (probabilistic) accumulation + sub-pixel resize.
MASK_PROB_THRESH = 0.5
# Always bridge tiny seg holes (cheap, stabilizes per-row edges).
_BASE_CLOSE_KERNEL = (3, 3)
# rail_rows: bridge gaps up to this many px when finding the contiguous bed run.
_ROW_GAP_PX = 14
# Line rendering style (lightweight: outlines only, no filled polygons).
_ZONE_LINE_THICK = 2
_RAIL_LINE_THICK = 2

# Distance model: minimum rows for a trustworthy line fit, robust-trim params.
_FIT_MIN_ROWS = 6
_FIT_ITERS = 4
_FIT_TRIM_SIGMA = 2.0


def load_model(path: Optional[Path] = None) -> YOLO:
    p = Path(path) if path else MODEL_PATH
    if not p.exists():
        print(
            f"[ERROR] Модель не найдена: {p}\n"
            "Запустите: python scripts/download_models.py"
        )
        sys.exit(1)
    return YOLO(str(p))


def _max_rail_confidence(res) -> Optional[float]:
    if res.boxes is None or len(res.boxes) == 0:
        return None
    cls = res.boxes.cls.cpu().numpy().astype(int)
    conf = res.boxes.conf.cpu().numpy()
    rail = np.isin(cls, list(RAIL_CLS))
    if not np.any(rail):
        return None
    return float(np.max(conf[rail]))


def _next_low_conf_path() -> Path:
    LOW_CONF_DIR.mkdir(parents=True, exist_ok=True)
    stem = datetime.now().strftime("%d_%m_%Y")
    p = LOW_CONF_DIR / f"{stem}.png"
    if not p.exists():
        return p
    n = 1
    while True:
        p = LOW_CONF_DIR / f"{stem}_{n}.png"
        if not p.exists():
            return p
        n += 1


def _maybe_save_low_conf_frame(frame: np.ndarray, max_rail_conf: Optional[float]) -> None:
    if max_rail_conf is None or max_rail_conf >= LOW_CONF_NORM:
        return
    path = _next_low_conf_path()
    cv2.imwrite(str(path), frame)


def get_focal_px(frame_w: int, fov_deg: float = DEFAULT_FOV_DEG) -> float:
    return frame_w / (2.0 * np.tan(np.radians(fov_deg) / 2.0))


def extract_rail_mask(results, h: int, w: int, morph: bool = False) -> np.ndarray:
    """Accumulate rail-class masks as probabilities, then threshold.

    Soft (float) accumulation + bilinear upscale yields sub-pixel-accurate,
    less jagged boundaries than per-mask NEAREST + binary OR. A light close is
    always applied to bridge tiny segmentation holes for stable per-row edges.
    """
    if results.masks is None:
        return np.zeros((h, w), dtype=np.uint8)

    prob = np.zeros((h, w), dtype=np.float32)
    has_any = False
    for m, c in zip(
        results.masks.data.cpu().numpy(),
        results.boxes.cls.cpu().numpy().astype(int),
    ):
        if c in RAIL_CLS:
            mr = cv2.resize(
                m.astype(np.float32), (w, h), interpolation=cv2.INTER_LINEAR
            )
            np.maximum(prob, mr, out=prob)
            has_any = True

    if not has_any:
        return np.zeros((h, w), dtype=np.uint8)

    mask = (prob >= MASK_PROB_THRESH).astype(np.uint8) * 255
    base_k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, _BASE_CLOSE_KERNEL)
    mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, base_k, iterations=1)
    return _clean_rail_mask(mask) if morph else mask


def _fill_holes(mask: np.ndarray) -> np.ndarray:
    """Fill enclosed background holes inside the rail blob (flood from border)."""
    if not np.any(mask):
        return mask
    h, w = mask.shape
    ff = mask.copy()
    flood = np.zeros((h + 2, w + 2), dtype=np.uint8)
    cv2.floodFill(ff, flood, (0, 0), 255)
    holes = cv2.bitwise_not(ff)
    return cv2.bitwise_or(mask, holes)


def _clean_rail_mask(mask: np.ndarray) -> np.ndarray:
    if not np.any(mask):
        return mask
    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, MORPH_KERNEL)
    mask = cv2.morphologyEx(mask, cv2.MORPH_OPEN, kernel, iterations=MORPH_OPEN_ITER)
    mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel, iterations=MORPH_CLOSE_ITER)
    mask = _fill_holes(mask)
    return mask


def _medfilt1d(a: np.ndarray, k: int = 7) -> np.ndarray:
    n = len(a)
    if n < k:
        return a.copy()
    if k % 2 == 0:
        k += 1
    pad = k // 2
    padded = np.pad(a.astype(np.float64), pad, mode="edge")
    out = np.empty(n, dtype=a.dtype)
    for i in range(n):
        out[i] = np.median(padded[i : i + k])
    return out


def _smooth1d(a: np.ndarray, k: int = 21) -> np.ndarray:
    n = len(a)
    if n < 5:
        return a.copy()
    k = min(k, n // 2 * 2 + 1)
    if k < 3:
        k = 3
    if k % 2 == 0:
        k += 1
    pad = k // 2
    padded = np.pad(a.astype(np.float64), pad, mode="edge")
    return np.convolve(padded, np.ones(k) / k, mode="valid")[:n].astype(a.dtype)


def _row_extent(cols: np.ndarray, gap_px: int = _ROW_GAP_PX) -> Tuple[int, int]:
    """Edges of the dominant contiguous run in a row (bridging small gaps).

    Robust to stray segmentation pixels far from the rail bed: instead of the
    global min/max (which any speck corrupts), pick the run holding the most
    pixels and return its span.
    """
    if len(cols) == 1:
        return int(cols[0]), int(cols[0])
    splits = np.where(np.diff(cols) > gap_px)[0]
    starts = np.concatenate(([0], splits + 1))
    ends = np.concatenate((splits, [len(cols) - 1]))
    counts = ends - starts + 1
    best = int(np.argmax(counts))
    return int(cols[starts[best]]), int(cols[ends[best]])


def rail_rows(
    mask: np.ndarray,
    min_px: int = 3,
    smooth: int = 21,
    max_w_frac: float = 0.7,
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
    h, w = mask.shape
    ys, ls, rs = [], [], []
    for y in range(h):
        cols = np.where(mask[y] > 0)[0]
        if len(cols) < min_px:
            continue
        l, r = _row_extent(cols)
        if (r - l) > max_w_frac * w:
            continue
        ys.append(y)
        ls.append(l)
        rs.append(r)
    if len(ys) < 4:
        return (
            np.array(ys, dtype=int),
            np.array(ls, dtype=int),
            np.array(rs, dtype=int),
        )
    ys = np.array(ys, dtype=int)
    ls_a = np.array(ls, dtype=int)
    rs_a = np.array(rs, dtype=int)

    # Width must shrink monotonically with distance (rows higher up are
    # narrower). Reject rows whose width deviates grossly from the local
    # median before smoothing — kills single-row blow-ups from merged blobs.
    widths = np.maximum(rs_a - ls_a, 1)
    wmed = _medfilt1d(widths, 9)
    good = np.abs(widths - wmed) <= np.maximum(0.6 * wmed, 8)
    if int(good.sum()) >= 4:
        ys, ls_a, rs_a = ys[good], ls_a[good], rs_a[good]

    ls = _smooth1d(_medfilt1d(ls_a, 9), smooth)
    rs = _smooth1d(_medfilt1d(rs_a, 9), smooth)
    return ys, ls, rs


def zone_bounds(
    ys: np.ndarray,
    ls: np.ndarray,
    rs: np.ndarray,
    w: int,
) -> dict:
    if len(ys) == 0:
        return {}
    widths = np.maximum(rs - ls, 1).astype(float)
    scale = GAUGE_MM / widths
    out = {}
    for name in ("yellow", "orange", "red"):
        off = (ZONE_DISTS_MM[name] / scale).astype(int)
        out[name] = (
            np.clip(ls - off, 0, w - 1),
            np.clip(rs + off, 0, w - 1),
        )
    return out

def _fit_width_line(
    ys: np.ndarray, ls: np.ndarray, rs: np.ndarray
) -> Optional[Tuple[float, float]]:
    """Robust linear fit width(y)=a*y+b via iterative residual trimming."""
    if len(ys) < _FIT_MIN_ROWS:
        return None
    yf = ys.astype(np.float64)
    wf = np.maximum(rs - ls, 1).astype(np.float64)
    try:
        a, b = np.polyfit(yf, wf, 1)
    except (np.linalg.LinAlgError, ValueError):
        return None
    for _ in range(_FIT_ITERS):
        res = wf - (a * yf + b)
        sigma = float(np.std(res))
        if sigma < 1e-6:
            break
        keep = np.abs(res) <= _FIT_TRIM_SIGMA * sigma
        if int(keep.sum()) < max(4, _FIT_MIN_ROWS // 2):
            break
        try:
            a, b = np.polyfit(yf[keep], wf[keep], 1)
        except (np.linalg.LinAlgError, ValueError):
            break
    # Width must grow toward the bottom of the image (a > 0) for a valid model.
    if a <= 1e-6:
        return None
    return float(a), float(b)

class _DistanceModel:
    """Closed-form row<->metric-distance mapping from the fitted width line."""

    __slots__ = ("a", "b", "y_vp", "C")

    def __init__(self, a: float, b: float, f_px: float) -> None:
        self.a = a
        self.b = b
        self.y_vp = -b / a
        self.C = f_px * GAUGE_MM / a  # mm * px

    def width_at(self, y: float) -> float:
        return self.a * float(y) + self.b

    def dist_mm(self, y: float) -> float:
        dy = max(float(y) - self.y_vp, 1e-3)
        return self.C / dy

    def row_at_dist(self, dist_mm: float) -> float:
        return self.y_vp + self.C / max(dist_mm, 1e-3)


def build_distance_model(
    ys: np.ndarray, ls: np.ndarray, rs: np.ndarray, f_px: float
) -> Optional[_DistanceModel]:
    fit = _fit_width_line(ys, ls, rs)
    if fit is None:
        return None
    a, b = fit
    return _DistanceModel(a, b, f_px)

def estimate_distance_m(
    y_px: int,
    ys: np.ndarray,
    ls: np.ndarray,
    rs: np.ndarray,
    f_px: float,
) -> Optional[float]:
    """Ground distance (m) of image row y_px, relative to nearest rail row."""
    if len(ys) < 2:
        return None
    y_bot = float(np.max(ys))
    if y_px >= y_bot:
        return 0.0

    model = build_distance_model(ys, ls, rs, f_px)
    if model is not None:
        d_obj = model.dist_mm(float(y_px))
        d_bot = model.dist_mm(y_bot)
        return max(0.0, float(d_obj - d_bot) / 1000.0)

    widths = np.maximum(rs - ls, 1).astype(float)
    w_at_y = float(np.interp(y_px, ys, widths))
    if w_at_y < 1:
        return None
    d_obj = f_px * GAUGE_MM / w_at_y
    d_bot = f_px * GAUGE_MM / float(widths[int(np.argmax(ys))])
    return max(0.0, (d_obj - d_bot) / 1000.0)

def _grid_positions_fallback(
    ys: np.ndarray, ls: np.ndarray, rs: np.ndarray, f_px: float
) -> list:
    widths = np.maximum(rs - ls, 1).astype(float)
    d_mm = f_px * GAUGE_MM / widths
    d_ref = d_mm[int(np.argmax(ys))]
    marks = []
    for i in range(1, 50):
        target = d_ref + i * GRID_STEP_MM
        above = np.where(d_mm >= target)[0]
        if len(above) == 0:
            break
        best = above[int(np.argmin(np.abs(d_mm[above] - target)))]
        marks.append((int(ys[best]), i * GRID_STEP_MM / 1000.0))
    return marks

def grid_positions(
    ys: np.ndarray,
    ls: np.ndarray,
    rs: np.ndarray,
    f_px: Optional[float] = None,
) -> list:
    """Rows for distance grid lines. Returns (row_px, distance_m) per mark."""
    if len(ys) < 2:
        return []
    if f_px is None:
        f_px = get_focal_px(1920)

    model = build_distance_model(ys, ls, rs, f_px)
    if model is None:
        return _grid_positions_fallback(ys, ls, rs, f_px)

    y_bot = float(np.max(ys))
    y_top = float(np.min(ys))
    d_ref = model.dist_mm(y_bot)

    marks = []
    for i in range(1, 80):
        target = d_ref + i * GRID_STEP_MM
        yy = model.row_at_dist(target)
        # Stop once the mark climbs past the visible rail / approaches horizon.
        if yy <= y_top - 1.0 or yy <= model.y_vp + 1.0:
            break
        if yy >= y_bot:
            continue
        marks.append((int(round(yy)), i * GRID_STEP_MM / 1000.0))
    return marks

def _outline_pts(ys: np.ndarray, xs: np.ndarray) -> np.ndarray:
    """Single boundary polyline (one side) for crisp anti-aliased outlines."""
    return np.column_stack([xs, ys]).astype(np.int32).reshape(-1, 1, 2)

def _put_label(frame: np.ndarray, text: str, org: Tuple[int, int]) -> None:
    """Distance label with a dark outline so it stays readable on any zone."""
    x, y = org
    cv2.putText(frame, text, (x, y), cv2.FONT_HERSHEY_PLAIN, 1,
               (0, 0, 0), 3, cv2.LINE_AA)
    cv2.putText(frame, text, (x, y), cv2.FONT_HERSHEY_PLAIN, 1,
               TEXT_COLOR, 1, cv2.LINE_AA)

def draw_zones(
    frame: np.ndarray,
    ys: np.ndarray,
    zones: dict,
    ls: np.ndarray,
    rs: np.ndarray,
    grid: list,
) -> None:
    """Lightweight rendering: zone boundary lines + rails + distance markup.

    No filled / translucent polygons — only crisp anti-aliased outlines drawn
    directly on the frame, plus the distance grid lines and labels.
    """
    # Zone boundary lines (left + right edge of each zone).
    for name in ("yellow", "orange", "red"):
        if name not in zones:
            continue
        zl, zr = zones[name]
        c = ZONE_COLORS[name]
        cv2.polylines(frame, [_outline_pts(ys, zl)], False, c,
                     _ZONE_LINE_THICK, cv2.LINE_AA)
        cv2.polylines(frame, [_outline_pts(ys, zr)], False, c,
                     _ZONE_LINE_THICK, cv2.LINE_AA)

    # Rail edges.
    cv2.polylines(frame, [_outline_pts(ys, ls)], False, ZONE_COLORS["red"],
                 _RAIL_LINE_THICK, cv2.LINE_AA)
    cv2.polylines(frame, [_outline_pts(ys, rs)], False, ZONE_COLORS["red"],
                 _RAIL_LINE_THICK, cv2.LINE_AA)

    # Distance grid lines + labels across the outer zone.
    outer = zones.get("yellow") or zones.get("orange") or zones.get("red")
    if outer and grid:
        gy_arr = np.fromiter((g[0] for g in grid), dtype=np.int32, count=len(grid))
        ii = np.argmin(np.abs(ys[:, None] - gy_arr[None, :]), axis=0)
        z0, z1 = outer[0], outer[1]
        for k, idx in enumerate(ii):
            gy = int(gy_arr[k])
            dist_m = grid[k][1]
            x1, x2 = int(z0[idx]), int(z1[idx])
            cv2.line(frame, (x1, gy), (x2, gy), GRID_COLOR, 1, cv2.LINE_AA)
            _put_label(frame, f"{dist_m:.0f} m", (x2 + 5, gy + 4))

def draw_legend(frame: np.ndarray) -> None:
    h = frame.shape[0]
    lx, ly = 10, h - 90
    lt = cv2.LINE_8
    for i, name in enumerate(("red", "orange", "yellow")):
        yy = ly + i * 24
        cv2.rectangle(frame, (lx, yy), (lx + 18, yy + 16), ZONE_COLORS[name], -1)
        dist = ZONE_DISTS_MM[name] / 1000
        if name == "red":
            label = f"Red: rails + {dist:.1f} m"
        elif name == "orange":
            label = f"Orange: {dist:.1f} m"
        else:
            label = f"Yellow: {dist:.1f} m"
        cv2.putText(
            frame,
            label,
            (lx + 24, yy + 13),
            cv2.FONT_HERSHEY_PLAIN,
            1,
            TEXT_COLOR,
            1,
            lt,
        )

def find_groups(mask: np.ndarray, min_area: int = 300) -> list:
    n, labels = cv2.connectedComponents(mask)
    groups = []
    for lab in range(1, n):
        comp = (labels == lab).astype(np.uint8) * 255
        if np.sum(comp > 0) < min_area:
            continue
        groups.append(comp)
    return groups if groups else [mask]

def _branch_sort_key(branch: int):
    if branch == 0:
        return lambda c: (-c[1], -c[2])
    return lambda c: (c[1], -c[2])

def _select_rail_group(
    groups: list, h: int, prev_center: Optional[float] = None
) -> np.ndarray:
    if len(groups) == 1:
        return groups[0]
    band = slice(int(0.55 * h), h)
    candidates: list[Tuple[np.ndarray, float, float]] = []
    for g in groups:
        gy, gx = np.where(g > 0)
        if len(gx) == 0:
            continue
        candidates.append((g, float(np.median(gx)), float(np.sum(g[band] > 0))))
    if not candidates:
        return groups[0]

    if prev_center is not None:
        candidates.sort(key=lambda c: abs(c[1] - prev_center))
        if abs(candidates[0][1] - prev_center) <= _MAX_CENTER_JUMP_PX:
            return candidates[0][0]

    candidates.sort(key=_branch_sort_key(RAIL_BRANCH))
    return candidates[0][0]

class _TemporalState:
    EMA_ALPHA = 0.30
    MAX_AREA_RATIO = 1.8
    FALLBACK_FRAMES = 10
    DECAY_SHRINK = 0.97
    TURN_SHIFT_THRESHOLD = 30
    TURN_ALPHA = 0.50

    def __init__(self) -> None:
        self.reset()

    def reset(self) -> None:
        self._ys: Optional[np.ndarray] = None
        self._ls: Optional[np.ndarray] = None
        self._rs: Optional[np.ndarray] = None
        self._area: float = 0.0
        self._miss: int = 0

    @property
    def active(self) -> bool:
        return self._ys is not None and len(self._ys) >= 4

    def center(self) -> Optional[float]:
        if not self.active:
            return None
        return float(np.median((self._ls + self._rs) / 2.0))

    def _area_of(self, ls: np.ndarray, rs: np.ndarray) -> float:
        return float(np.sum(np.maximum(rs - ls, 0)))

    def update(
        self,
        ys: np.ndarray,
        ls: np.ndarray,
        rs: np.ndarray,
        w: int,
    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
        if len(ys) < 4:
            return self._fallback(w)

        area = self._area_of(ls, rs)

        if not self.active:
            self._ys, self._ls, self._rs = ys.copy(), ls.copy(), rs.copy()
            self._area = area
            self._miss = 0
            return ys, ls, rs

        prev_c = self.center()
        cur_c = float(np.median((ls + rs) / 2.0))
        shift = abs(cur_c - prev_c) if prev_c is not None else 0.0

        alpha = self.TURN_ALPHA if shift > self.TURN_SHIFT_THRESHOLD else self.EMA_ALPHA

        if self._area > 0:
            ratio = area / self._area
            if ratio > self.MAX_AREA_RATIO:
                alpha *= 0.2
            elif ratio < 1.0 / self.MAX_AREA_RATIO:
                alpha *= 0.4

        ls_prev = np.interp(ys, self._ys, self._ls).astype(np.float64)
        rs_prev = np.interp(ys, self._ys, self._rs).astype(np.float64)

        ls_new = np.clip(
            (alpha * ls.astype(np.float64) + (1 - alpha) * ls_prev), 0, w - 1
        ).astype(np.int32)
        rs_new = np.clip(
            (alpha * rs.astype(np.float64) + (1 - alpha) * rs_prev), 0, w - 1
        ).astype(np.int32)

        bad = ls_new >= rs_new
        if np.any(bad):
            ls_new[bad] = np.minimum(ls[bad], ls_prev[bad].astype(np.int32))
            rs_new[bad] = np.maximum(rs[bad], rs_prev[bad].astype(np.int32))

        self._ys, self._ls, self._rs = ys.copy(), ls_new.copy(), rs_new.copy()
        self._area = self._area_of(ls_new, rs_new)
        self._miss = 0
        return ys, ls_new, rs_new

    def _fallback(self, w: int) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
        empty = np.array([], dtype=np.int32)
        if not self.active:
            return empty, empty, empty
        self._miss += 1
        if self._miss > self.FALLBACK_FRAMES:
            self.reset()
            return empty, empty, empty
        c = (self._ls + self._rs) / 2.0
        hw = (self._rs - self._ls) / 2.0 * self.DECAY_SHRINK
        self._ls = np.clip((c - hw).astype(np.int32), 0, w - 1)
        self._rs = np.clip((c + hw).astype(np.int32), 0, w - 1)
        self._area = self._area_of(self._ls, self._rs)
        return self._ys.copy(), self._ls.copy(), self._rs.copy()

_temporal = _TemporalState()

def process_frame(frame: np.ndarray, model: YOLO) -> np.ndarray:
    h, w = frame.shape[:2]
    res = model(frame, conf=CONF, device=DEVICE, verbose=False)[0]
    _maybe_save_low_conf_frame(frame, _max_rail_confidence(res))
    full_mask = extract_rail_mask(res, h, w, morph=USE_MORPH)

    if np.any(full_mask):
        groups = find_groups(full_mask)
        gmask = _select_rail_group(groups, h, _temporal.center())
        ys, ls, rs = rail_rows(gmask)
    else:
        ys = np.array([], dtype=np.int32)
        ls = np.array([], dtype=np.int32)
        rs = np.array([], dtype=np.int32)

    ys, ls, rs = _temporal.update(ys, ls, rs, w)

    if len(ys) < 4:
        return frame

    f_px = get_focal_px(w)
    zones = zone_bounds(ys, ls, rs, w)
    grid = grid_positions(ys, ls, rs, f_px)
    draw_zones(frame, ys, zones, ls, rs, grid)
    return frame

def reset_temporal_state() -> None:
    _temporal.reset()

def main() -> None:
    global USE_MORPH
    ap = argparse.ArgumentParser(
        description="Building danger zones around rails"
    )
    ap.add_argument(
        "-m", "--morph", action="store_true", dest="morph",
        help="Enable morphological cleaning of the segmentation mask",
    )
    args = ap.parse_args()
    USE_MORPH = args.morph

    if USE_MORPH:
        print("Morphology enabled: clean rail mask after segmentation")

    model = load_model()
    cap = cv2.VideoCapture(VIDEO_PATH)
    if not cap.isOpened():
        raise OSError(f"Failed to open video: {VIDEO_PATH}")

    fps = int(cap.get(cv2.CAP_PROP_FPS)) or 25
    W = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    H = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

    out = None
    if OUTPUT_PATH:
        Path(OUTPUT_PATH).parent.mkdir(parents=True, exist_ok=True)
        out = cv2.VideoWriter(
            OUTPUT_PATH, cv2.VideoWriter_fourcc(*"mp4v"), fps, (W, H)
        )

    print(f"Video {W}×{H} @ {fps} fps  -  'q' for exit")
    n = 0
    while True:
        ok, frame = cap.read()
        if not ok:
            break
        frame = process_frame(frame, model)
        cv2.imshow("Rail Danger Zones", frame)
        if out:
            out.write(frame)
        if cv2.waitKey(1) & 0xFF == ord("q"):
            break
        n += 1

    cap.release()
    if out:
        out.release()
    cv2.destroyAllWindows()
    print(f"Done - {n} frames processed.")

if __name__ == "__main__":
    main()