fix: clip pad ±3s + OSD-clock alignment at cut points (single-frame OCR)

.env.example

@@ -31,6 +31,8 @@ RTSP_URL=rtsp://user:pass@192.168.192.209:8554/camera_stream_640
 # --- Batch clip download (motionEye; dashboard /api/batch-clip only) ---
 MOTIONEYE_URL=http://192.168.192.253:8765   # motionEye base URL for batch clip download; empty disables
 MOTIONEYE_CAMERA_ID=2                          # motionEye camera id (1 = default camera)
+MOTIONEYE_CLIP_PAD=3                           # seconds trimmed before/after the batch window (float)
+MOTIONEYE_OSD_ALIGN=1                          # 1 = OCR-correct OSD clock drift at cut points, 0 = filename-based offsets only
 
 # --- Deprecated: model pipeline mode now lives in config.yaml models.active_mode.
 # Dashboard switches persist there (atomic write, comments preserved) and apply

CHANGELOG.md

@@ -32,6 +32,8 @@ version tags are cut in this repo — POC stage, `main` is the release line).
 - History `Klip` button → office-only `GET /api/batch-clip/<date>/<batch_number>`
   trims+concatenates motionEye recordings to the batch window (`src/clip.py`),
   download `b<n>-<date>-<HH-MM-SS>.mp4`.
+  - `MOTIONEYE_CLIP_PAD` (default ±3 s window padding) + OSD-clock alignment via
+    single-frame OCR at each cut point (`MOTIONEYE_OSD_ALIGN`, default on).
 
 ### Changed
 - OCR engine default **`rapid`** (was `tesseract`) across `config.yaml`,

counter_dashboard.py

@@ -1575,10 +1575,25 @@ def api_batch_clip(date, batch_number):
     except ValueError:
         camera_id = 2
 
+    try:
+        pad_seconds = float((os.getenv("MOTIONEYE_CLIP_PAD") or "").strip())
+    except (TypeError, ValueError):
+        pad_seconds = 3.0
+    _align_raw = os.getenv("MOTIONEYE_OSD_ALIGN")
+    align_osd = (
+        True if _align_raw is None
+        else _align_raw.strip().lower() not in ("0", "false", "no", "")
+    )
+
     if not str(batch_number).isdigit():
         return jsonify({"success": False, "error": f"batch_number tidak valid: {batch_number}"}), 400
     batch_number = int(batch_number)
 
+    print(
+        f"[CLIP] batch={date}/{batch_number} pad={pad_seconds} align={align_osd}",
+        flush=True,
+    )
+
     try:
         conn = get_db()
         cur = conn.cursor()
@@ -1614,7 +1629,10 @@ def api_batch_clip(date, batch_number):
         return jsonify({"success": False, "error": f"Modul klip belum tersedia: {e}"}), 503
 
     try:
-        clip_batch(motion_url, camera_id, str(start_iso), str(end_iso), tmp_path)
+        clip_batch(
+            motion_url, camera_id, str(start_iso), str(end_iso), tmp_path,
+            pad_seconds=pad_seconds, align_osd=align_osd,
+        )
     except ClipError as ce:
         _unlink_quiet(tmp_path)
         msg = str(ce)

docs/configuration.md

@@ -80,6 +80,8 @@ when set.
 | `RTSP_URL` | — | Camera stream URL (env-only, never in YAML) |
 | `MOTIONEYE_URL` | `""` | motionEye base URL for batch clip download (empty disables the endpoint) |
 | `MOTIONEYE_CAMERA_ID` | `2` | motionEye camera id used for batch clip download |
+| `MOTIONEYE_CLIP_PAD` | `3` | Seconds padded before/after the batch window when trimming the clip (float) |
+| `MOTIONEYE_OSD_ALIGN` | `1` | `1` = OCR-correct burned-in OSD clock drift at cut points, `0` = filename-based offsets only |
 | `MODEL_PATH` | — (deprecated) | Single-file v4 override, folded into `models.paths` |
 | `MODEL_MODE` | — (deprecated) | One-run override of `models.active_mode` (warned) |
 | `BATCH_MERGE_THRESHOLD_SECONDS` | `300` | Merge window for adjacent batches |

docs/deployment.md

@@ -103,6 +103,19 @@ under the hood) and returns the file as an attachment.
 - URL keys (`.env`): `MOTIONEYE_URL` (base URL; empty → 404
   `motionEye belum dikonfigurasi (MOTIONEYE_URL)`), `MOTIONEYE_CAMERA_ID`
   (default `2`).
+- `MOTIONEYE_CLIP_PAD` — float seconds padded before/after the batch window
+  at both cut points (default `3` → ±3 s; empty/invalid falls back to `3.0`).
+- `MOTIONEYE_OSD_ALIGN` — OSD-clock alignment, default **on** (unset = on;
+  `0` / `false` / `no` / empty = off). Why: the camera's burned-in OSD clock
+  and the recording clock drift — up to ~+22 s inside a long clip — because
+  of dropped frames, so filename-based offsets cut at the wrong media time.
+  When on, `clip_batch` OCRs **1 frame** at each cut point (~2–6 s extra per
+  clip, result cached per clip+second) and corrects the cut times. `0` →
+  filename-based offsets only. If OCR fails, it silently falls back to
+  filename-based offsets; both cases log one
+  `[CLIP] batch=... pad=... align=...` line at request start (plus any
+  `[CLIP]` alignment detail lines) — check with `journalctl -u
+  karung-counter-dashboard`.
 - motionEye HTTP API used: `GET /movie/<id>/list/` (recording segments in the
   window) and `GET /movie/<id>/download<path>` (segment bytes).
 - Requires `ffmpeg` / `ffprobe` on PATH — both already installed on the Jetson.

src/clip.py

@@ -12,6 +12,12 @@ import urllib.request
 from datetime import datetime, timedelta
 
 _PATH_RE = re.compile(r"(\d{4}-\d{2}-\d{2})/(\d{2})-(\d{2})-(\d{2})\.mp4$")
+# OSD clock burned into pixels: "2026-09-26 08:52:47" (OCR sometimes glues date+time)
+_OSD_RE = re.compile(r"(\d{4}-\d{2}-\d{2})\s*(\d{1,2}:\d{2}:\d{2})")
+_OSD_CACHE = {}  # (url, round(media_offset s)) -> datetime
+_OCR = None      # lazy rapidocr singleton
+_OCR_FAILED = False
+_FRAME_W, _FRAME_H = 960, 540
 _CANDIDATE_BACK = timedelta(minutes=90)
 _OVERLAP_MIN = 0.5
 _NO_CLIP_MSG = (
@@ -130,6 +136,91 @@ def _fmt(seconds: float) -> str:
     return "{:.3f}".format(seconds)
 
 
+def osd_time_at(url: str, media_offset: float, timeout: int = 45):
+    """Extract ONE frame at media_offset, OCR the burned-in OSD clock.
+
+    Returns naive datetime or None (never raises). Lazy rapidocr singleton.
+    """
+    global _OCR, _OCR_FAILED
+    if _OCR_FAILED:
+        return None
+    try:
+        proc = subprocess.run(
+            [
+                "ffmpeg", "-v", "error",
+                "-ss", _fmt(media_offset),
+                "-i", url,
+                "-frames:v", "1",
+                "-s", "{}x{}".format(_FRAME_W, _FRAME_H),
+                "-f", "rawvideo", "-pix_fmt", "bgr24", "-",
+            ],
+            capture_output=True, timeout=timeout,
+        )
+        need = _FRAME_W * _FRAME_H * 3
+        if proc.returncode != 0 or len(proc.stdout) < need:
+            return None
+        import numpy as np  # CI-safe: numpy is a test dep
+
+        frame = np.frombuffer(proc.stdout[:need], np.uint8).reshape(
+            _FRAME_H, _FRAME_W, 3
+        )
+        if _OCR is None:
+            try:
+                from rapidocr_onnxruntime import RapidOCR  # type: ignore
+            except ImportError:
+                _OCR_FAILED = True
+                print("[CLIP] rapidocr_onnxruntime missing — OSD align off "
+                      "(filename-based offsets).", flush=True)
+                return None
+            _OCR = RapidOCR()
+        result, _ = _OCR(frame)
+        text = " ".join(str(item[1]) for item in (result or []))
+        match = _OSD_RE.search(text)
+        if not match:
+            return None
+        return datetime.strptime(
+            "{} {}".format(match.group(1), match.group(2)), "%Y-%m-%d %H:%M:%S"
+        )
+    except Exception:
+        return None
+
+
+def _align_offset(url: str, cur: float, desired: datetime,
+                  duration: float, label: str) -> float:
+    """Correct one cut offset so the OSD at that frame reads `desired`.
+
+    Max 2 probes (frame+OCR each); OCR failure keeps the filename-based
+    offset. Media time drifts vs wall clock inside a clip (dropped frames),
+    so the correction is measured AT the cut point, not assumed constant.
+    """
+    original = cur
+    corrected = False
+    last_osd = None
+    for _ in range(2):
+        key = (url, int(round(cur)))
+        osd = _OSD_CACHE.get(key)
+        if osd is None:
+            osd = osd_time_at(url, cur)
+            if osd is not None:
+                _OSD_CACHE[key] = osd
+        if osd is None:
+            break
+        last_osd = osd
+        delta = (osd - desired).total_seconds()
+        if abs(delta) <= 1.0:
+            break
+        if corrected and osd.date() != desired.date():
+            break  # suspicious day jump — keep the one correction already made
+        new = min(max(cur - delta, 0.0), duration)
+        if new == cur:
+            break
+        cur = new
+        corrected = True
+    print("[CLIP] align {}: {:.1f} -> {:.1f} (osd={}, desired={})".format(
+        label, original, cur, last_osd, desired), flush=True)
+    return cur
+
+
 def _piece_argv(clip_url: str, off0: float, off1: float, piece: str) -> list:
     return [
         "ffmpeg", "-y",
@@ -166,11 +257,23 @@ def _run_ffmpeg(args: list, timeout: int):
 
 
 def clip_batch(motion_url: str, camera_id: int, start_iso: str, end_iso: str,
-               out_path: str, timeout: int = 180) -> None:
+               out_path: str, timeout: int = 180,
-    """List → cover → trim pieces → concat into out_path. Raises ClipError only."""
+               pad_seconds: float = 3.0, align_osd: bool = True) -> None:
+    """List → cover → trim pieces → concat into out_path. Raises ClipError only.
+
+    pad_seconds widens the target window (trim a little before/after the
+    batch). align_osd OCR-corrects the burned-in OSD clock at the cut
+    points (media-vs-wall drift; ~1-2 s per probe, cached).
+    """
     try:
         start = datetime.fromisoformat(start_iso)
         end = datetime.fromisoformat(end_iso)
+        try:
+            pad = float(pad_seconds)
+        except (TypeError, ValueError):
+            pad = 3.0
+        window_start = start - timedelta(seconds=pad)
+        window_end = end + timedelta(seconds=pad)
         out_path = os.path.abspath(out_path)
         os.makedirs(os.path.dirname(out_path), exist_ok=True)
         tmpdir = None
@@ -180,20 +283,29 @@ def clip_batch(motion_url: str, camera_id: int, start_iso: str, end_iso: str,
             def probe(path):
                 return probe_duration(_download_url(motion_url, camera_id, path))
 
-            clips = find_clips(movies, start, end, probe)
+            clips = find_clips(movies, window_start, window_end, probe)
             if not clips:
                 raise ClipError(_NO_CLIP_MSG)
             tmpdir = tempfile.mkdtemp(dir=os.path.dirname(out_path))
             pieces = []
             for index, clip in enumerate(clips):
                 piece = os.path.join(tmpdir, "piece_{:03d}.mp4".format(index))
+                clip_url = _download_url(motion_url, camera_id, clip["path"])
+                off0, off1 = clip["off0"], clip["off1"]
+                if align_osd:
+                    orig0, orig1 = off0, off1
+                    if index == 0:
+                        off0 = _align_offset(clip_url, off0, window_start,
+                                             clip["duration"], "start")
+                    if index == len(clips) - 1:
+                        off1 = _align_offset(clip_url, off1, window_end,
+                                             clip["duration"], "end")
+                    if off1 - off0 < 1.0:
+                        off0, off1 = orig0, orig1
+                        print("[CLIP] align rejected (span < 1s); "
+                              "filename-based offsets kept.", flush=True)
                 _run_ffmpeg(
-                    _piece_argv(
+                    _piece_argv(clip_url, off0, off1, piece),
-                        _download_url(motion_url, camera_id, clip["path"]),
-                        clip["off0"],
-                        clip["off1"],
-                        piece,
-                    ),
                     timeout,
                 )
                 pieces.append(piece)

tests/test_clip.py

@@ -201,7 +201,8 @@ def test_clip_batch_single_piece_argv(monkeypatch, tmp_path):
     calls, captured = [], {}
     monkeypatch.setattr(clip, "_run_ffmpeg", fake_run_factory(calls, captured))
     out = tmp_path / "sub" / "batch.mp4"
-    clip.clip_batch(BASE, 1, "2026-09-26T09:00:10", "2026-09-26T09:00:50", str(out))
+    clip.clip_batch(BASE, 1, "2026-09-26T09:00:10", "2026-09-26T09:00:50", str(out),
+                    pad_seconds=0, align_osd=False)
     assert len(calls) == 1
     args, timeout = calls[0]
     assert timeout == 180
@@ -229,7 +230,8 @@ def test_clip_batch_two_pieces_trim_and_concat(monkeypatch, tmp_path):
     calls, captured = [], {}
     monkeypatch.setattr(clip, "_run_ffmpeg", fake_run_factory(calls, captured))
     out = tmp_path / "batch.mp4"
-    clip.clip_batch(BASE, 1, START_ISO, END_ISO, str(out), timeout=60)
+    clip.clip_batch(BASE, 1, START_ISO, END_ISO, str(out), timeout=60,
+                    pad_seconds=0, align_osd=False)
     assert len(calls) == 3
     (piece_a, timeout_a), (piece_b, _), (concat_args, timeout_b) = calls
     assert timeout_a == timeout_b == 60
@@ -273,3 +275,154 @@ def test_clip_batch_bad_iso_raises_cliperror(tmp_path):
     with pytest.raises(ClipError):
         clip.clip_batch(BASE, 1, "bukan-tanggal", "2026-09-26T10:00:00",
                         str(tmp_path / "out.mp4"))
+
+
+# ------------------------------------------------------- pad + OSD alignment
+
+@pytest.fixture(autouse=True)
+def _clear_osd_cache():
+    clip._OSD_CACHE.clear()
+    yield
+    clip._OSD_CACHE.clear()
+
+
+def _one_movie(monkeypatch, start_iso, duration):
+    movies = [{"path": CLIP_A, "start": dt("2026-09-26 09:00:00")}]
+    patch_batch_env(monkeypatch, movies, duration=duration)
+
+
+def test_pad_extends_selection_window(monkeypatch, tmp_path):
+    # clip 09:00:00 + 5 s ends exactly at batch start → pad 0 finds nothing
+    movies = [{"path": CLIP_A, "start": dt("2026-09-26 09:00:00")}]
+    patch_batch_env(monkeypatch, movies, duration=5.0)
+    calls, captured = [], {}
+    monkeypatch.setattr(clip, "_run_ffmpeg", fake_run_factory(calls, captured))
+    with pytest.raises(ClipError, match="Tidak ada rekaman"):
+        clip.clip_batch(BASE, 1, "2026-09-26T09:00:05", "2026-09-26T09:00:10",
+                        str(tmp_path / "a.mp4"), pad_seconds=0, align_osd=False)
+    clip.clip_batch(BASE, 1, "2026-09-26T09:00:05", "2026-09-26T09:00:10",
+                    str(tmp_path / "b.mp4"), pad_seconds=3, align_osd=False)
+    args = calls[0][0]
+    assert args[args.index("-ss") + 1] == "2.000"   # window_start 09:00:02
+    assert args[args.index("-to") + 1] == "5.000"   # clip ends 09:00:05
+
+
+def test_osd_alignment_corrects_start_offset(monkeypatch, tmp_path):
+    _one_movie(monkeypatch, None, 600.0)
+    calls, captured = [], {}
+    monkeypatch.setattr(clip, "_run_ffmpeg", fake_run_factory(calls, captured))
+    # batch 09:01:10→09:02:00 (pad 0): filename off0=70, off1=120.
+    # OSD at media 70 reads desired+14 → shift to 56; at 56 reads desired.
+    reads = {
+        70: dt("2026-09-26 09:01:24"),
+        56: dt("2026-09-26 09:01:10"),
+        120: dt("2026-09-26 09:02:00"),
+    }
+
+    def fake_osd(url, media_offset, timeout=45):
+        return reads.get(int(round(media_offset)))
+
+    monkeypatch.setattr(clip, "osd_time_at", fake_osd)
+    out = tmp_path / "out.mp4"
+    clip.clip_batch(BASE, 1, "2026-09-26T09:01:10", "2026-09-26T09:02:00",
+                    str(out), pad_seconds=0, align_osd=True)
+    args = calls[0][0]
+    assert args[args.index("-ss") + 1] == "56.000"
+    assert args[args.index("-to") + 1] == "120.000"
+
+
+def test_align_disabled_never_probes_osd(monkeypatch, tmp_path):
+    _one_movie(monkeypatch, None, 600.0)
+    calls, captured = [], {}
+    monkeypatch.setattr(clip, "_run_ffmpeg", fake_run_factory(calls, captured))
+
+    def boom(*a, **k):
+        raise AssertionError("osd_time_at must not be called")
+
+    monkeypatch.setattr(clip, "osd_time_at", boom)
+    clip.clip_batch(BASE, 1, "2026-09-26T09:00:10", "2026-09-26T09:00:50",
+                    str(tmp_path / "o.mp4"), pad_seconds=0, align_osd=False)
+    args = calls[0][0]
+    assert args[args.index("-ss") + 1] == "10.000"
+
+
+def test_osd_failure_keeps_filename_offsets(monkeypatch, tmp_path):
+    _one_movie(monkeypatch, None, 600.0)
+    calls, captured = [], {}
+    monkeypatch.setattr(clip, "_run_ffmpeg", fake_run_factory(calls, captured))
+    monkeypatch.setattr(clip, "osd_time_at", lambda *a, **k: None)
+    clip.clip_batch(BASE, 1, "2026-09-26T09:00:10", "2026-09-26T09:00:50",
+                    str(tmp_path / "o.mp4"), pad_seconds=0, align_osd=True)
+    args = calls[0][0]
+    assert args[args.index("-ss") + 1] == "10.000"
+    assert args[args.index("-to") + 1] == "50.000"
+
+
+def test_osd_alignment_probe_iteration_cap(monkeypatch, tmp_path):
+    _one_movie(monkeypatch, None, 600.0)
+    calls, captured = [], {}
+    monkeypatch.setattr(clip, "_run_ffmpeg", fake_run_factory(calls, captured))
+    probes = []
+
+    def always_off(url, media_offset, timeout=45):
+        probes.append(int(round(media_offset)))
+        return dt("2026-09-26 09:01:24")  # always +14 vs desired 09:01:10
+
+    monkeypatch.setattr(clip, "osd_time_at", always_off)
+    clip.clip_batch(BASE, 1, "2026-09-26T09:01:10", "2026-09-26T09:02:00",
+                    str(tmp_path / "o.mp4"), pad_seconds=0, align_osd=True)
+    # start align: at most 2 probes; end align: 2 more (same cap)
+    assert len(probes) <= 4
+    start_probes = [p for p in probes if p in (70, 56, 42)]
+    assert len(start_probes) <= 2
+    args = calls[0][0]
+    # after two corrections 70 -> 56 -> 42
+    assert args[args.index("-ss") + 1] == "42.000"
+
+
+def test_osd_date_mismatch_corrects_once_then_stops(monkeypatch, tmp_path):
+    # batch spans New Year's Eve; OSD read lands on the wrong calendar day
+    movies = [{"path": "/2026-12-31/23-00-00.mp4",
+               "start": dt("2026-12-31 23:00:00")}]
+    monkeypatch.setattr(clip, "list_movies", lambda *a, **k: list(movies))
+    monkeypatch.setattr(clip, "probe_duration", lambda url, timeout=30: 7200.0)
+    calls, captured = [], {}
+    monkeypatch.setattr(clip, "_run_ffmpeg", fake_run_factory(calls, captured))
+    probes = []
+
+    def wrong_day(url, media_offset, timeout=45):
+        off = int(round(media_offset))
+        probes.append(off)
+        if off == 1924:
+            return dt("2027-01-01 00:32:04")   # +1 h vs desired Dec31 23:32:04
+        if off == 0:
+            return dt("2027-01-01 00:31:50")   # still Jan 1 after correction
+        return None                            # end-offset probes do nothing
+
+    monkeypatch.setattr(clip, "osd_time_at", wrong_day)
+    clip.clip_batch(BASE, 1, "2026-12-31T23:32:04", "2026-12-31T23:35:00",
+                    str(tmp_path / "o.mp4"), pad_seconds=0, align_osd=True)
+    # start align: probe at 1924 → corrects (clamped to 0), probe at 0 has
+    # wrong date again → rejected by date check (2 probes max).
+    assert probes[:2] == [1924, 0]
+    args = calls[0][0]
+    assert args[args.index("-ss") + 1] == "0.000"
+    assert args[args.index("-to") + 1] == "2100.000"  # end offset untouched
+
+
+def test_alignment_span_collapse_falls_back(monkeypatch, tmp_path):
+    _one_movie(monkeypatch, None, 600.0)
+    calls, captured = [], {}
+    monkeypatch.setattr(clip, "_run_ffmpeg", fake_run_factory(calls, captured))
+
+    def collapse(url, media_offset, timeout=45):
+        # readings that would push off1 before off0
+        return dt("2026-09-26 09:02:59")  # end desired 09:00:50 → huge shift
+
+    monkeypatch.setattr(clip, "osd_time_at", collapse)
+    clip.clip_batch(BASE, 1, "2026-09-26T09:00:10", "2026-09-26T09:00:50",
+                    str(tmp_path / "o.mp4"), pad_seconds=0, align_osd=True)
+    args = calls[0][0]
+    # fallback keeps filename-based offsets
+    assert args[args.index("-ss") + 1] == "10.000"
+    assert args[args.index("-to") + 1] == "50.000"