video_processor.py

"""
video_processor.py
==================
Processes a folder of dashcam videos:
  1. Splits each 20-second video into 3-second clips
  2. Runs the LK optical flow pipeline on each clip
  3. Encodes each clip as a 53-dim trajectory embedding
  4. Saves everything to a .pkl file

Usage:
    python video_processor.py \
        --video_dir /path/to/videos \
        --output_pkl trajectory_db.pkl \
        --max_videos 10
"""

import os
import cv2
import pickle
import argparse
import numpy as np
from pathlib import Path
from tqdm import tqdm
from typing import Dict, List, Optional

from trajectory_extractor import (
    extract_clip_frames,
    compute_lk_flow_clip,
    encode_trajectory,
    CLIP_FRAMES,
    FPS,
    CLIP_SEC,
)


# ─────────────────────────────────────────────────────────────
# Helpers
# ─────────────────────────────────────────────────────────────
SUPPORTED_EXTS = {".mp4", ".avi", ".mov", ".mkv", ".MP4", ".AVI", ".MOV"}


# def get_video_files(video_dir: str, max_videos: Optional[int] = None) -> List[str]:
#     """Return sorted list of video file paths from directory."""
#     video_dir = Path(video_dir)
#     files = [
#         str(f) for f in sorted(video_dir.iterdir())
#         if f.suffix in SUPPORTED_EXTS
#     ]
#     if max_videos is not None:
#         files = files[:max_videos]
#     return files
def get_video_files(video_dir: str, max_videos: Optional[int] = None) -> List[str]:
    """Return sorted list of video file paths from directory and all subfolders."""
    video_dir = Path(video_dir)
    files = sorted([
        str(f) for f in video_dir.rglob("*")
        if f.suffix in SUPPORTED_EXTS
    ])
    if max_videos is not None:
        files = files[:max_videos]
    return files

def get_video_total_frames(video_path: str) -> int:
    """Return total frame count of a video."""
    cap = cv2.VideoCapture(video_path)
    total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    cap.release()
    return total


def get_clip_boundaries(total_frames: int, clip_frames: int = CLIP_FRAMES) -> List[tuple]:
    """
    Split video into non-overlapping 3-second clips.
    Returns list of (start_frame, end_frame) tuples.
    Last partial clip is discarded if < 1.5 seconds.
    """
    boundaries = []
    start = 0
    while start + clip_frames <= total_frames:
        boundaries.append((start, start + clip_frames))
        start += clip_frames

    # Include last partial clip only if it's at least half a clip
    remaining = total_frames - start
    if remaining >= clip_frames // 2:
        boundaries.append((start, total_frames))

    return boundaries


# ─────────────────────────────────────────────────────────────
# Per-clip Processing
# ─────────────────────────────────────────────────────────────
def process_clip(video_path: str, start_frame: int, end_frame: int, clip_id: str) -> Optional[Dict]:
    """
    Process a single clip and return its database entry.

    Returns None if clip is too short or processing fails.
    """
    try:
        frames = extract_clip_frames(video_path, start_frame, end_frame)

        if len(frames) < 10:
            return None

        # Run LK optical flow pipeline
        flow_result = compute_lk_flow_clip(frames)

        lateral_signals = flow_result["lateral_signals"]
        forward_signals = flow_result["forward_signals"]

        if len(lateral_signals) < 5:
            return None

        # Encode to 53-dim embedding
        encoding = encode_trajectory(lateral_signals, forward_signals)

        start_sec = start_frame / FPS
        end_sec   = end_frame   / FPS

        return {
            "clip_id"        : clip_id,
            "video_path"     : video_path,
            "video_name"     : Path(video_path).name,
            "start_frame"    : start_frame,
            "end_frame"      : end_frame,
            "start_sec"      : round(start_sec, 2),
            "end_sec"        : round(end_sec,   2),
            "embedding"      : encoding["embedding"],
            "turn_ratio"     : encoding["turn_ratio"],
            "peak_lateral"   : encoding["peak_lateral"],
            "direction"      : encoding["direction"],
            "trajectory_raw" : encoding["trajectory_raw"],
            "lateral_signals": encoding["lateral_signals"],
            "n_frames"       : encoding["n_frames"],
            "n_features_avg" : float(np.mean(flow_result["n_features"])) if flow_result["n_features"] else 0.0,
        }

    except Exception as e:
        print(f"  [ERROR] clip {clip_id}: {e}")
        return None


# ─────────────────────────────────────────────────────────────
# Main Processing Loop
# ─────────────────────────────────────────────────────────────
def process_video_folder(
    video_dir: str,
    output_pkl: str,
    max_videos: Optional[int] = None,
    verbose: bool = True,
) -> Dict:
    """
    Process all videos in a folder and save trajectory database to pkl.

    Returns the database dict.
    """
    video_files = get_video_files(video_dir, max_videos)

    if not video_files:
        raise FileNotFoundError(f"No video files found in: {video_dir}")

    print(f"\n{'='*60}")
    print(f"Trajectory Extraction Pipeline")
    print(f"{'='*60}")
    print(f"Video directory : {video_dir}")
    print(f"Videos to process: {len(video_files)}")
    print(f"Clip length     : {CLIP_SEC}s ({CLIP_FRAMES} frames)")
    print(f"Output pkl      : {output_pkl}")
    print(f"{'='*60}\n")

    database = {}       # clip_id → entry dict
    stats = {
        "total_clips"   : 0,
        "failed_clips"  : 0,
        "left_clips"    : 0,
        "right_clips"   : 0,
        "straight_clips": 0,
    }

    for vid_idx, vpath in enumerate(video_files):
        vname = Path(vpath).stem
        total_frames = get_video_total_frames(vpath)
        clips = get_clip_boundaries(total_frames)

        print(f"[{vid_idx+1}/{len(video_files)}] {Path(vpath).name}")
        print(f"  Total frames: {total_frames}  |  Clips: {len(clips)}")

        for clip_idx, (start_f, end_f) in enumerate(
            tqdm(clips, desc=f"  Clips", leave=False, disable=not verbose)
        ):
            clip_id = f"{vname}__clip_{clip_idx:03d}"

            entry = process_clip(vpath, start_f, end_f, clip_id)

            if entry is None:
                stats["failed_clips"] += 1
                continue

            database[clip_id] = entry
            stats["total_clips"] += 1
            stats[f"{entry['direction']}_clips"] += 1

            if verbose:
                tqdm.write(
                    f"  [{clip_idx}] {entry['start_sec']:.1f}s–{entry['end_sec']:.1f}s"
                    f" | {entry['direction']:8s}"
                    f" | turn_ratio={entry['turn_ratio']:+.3f}"
                    f" | features≈{entry['n_features_avg']:.0f}"
                )

        print()

    # ── Save to pkl ───────────────────────────────────────────
    os.makedirs(os.path.dirname(os.path.abspath(output_pkl)), exist_ok=True)
    with open(output_pkl, "wb") as f:
        pickle.dump(database, f, protocol=pickle.HIGHEST_PROTOCOL)

    # ── Summary ───────────────────────────────────────────────
    print(f"\n{'='*60}")
    print(f"Processing Complete")
    print(f"{'='*60}")
    print(f"Total clips saved : {stats['total_clips']}")
    print(f"Failed clips      : {stats['failed_clips']}")
    print(f"Left turns        : {stats['left_clips']}")
    print(f"Right turns       : {stats['right_clips']}")
    print(f"Straight          : {stats['straight_clips']}")
    print(f"Saved to          : {output_pkl}")
    print(f"{'='*60}\n")

    return database


# ─────────────────────────────────────────────────────────────
# CLI Entry Point
# ─────────────────────────────────────────────────────────────
if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Trajectory extraction pipeline")
    parser.add_argument(
        "--video_dir",
        type=str,
        default="/media/RTCIN15TBD/AllUsers/sjl3kor/trajectory_retrieval_system/data/videos",
        help="Path to folder containing dashcam videos",
    )
    parser.add_argument(
        "--output_pkl",
        type=str,
        default="trajectory_db.pkl",
        help="Output pickle file path",
    )
    parser.add_argument(
        "--max_videos",
        type=int,
        default=10,
        help="Max number of videos to process (use None for all)",
    )
    args = parser.parse_args()

    process_video_folder(
        video_dir   = args.video_dir,
        output_pkl  = args.output_pkl,
        max_videos  = args.max_videos,
    )