| """Estimate fixed per-vehicle lengths using six-column pixel detections and calibration.""" |
| import numpy as np |
| import pandas as pd |
| from _common import numeric, parser, read_csv, require, save_csv, unique |
| from coordinate_transform import load_calibration |
|
|
|
|
| def estimate(df, calibration, minimum_frames=5, iqr_factor=1.5): |
| df = df.copy() |
| numeric(df, ["Frame", "Vehicle_ID"], integer=True) |
| unique(df, ["Frame", "Vehicle_ID"]) |
| cols = ["x_center_px", "y_center_px", "width_px", "height_px"] |
| require(df, cols) |
| if (not np.isfinite([minimum_frames, iqr_factor]).all() or minimum_frames < 5 or |
| int(minimum_frames) != minimum_frames or iqr_factor < 0): |
| raise ValueError("Require integer minimum_frames >= 5 and finite iqr_factor >= 0") |
| data = df[cols].apply(pd.to_numeric, errors="coerce").to_numpy(float) |
| x, y, w, h = data.T |
| valid = np.isfinite(data).all(axis=1) & (w > 0) & (h > 0) |
| length, width = np.full(len(df), np.nan), np.full(len(df), np.nan) |
| indexes = np.flatnonzero(valid) |
| for block in np.array_split(indexes, max(1, int(np.ceil(len(indexes) / 100000)))): |
| if not len(block): |
| continue |
| corners = np.stack([np.column_stack((x[block] + sx * w[block] / 2, |
| y[block] + sy * h[block] / 2)) |
| for sx, sy in [(-1, -1), (1, -1), (1, 1), (-1, 1)]], axis=1) |
| metric = calibration.transform(corners.reshape(-1, 2).astype(np.float32)).reshape(-1, 4, 2) |
| length[block], width[block] = np.ptp(metric, axis=1).T |
| valid &= np.isfinite(length) & np.isfinite(width) & (length > 0) & (width > 0) |
| rows = [] |
| for vehicle, indexes in df.groupby("Vehicle_ID", sort=True).indices.items(): |
| selected = indexes[valid[indexes]] |
| before = len(selected) |
| if before >= 4: |
| q1, q3 = np.quantile(length[selected], [0.25, 0.75]) |
| margin = iqr_factor * (q3 - q1) |
| selected = selected[(length[selected] >= q1 - margin) & (length[selected] <= q3 + margin)] |
| sufficient = len(selected) >= minimum_frames |
| rows.append({"Source_ID": calibration.source, "Vehicle_ID": vehicle, |
| "Vehicle_Length_m": np.quantile(length[selected], 0.25) if sufficient else np.nan, |
| "BBox_Lateral_Extent_Q25_m": np.quantile(width[selected], 0.25) if sufficient else np.nan, |
| "Input_Frames": len(indexes), "Valid_Before_IQR": before, |
| "Retained_Frames": len(selected), "Removed_By_IQR": before - len(selected), |
| "Quality": "good" if len(selected) >= 20 else "limited" if sufficient else "insufficient"}) |
| return pd.DataFrame(rows) |
|
|
|
|
| def main(): |
| p = parser(__doc__) |
| p.add_argument("--input", required=True, help="Six-column raw detection CSV; one source video") |
| p.add_argument("--calibration", required=True, help="Same scene calibration JSON used for centers") |
| p.add_argument("--source", required=True, help="Source video key in calibration JSON") |
| p.add_argument("--output", required=True, help="New per-vehicle metric length and QC CSV") |
| a = p.parse_args() |
| df = read_csv(a.input) |
| if "Source_ID" in df and not df.Source_ID.eq(a.source).all(): |
| raise ValueError("Input must contain only the requested source") |
| save_csv(estimate(df, load_calibration(a.calibration, a.source)), a.output) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
