awe/radar_system
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
eacea436a42186d56a87df11f16130c3478a0ce3
radar_system/Horns_clean.py
T
Ayzen 0da8b1283c working version
2026-05-18 17:27:00 +03:00

1087 lines
43 KiB
Python
Raw Blame History

"""
MIMO GPR — coherent time-domain BackProjection
=================================================
Эта ячейка полностью независима от верхнего эллипсного алгоритма:
1. загружает те же S21-данные;
2. строит oversampled A-сканы через тот же частотный сдвиг перед IFFT;
3. для каждой точки (x,z) вычисляет tau_ij = (Rtx + Rrx) / v;
4. интерполирует комплексный A-скан в этой задержке;
5. когерентно суммирует комплексные вклады всех Tx/Rx-пар с компенсацией geo·pattern.
Это coherent BP: суммируются комплексные h_ij(tau), затем строится |sum h_ij|.
"""
import numpy as np
import matplotlib.pyplot as plt
from scipy.ndimage import gaussian_filter, label
from pathlib import Path
import contextlib
import io
# ══════════════════════════════════════════════════════
# 0.1 ИЗМЕНЯЕМЫЕ ПАРАМЕТРЫ
# ══════════════════════════════════════════════════════
INPUT_IDX = [0, 1, 2, 3]
OUTPUT_IDX = [2, 3]
# Частотный диапазон и глубинный gate.
F_START = 7 * 1e8
F_STOP = 6 * 1e9
MIN_DEPTH = 4.0
MAX_DEPTH = 12.0
# Данные и вычитание среднего фона.
BG_SUBTRACT = True
BG_PATH = Path('/Users/ivan_root/Downloads/Telegram_dwnld/20260514/20260514_s1_10-6000_1201_50k/preprocessed')
DATA_PATH = Path('/Users/ivan_root/Downloads/Telegram_dwnld/20260514/20260514_s1_10-6000_1201_50k/preprocessed/0004_id1_ns11951298408967')
# Убрать паразитные боковые лепестки с 2D карты.
BP_REMOVE_SIDELOBE_OBJECTS = True # True: убрать SL-кандидаты из финальной таблицы и разметки
BP_MAX_DETECTED_OBJECTS_TO_DRAW = 8 # N: если найдено больше объектов, цели на карте не рисуются
BP_DRAW_TOP_M_OBJECTS = 4 # M: если найдено <= N, рисуются только первые M по текущему score
BP_OBJECT_MIN_FRAC = 0.7 # остановка: пик ниже этой доли от глобального максимума
# Компенсация затухания разделена на 2 части:
# range: геометрическое расхождение 1/(Rtx*Rrx)
# angle: диаграмма направленности cos_tx^2 * cos_rx^2
COMP_RANGE_POWER = 0.1
# Метрика для ранжирования целей.
# 'peak' — порядок по максимумам coherent BP;
# 'combined' — coherent peak + coherence factor + prominence + contrast.
BP_SCORE_MODE = 'combined'
# ══════════════════════════════════════════════════════
# 0.2 КОНФИГИ АНТЕНН
# ══════════════════════════════════════════════════════
# Физические координаты антенн по их реальным индексам, [м].
# Формат: (x, y, z), где X - поперечная ось, Z - дальность вдоль оси радара,
# Y - нормаль к плоскости XOZ. BP-карта строится в плоскости y=BP_PLANE_Y.
# Для индивидуальных смещений Tx меняйте второй/третий элемент tuple у каждой Tx.
BP_PLANE_Y = 0.0
TX_POSITIONS = {
2: (-128.7 * 0.01, 0.375, 0.0),
3: ( 130.7 * 0.01, 0.375, 0.0),
}
RX_POSITIONS = {
0: ( 25.8 * 0.01, 0.0, 0.0),
1: ( 75.0 * 0.01, 0.0, 0.0),
2: (-75.0 * 0.01, 0.0, 0.0),
3: (-24.5 * 0.01, 0.0, 0.0),
}
# ══════════════════════════════════════════════════════
# 0.3 НЕИЗМЕНЯЕМЫЕ ПАРАМЕТРЫ (ЛУЧШЕ НЕ ТРОГАТЬ)
# ══════════════════════════════════════════════════════
eps_r = 1.0
v = 3e8 / np.sqrt(eps_r)
def positions_to_xyz(position_dict):
coords = []
for idx in sorted(position_dict):
pos = np.asarray(position_dict[idx], dtype=float)
if pos.ndim == 0:
pos = np.array([float(pos), 0.0, 0.0], dtype=float)
elif pos.shape == (2,):
pos = np.array([pos[0], pos[1], 0.0], dtype=float)
elif pos.shape != (3,):
raise ValueError(f'Позиция антенны {idx} должна быть x, (x,y) или (x,y,z), получено {pos}')
coords.append(pos)
return np.vstack(coords)
tx_xyz = positions_to_xyz(TX_POSITIONS)
rx_xyz = positions_to_xyz(RX_POSITIONS)
x_tx, y_tx, z_tx = tx_xyz.T
x_rx, y_rx, z_rx = rx_xyz.T
# Сетка BP-карты.
x_ant = np.concatenate([x_tx, x_rx])
x_min, x_max = x_ant.min() - 2.0, x_ant.max() + 2.0
z_min, z_max = 0.1, MAX_DEPTH
NX_BP = 300
NZ_BP = 300
# Oversampling A-сканов: повышает плотность точек по t, но не физическое разрешение.
BP_OVERSAMPLE = 8
BP_WINDOW = True
# Нормировка каналов Tx/Rx перед когерентным сложением.
# Масштаб считается по |complex A-scan| в выбранном диапазоне глубин.
PAIR_NORMALIZE = True
PAIR_NORM_PERCENTILE = 50.0
PAIR_NORM_DEPTH_MIN = MIN_DEPTH
PAIR_NORM_DEPTH_MAX = MAX_DEPTH
PAIR_NORM_EPS = 1e-15
# Компенсация нормируется на точку под виртуальным центром пары на глубине COMP_REF_DEPTH.
COMP_ANGLE_POWER = 0.0
COMP_RANGE_WEIGHT_MAX = 5.0
COMP_ANGLE_WEIGHT_MAX = 2.0
COMP_WEIGHT_MAX = 8.0
COMP_REF_DEPTH = 5.0
BP_VALIDATE_COMPENSATION = False # clean mode: не считаем отдельную карту без compensation
# Сглаживание только для удобства поиска/визуализации максимума.
BP_SMOOTH_SIGMA = 1.5
# Параметры поиска объектов на BP-карте.
MAX_OBJECTS = 10
BP_ONLY_MAP_OUTPUT = True # True: подавить текстовые выводы и оставить только BP-карту
BP_REGION_THRESH_FRAC = 0.75 # область объекта: связная область выше этой доли от локального пика
BP_SUPPRESS_THRESH_FRAC = 0.2 # подавление: более широкая связная область вокруг найденного пика
BP_SUPPRESS_USE_WINDOW = True # False: подавлять всю связанную область; True: ограничить окно вокруг пика
BP_SUPPRESS_RX_CM = 80.0 # используется только если BP_SUPPRESS_USE_WINDOW=True
BP_SUPPRESS_RZ_CM = 40.0 # используется только если BP_SUPPRESS_USE_WINDOW=True
BP_MIN_REGION_AREA_CM2 = 10.0 # отсечение совсем мелких шумовых пятен
# Компактный центроид вокруг локального максимума, устойчивее центроида всей вытянутой области.
BP_CENTER_USE_COMPACT = True
BP_CENTER_RX_CM = 60.0
BP_CENTER_RZ_CM = 25.0
BP_CENTER_THRESH_FRAC = 0.88
BP_CENTER_WEIGHT_POWER = 2.0
# Диагностика боковых лепестков coherent BP.
BP_SIDELOBE_DETECT = True
BP_SIDELOBE_RANGE_RMS_TOL_CM = 20.0 # RMS-разница бистатических глубин по всем парам
BP_SIDELOBE_MIN_DX_CM = 35.0 # боковой лепесток должен быть заметно смещен по X
BP_SIDELOBE_MAX_DZ_CM = 70.0 # но находиться примерно на той же глубине
BP_SIDELOBE_MAX_REL_PEAK = 0.85 # кандидат должен быть слабее родительского максимума
# Фазовая метрика внутри compact-области объекта.
BP_PHASE_WEIGHT_POWER = 1.0
# Локальная выраженность объекта над окружающим фоном.
BP_LOCAL_BG_RX_CM = 120.0
BP_LOCAL_BG_RZ_CM = 80.0
BP_LOCAL_BG_PERCENTILE = 50.0
BP_LOCAL_CONTRAST_EPS = 1e-12
# Экспериментальный score для ранжирования найденных объектов.
BP_SCORE_COMPUTE_INCOHERENT = True
BP_SCORE_COH_PEAK_WEIGHT = 0.45
BP_SCORE_COHERENCE_FACTOR_WEIGHT = 0.25
BP_SCORE_PROMINENCE_WEIGHT = 0.20
BP_SCORE_CONTRAST_WEIGHT = 0.10
BP_SCORE_CONTRAST_CAP = 6.0
BP_SCORE_CF_EPS = 1e-12
if BP_ONLY_MAP_OUTPUT:
_bp_stdout_buffer = io.StringIO()
_bp_stdout_redirect = contextlib.redirect_stdout(_bp_stdout_buffer)
_bp_stdout_redirect.__enter__()
# ══════════════════════════════════════════════════════
# 1. ЗАГРУЗКА ДАННЫХ
# ══════════════════════════════════════════════════════
def load_mimo_data(data_path, input_idx, output_idx):
data_path = Path(data_path)
s21_data = {}
freq_data = {}
for f in data_path.glob('i*_o*_s21.npy'):
name = f.stem
parts = name.split('_')
i_tx_phys = int(parts[1][1:])
i_rx_phys = int(parts[0][1:])
if i_tx_phys not in output_idx or i_rx_phys not in input_idx:
continue
i_tx = sorted(output_idx).index(i_tx_phys)
i_rx = sorted(input_idx).index(i_rx_phys)
s21_data[(i_tx, i_rx)] = np.load(f)
freq_file = data_path / f'i{i_rx_phys}_o{i_tx_phys}_freq.npy'
freq_data[(i_tx, i_rx)] = np.load(freq_file)
return s21_data, freq_data
def compute_background(bg_path, input_idx, output_idx):
bg_path = Path(bg_path)
snapshots = [s for s in sorted(bg_path.glob('*/')) if s.is_dir()]
if len(snapshots) == 0:
print('Фоновые снимки не найдены, BG_SUBTRACT отключен.')
return None
print(f'Вычисление фона по {len(snapshots)} снимкам...', end=' ', flush=True)
bg_sum = {}
bg_count = {}
for snap_dir in snapshots:
for f in snap_dir.glob('i*_o*_s21.npy'):
name = f.stem
parts = name.split('_')
i_tx_phys = int(parts[1][1:])
i_rx_phys = int(parts[0][1:])
if i_tx_phys not in output_idx or i_rx_phys not in input_idx:
continue
i_tx = sorted(output_idx).index(i_tx_phys)
i_rx = sorted(input_idx).index(i_rx_phys)
key = (i_tx, i_rx)
s21 = np.load(f)
if key not in bg_sum:
bg_sum[key] = np.zeros_like(s21, dtype=np.complex128)
bg_count[key] = 0
bg_sum[key] += s21
bg_count[key] += 1
bg = {key: bg_sum[key] / bg_count[key] for key in bg_sum}
print(f'готово. Пар: {len(bg)}')
return bg
s21_data, freq_data = load_mimo_data(DATA_PATH, INPUT_IDX, OUTPUT_IDX)
N_tx = len(OUTPUT_IDX)
N_rx = len(INPUT_IDX)
N_pairs = len(s21_data)
assert len(x_tx) == N_tx, f'x_tx должен содержать {N_tx} элементов'
assert len(x_rx) == N_rx, f'x_rx должен содержать {N_rx} элементов'
assert N_pairs > 0, 'Не найдено ни одной Tx/Rx-пары. Проверьте DATA_PATH.'
print('Геометрия антенн, [м]:')
for local_i, phys_i in enumerate(sorted(TX_POSITIONS)):
print(f' Tx{local_i} / o{phys_i}: x={x_tx[local_i]:+.3f}, y={y_tx[local_i]:+.3f}, z={z_tx[local_i]:+.3f}')
for local_j, phys_j in enumerate(sorted(RX_POSITIONS)):
print(f' Rx{local_j} / i{phys_j}: x={x_rx[local_j]:+.3f}, y={y_rx[local_j]:+.3f}, z={z_rx[local_j]:+.3f}')
if BG_SUBTRACT:
background = compute_background(BG_PATH, INPUT_IDX, OUTPUT_IDX)
if background is None:
BG_SUBTRACT = False
else:
background = None
print('BG_SUBTRACT = False, вычитание фона отключено.')
first_key = list(freq_data.keys())[0]
freqs = freq_data[first_key]
freq_mask = (freqs >= F_START) & (freqs <= F_STOP)
freqs_bp = freqs[freq_mask]
if len(freqs_bp) < 2:
raise ValueError('В выбранном частотном диапазоне меньше двух точек.')
f_min = float(freqs_bp[0])
f_max = float(freqs_bp[-1])
BW = f_max - f_min
df_values = np.diff(freqs_bp)
df_median = float(np.median(df_values))
df_min = float(np.min(df_values))
df_max = float(np.max(df_values))
df_rel_spread = (df_max - df_min) / (df_median + 1e-30)
range_resolution = v / (2 * BW)
unambiguous_depth = v / (2 * df_median)
unambiguous_total_path = v / df_median
print(f'Загружено пар Tx/Rx: {N_pairs}')
print(f'Частотный диапазон BP: {f_min/1e9:.3f} - {f_max/1e9:.3f} ГГц')
print(f'Точек частоты в BP-диапазоне: {len(freqs_bp)}')
print(f'Шаг частоты df: median={df_median/1e6:.3f} МГц, '
f'min={df_min/1e6:.3f} МГц, max={df_max/1e6:.3f} МГц')
print(f'Неравномерность df: {(df_rel_spread*100):.3f}% от median')
print(f'Полоса B: {BW/1e9:.3f} ГГц')
print(f'Теоретический предел разрешения по глубине deltaZ = {range_resolution*100:.2f} см')
print(f'Unambiguous range по глубине = {unambiguous_depth:.2f} м '
f'(max total path = {unambiguous_total_path:.2f} м)')
print(f'BP_OVERSAMPLE = {BP_OVERSAMPLE}')
# ══════════════════════════════════════════════════════
# 2. OVERSAMPLED A-СКАНЫ
# ══════════════════════════════════════════════════════
def compute_ascan_bp(s21, freq, f_start, f_stop, window=True, oversample=8):
"""
S21(f) -> A-скан с правильным частотным сдвигом и oversampling.
Частотный шаг df остается тем же, а n_fft увеличивается в oversample раз.
Поэтому временная сетка становится плотнее: dt = 1 / (n_fft * df).
"""
mask = (freq >= f_start) & (freq <= f_stop)
freq_cut = freq[mask]
s21_cut = s21[mask]
if len(freq_cut) < 2:
raise ValueError('После обрезки по частоте осталось меньше двух точек.')
df = float(np.median(np.diff(freq_cut)))
n = len(freq_cut)
k0 = int(round(freq_cut[0] / df))
min_len = 2 * (k0 + n - 1)
n_fft_base = 1 << int(np.ceil(np.log2(min_len)))
n_fft = int(n_fft_base * oversample)
dt = 1.0 / (n_fft * df)
t_sec = np.arange(n_fft, dtype=float) * dt
s = s21_cut * np.hanning(n) if window else s21_cut.copy()
H = np.zeros(n_fft, dtype=np.complex128)
H[k0:k0 + n] = s
h_complex = np.fft.ifft(H)
a_abs = np.abs(h_complex)
return t_sec, a_abs, h_complex, n_fft_base, n_fft
print('Вычисление oversampled A-сканов...', end=' ', flush=True)
A_bp = {}
H_bp = {}
T_bp = {}
Z_bp = {}
n_fft_info = {}
pair_norm_info = {}
for (i, j), s21 in s21_data.items():
s21_proc = s21.copy()
if BG_SUBTRACT and background is not None and (i, j) in background:
s21_proc = s21_proc - background[(i, j)]
t_pair, a_pair, h_pair, n_fft_base, n_fft = compute_ascan_bp(
s21_proc,
freq_data[(i, j)],
f_start=F_START,
f_stop=F_STOP,
window=BP_WINDOW,
oversample=BP_OVERSAMPLE,
)
T_bp[(i, j)] = t_pair
Z_bp[(i, j)] = t_pair * v / 2
A_bp[(i, j)] = a_pair
H_bp[(i, j)] = h_pair
n_fft_info[(i, j)] = (n_fft_base, n_fft)
# Robust per-pair amplitude normalization. It equalizes channel scale, not phase.
for key in sorted(H_bp.keys()):
z_axis = Z_bp[key]
h_abs = np.abs(H_bp[key])
norm_mask = (z_axis >= PAIR_NORM_DEPTH_MIN) & (z_axis <= PAIR_NORM_DEPTH_MAX)
if not np.any(norm_mask):
norm_mask = np.ones_like(z_axis, dtype=bool)
median_val = float(np.median(h_abs[norm_mask]))
p75_val = float(np.percentile(h_abs[norm_mask], 75))
p95_val = float(np.percentile(h_abs[norm_mask], 95))
scale = float(np.percentile(h_abs[norm_mask], PAIR_NORM_PERCENTILE))
if not np.isfinite(scale) or scale <= PAIR_NORM_EPS:
scale = 1.0
pair_norm_info[key] = {
'median': median_val,
'p75': p75_val,
'p95': p95_val,
'scale': scale,
}
if PAIR_NORMALIZE:
H_bp[key] = H_bp[key] / (scale + PAIR_NORM_EPS)
A_bp[key] = np.abs(H_bp[key])
z_h_bp = Z_bp[first_key]
t_h_bp = T_bp[first_key]
print('готово.')
print(f'dt = {(t_h_bp[1] - t_h_bp[0])*1e12:.2f} пс')
print(f'dz_sample = {(z_h_bp[1] - z_h_bp[0])*100:.3f} см')
print('\n' + '=' * 86)
print(f' НОРМИРОВКА КАНАЛОВ Tx/Rx: enabled={PAIR_NORMALIZE}, '
f'percentile={PAIR_NORM_PERCENTILE:.1f}, '
f'z=[{PAIR_NORM_DEPTH_MIN:.2f}, {PAIR_NORM_DEPTH_MAX:.2f}] м')
print('=' * 86)
print(f" {'Pair':<8} {'median':>12} {'p75':>12} {'p95':>12} {'scale':>12} {'rel_scale':>12}")
print('-' * 86)
scales = np.array([v['scale'] for v in pair_norm_info.values()], dtype=float)
scale_ref = float(np.median(scales)) if len(scales) else 1.0
for key in sorted(pair_norm_info.keys()):
info = pair_norm_info[key]
rel_scale = info['scale'] / (scale_ref + PAIR_NORM_EPS)
print(f" Tx{key[0]}-Rx{key[1]:<3} {info['median']:>12.4e} {info['p75']:>12.4e} "
f"{info['p95']:>12.4e} {info['scale']:>12.4e} {rel_scale:>12.3f}")
print('=' * 86)
# ══════════════════════════════════════════════════════
# 3. TIME-DOMAIN COHERENT BACKPROJECTION
# ══════════════════════════════════════════════════════
x_grid_bp = np.linspace(x_min, x_max, NX_BP)
z_grid_bp = np.linspace(z_min, z_max, NZ_BP)
XX_bp, ZZ_bp = np.meshgrid(x_grid_bp, z_grid_bp)
depth_gate = (ZZ_bp >= MIN_DEPTH) & (ZZ_bp <= MAX_DEPTH)
def bistatic_ranges(i_tx, i_rx, XX, ZZ, yy=BP_PLANE_Y):
Rtx = np.sqrt((XX - x_tx[i_tx])**2 + (yy - y_tx[i_tx])**2 + (ZZ - z_tx[i_tx])**2)
Rrx = np.sqrt((XX - x_rx[i_rx])**2 + (yy - y_rx[i_rx])**2 + (ZZ - z_rx[i_rx])**2)
return Rtx, Rrx
def antenna_boresight_cos_z(R, z_ant, ZZ):
# В локальных координатах радара все антенны смотрят вдоль +Z.
return (ZZ - z_ant) / (R + 1e-12)
def attenuation_components_map(i_tx, i_rx, XX, ZZ):
Rtx, Rrx = bistatic_ranges(i_tx, i_rx, XX, ZZ)
geo = 1.0 / (Rtx * Rrx + 1e-12)
cos_tx = antenna_boresight_cos_z(Rtx, z_tx[i_tx], ZZ)
cos_rx = antenna_boresight_cos_z(Rrx, z_rx[i_rx], ZZ)
angle = cos_tx**2 * cos_rx**2
return geo + 1e-30, angle + 1e-30
def attenuation_components_at_ref_depth(i_tx, i_rx, z_ref):
xc = (x_tx[i_tx] + x_rx[i_rx]) / 2.0
Rtx, Rrx = bistatic_ranges(i_tx, i_rx, xc, z_ref)
geo = 1.0 / (Rtx * Rrx + 1e-12)
cos_tx = antenna_boresight_cos_z(Rtx, z_tx[i_tx], z_ref)
cos_rx = antenna_boresight_cos_z(Rrx, z_rx[i_rx], z_ref)
angle = cos_tx**2 * cos_rx**2
return geo + 1e-30, angle + 1e-30
def bp_compensation_weight(i_tx, i_rx, XX, ZZ):
geo, angle = attenuation_components_map(i_tx, i_rx, XX, ZZ)
geo_ref, angle_ref = attenuation_components_at_ref_depth(i_tx, i_rx, COMP_REF_DEPTH)
geo_norm = geo / geo_ref
angle_norm = angle / angle_ref
range_weight = 1.0 / (geo_norm ** COMP_RANGE_POWER + 1e-12)
angle_weight = 1.0 / (angle_norm ** COMP_ANGLE_POWER + 1e-12)
range_weight = np.clip(range_weight, 0.0, COMP_RANGE_WEIGHT_MAX)
angle_weight = np.clip(angle_weight, 0.0, COMP_ANGLE_WEIGHT_MAX)
weight = range_weight * angle_weight
return np.clip(weight, 0.0, COMP_WEIGHT_MAX)
def interpolate_ascan_amplitude(tau, t_axis, a_axis):
return np.interp(tau.ravel(), t_axis, a_axis, left=0.0, right=0.0).reshape(tau.shape)
def interpolate_ascan_complex(tau, t_axis, h_axis):
h_real = np.interp(tau.ravel(), t_axis, h_axis.real, left=0.0, right=0.0)
h_imag = np.interp(tau.ravel(), t_axis, h_axis.imag, left=0.0, right=0.0)
return (h_real + 1j * h_imag).reshape(tau.shape)
def backproject_coherent(H, T, compensate=True):
bp_complex = np.zeros_like(XX_bp, dtype=np.complex128)
contribution_count = np.zeros_like(XX_bp, dtype=float)
for i in range(N_tx):
for j in range(N_rx):
key = (i, j)
if key not in H:
continue
Rtx, Rrx = bistatic_ranges(i, j, XX_bp, ZZ_bp)
tau = (Rtx + Rrx) / v
valid = depth_gate & (tau >= T[key][0]) & (tau <= T[key][-1])
h_tau = interpolate_ascan_complex(tau, T[key], H[key])
h_tau = np.where(valid, h_tau, 0.0 + 0.0j)
if compensate:
w = bp_compensation_weight(i, j, XX_bp, ZZ_bp)
w = np.where(valid, w, 0.0)
else:
w = np.where(valid, 1.0, 0.0)
bp_complex += h_tau * w
contribution_count += valid.astype(float)
bp_complex = bp_complex / (contribution_count + 1e-12)
bp_complex = np.where(depth_gate, bp_complex, 0.0 + 0.0j)
bp_abs = np.abs(bp_complex)
return bp_abs, bp_complex
def backproject_coherent_and_incoherent(H, T, compensate=True):
"""Одним проходом строит coherent |sum h| и incoherent sum |h| BP-карты."""
bp_complex = np.zeros_like(XX_bp, dtype=np.complex128)
bp_incoherent = np.zeros_like(XX_bp, dtype=float)
contribution_count = np.zeros_like(XX_bp, dtype=float)
for i in range(N_tx):
for j in range(N_rx):
key = (i, j)
if key not in H:
continue
Rtx, Rrx = bistatic_ranges(i, j, XX_bp, ZZ_bp)
tau = (Rtx + Rrx) / v
valid = depth_gate & (tau >= T[key][0]) & (tau <= T[key][-1])
h_tau = interpolate_ascan_complex(tau, T[key], H[key])
h_tau = np.where(valid, h_tau, 0.0 + 0.0j)
if compensate:
w = bp_compensation_weight(i, j, XX_bp, ZZ_bp)
w = np.where(valid, w, 0.0)
else:
w = np.where(valid, 1.0, 0.0)
bp_complex += h_tau * w
bp_incoherent += np.abs(h_tau) * w
contribution_count += valid.astype(float)
bp_complex = bp_complex / (contribution_count + 1e-12)
bp_complex = np.where(depth_gate, bp_complex, 0.0 + 0.0j)
bp_abs = np.abs(bp_complex)
bp_incoherent = bp_incoherent / (contribution_count + 1e-12)
bp_incoherent = np.where(depth_gate, bp_incoherent, 0.0)
bp_cf = bp_abs / (bp_incoherent + BP_SCORE_CF_EPS)
bp_cf = np.where(depth_gate, bp_cf, 0.0)
bp_cf = np.clip(bp_cf, 0.0, 1.0)
return bp_abs, bp_complex, bp_incoherent, bp_cf
def normalize_bp_map(bp_raw, smooth_sigma=BP_SMOOTH_SIGMA):
bp_norm = bp_raw / (bp_raw.max() + 1e-12)
bp_norm = gaussian_filter(bp_norm, sigma=smooth_sigma)
bp_norm = np.where(depth_gate, bp_norm, 0.0)
return bp_norm / (bp_norm.max() + 1e-12)
def component_containing_peak(image, iz, ix, threshold, window_mask=None):
mask = image >= threshold
if window_mask is not None:
mask &= window_mask
labels, n_labels = label(mask, structure=np.ones((3, 3), dtype=int))
if n_labels == 0 or labels[iz, ix] == 0:
fallback = np.zeros_like(image, dtype=bool)
fallback[iz, ix] = True
return fallback
return labels == labels[iz, ix]
def weighted_centroid(image, region_mask, threshold=0.0):
values = image[region_mask]
weights = np.clip(values - threshold, 0.0, None)
if weights.sum() <= 1e-15:
weights = values.copy()
if weights.sum() <= 1e-15:
iz, ix = np.argwhere(region_mask)[0]
return x_grid_bp[ix], z_grid_bp[iz]
x_vals = XX_bp[region_mask]
z_vals = ZZ_bp[region_mask]
return (x_vals * weights).sum() / weights.sum(), (z_vals * weights).sum() / weights.sum()
def compact_peak_centroid(image, iz, ix, peak):
x0 = x_grid_bp[ix]
z0 = z_grid_bp[iz]
window_mask = (
(np.abs(XX_bp - x0) <= BP_CENTER_RX_CM / 100.0) &
(np.abs(ZZ_bp - z0) <= BP_CENTER_RZ_CM / 100.0)
)
threshold = BP_CENTER_THRESH_FRAC * peak
center_mask = window_mask & (image >= threshold)
if center_mask.sum() == 0:
center_mask = window_mask.copy()
center_mask[iz, ix] = True
values = image[center_mask]
weights = np.clip(values - threshold, 0.0, None) ** BP_CENTER_WEIGHT_POWER
if weights.sum() <= 1e-15:
weights = values.copy()
if weights.sum() <= 1e-15:
return x0, z0, center_mask
x_vals = XX_bp[center_mask]
z_vals = ZZ_bp[center_mask]
x_c = (x_vals * weights).sum() / weights.sum()
z_c = (z_vals * weights).sum() / weights.sum()
return x_c, z_c, center_mask
def find_bp_objects(bp_image):
"""
CLEAN-подобный поиск объектов на BP-карте.
Для каждого шага берется максимум рабочей карты, вокруг него выделяется
связная область выше BP_REGION_THRESH_FRAC от локального пика, затем
считается взвешенный центроид этой области. После этого более широкая
область вокруг той же цели подавляется на рабочей карте.
"""
work = bp_image.copy()
objects = []
global_peak = float(work.max())
stop_level = BP_OBJECT_MIN_FRAC * global_peak
dx_cm = abs(x_grid_bp[1] - x_grid_bp[0]) * 100
dz_cm = abs(z_grid_bp[1] - z_grid_bp[0]) * 100
pixel_area_cm2 = dx_cm * dz_cm
for step in range(MAX_OBJECTS):
peak = float(work.max())
if peak <= stop_level or peak <= 0:
break
iz, ix = np.unravel_index(np.argmax(work), work.shape)
x_peak_local = x_grid_bp[ix]
z_peak_local = z_grid_bp[iz]
region_threshold = BP_REGION_THRESH_FRAC * peak
region_mask = component_containing_peak(work, iz, ix, region_threshold)
region_area_cm2 = float(region_mask.sum() * pixel_area_cm2)
if region_area_cm2 < BP_MIN_REGION_AREA_CM2:
work[iz, ix] = 0.0
continue
x_region_centroid, z_region_centroid = weighted_centroid(work, region_mask, threshold=region_threshold)
if BP_CENTER_USE_COMPACT:
x_centroid, z_centroid, center_mask = compact_peak_centroid(work, iz, ix, peak)
else:
x_centroid, z_centroid = x_region_centroid, z_region_centroid
center_mask = region_mask.copy()
center_area_cm2 = float(center_mask.sum() * pixel_area_cm2)
region_values = work[region_mask]
objects.append({
'index': len(objects) + 1,
'ix_peak': int(ix),
'iz_peak': int(iz),
'x_peak': float(x_peak_local),
'z_peak': float(z_peak_local),
'x': float(x_centroid),
'z': float(z_centroid),
'x_region': float(x_region_centroid),
'z_region': float(z_region_centroid),
'peak': peak,
'area_cm2': region_area_cm2,
'center_area_cm2': center_area_cm2,
'region_mask': region_mask.copy(),
'center_mask': center_mask.copy(),
'mean_value': float(region_values.mean()),
'sum_value': float(region_values.sum()),
})
if BP_SUPPRESS_USE_WINDOW:
suppress_window = (
(np.abs(XX_bp - x_peak_local) <= BP_SUPPRESS_RX_CM / 100.0) &
(np.abs(ZZ_bp - z_peak_local) <= BP_SUPPRESS_RZ_CM / 100.0)
)
else:
suppress_window = None
suppress_threshold = BP_SUPPRESS_THRESH_FRAC * peak
suppress_mask = component_containing_peak(
work, iz, ix, suppress_threshold, window_mask=suppress_window
)
# Если широкий порог дал слишком маленькую область, подавляем хотя бы область детекции.
if suppress_mask.sum() < region_mask.sum():
suppress_mask = region_mask
work[suppress_mask] = 0.0
return objects, work
def circular_phase_stats(bp_complex_map, mask):
if mask.sum() == 0:
return np.nan, np.nan, np.nan
h = bp_complex_map[mask]
amp = np.abs(h)
valid = amp > 0
if not np.any(valid):
return np.nan, np.nan, np.nan
phase = np.angle(h[valid])
weights = amp[valid] ** BP_PHASE_WEIGHT_POWER
if weights.sum() <= 1e-15:
weights = np.ones_like(phase)
vec = np.sum(weights * np.exp(1j * phase)) / (np.sum(weights) + 1e-15)
phase_mean = float(np.angle(vec))
phase_coherence = float(np.abs(vec))
phase_circular_variance = float(1.0 - phase_coherence)
return phase_mean, phase_coherence, phase_circular_variance
def add_phase_metrics(objects, bp_complex_map):
for obj in objects:
mask = obj.get('center_mask', obj['region_mask'])
phase_mean, phase_coh, phase_var = circular_phase_stats(bp_complex_map, mask)
obj['phase_mean_rad'] = phase_mean
obj['phase_coherence'] = phase_coh
obj['phase_circular_variance'] = phase_var
return objects
def add_local_prominence_metrics(objects, bp_image):
for obj in objects:
x0 = obj['x_peak']
z0 = obj['z_peak']
outer_mask = (
(np.abs(XX_bp - x0) <= BP_LOCAL_BG_RX_CM / 100.0) &
(np.abs(ZZ_bp - z0) <= BP_LOCAL_BG_RZ_CM / 100.0) &
depth_gate
)
bg_mask = outer_mask & (~obj['region_mask'])
if bg_mask.sum() < 10:
bg_mask = depth_gate & (~obj['region_mask'])
if bg_mask.sum() == 0:
bg_level = 0.0
bg_p75 = 0.0
else:
bg_values = bp_image[bg_mask]
bg_level = float(np.percentile(bg_values, BP_LOCAL_BG_PERCENTILE))
bg_p75 = float(np.percentile(bg_values, 75))
peak = float(obj['peak'])
obj['local_bg'] = bg_level
obj['local_bg_p75'] = bg_p75
obj['prominence'] = peak - bg_level
obj['contrast'] = peak / (bg_level + BP_LOCAL_CONTRAST_EPS)
return objects
def add_incoherent_support_metrics(objects, bp_incoherent_image, bp_cf_image=None):
for obj in objects:
if bp_incoherent_image is None:
obj['incoh_peak'] = 0.0
obj['incoh_mean'] = 0.0
obj['incoh_center_mean'] = 0.0
obj['coherence_factor_peak_raw'] = 0.0
obj['coherence_factor_center_raw'] = 0.0
obj['coherence_factor_peak'] = 0.0
obj['coherence_factor_center'] = 0.0
continue
region_mask = obj['region_mask']
center_mask = obj.get('center_mask', region_mask)
if region_mask.sum() == 0:
obj['incoh_peak'] = 0.0
obj['incoh_mean'] = 0.0
else:
region_values = bp_incoherent_image[region_mask]
obj['incoh_peak'] = float(region_values.max())
obj['incoh_mean'] = float(region_values.mean())
if center_mask.sum() == 0:
obj['incoh_center_mean'] = obj['incoh_mean']
else:
obj['incoh_center_mean'] = float(bp_incoherent_image[center_mask].mean())
if bp_cf_image is None:
obj['coherence_factor_peak_raw'] = float(
obj.get('peak', 0.0) / (obj['incoh_peak'] + BP_SCORE_CF_EPS)
)
obj['coherence_factor_center_raw'] = float(
obj.get('mean_value', 0.0) / (obj['incoh_center_mean'] + BP_SCORE_CF_EPS)
)
else:
iz_peak = int(obj.get('iz_peak', 0))
ix_peak = int(obj.get('ix_peak', 0))
obj['coherence_factor_peak_raw'] = float(bp_cf_image[iz_peak, ix_peak])
if center_mask.sum() == 0:
obj['coherence_factor_center_raw'] = obj['coherence_factor_peak_raw']
else:
obj['coherence_factor_center_raw'] = float(bp_cf_image[center_mask].mean())
obj['coherence_factor_peak'] = float(np.clip(obj['coherence_factor_peak_raw'], 0.0, 1.0))
obj['coherence_factor_center'] = float(np.clip(obj['coherence_factor_center_raw'], 0.0, 1.0))
return objects
def _contrast_score_unit(contrast):
if not np.isfinite(contrast):
return 0.0
if BP_SCORE_CONTRAST_CAP <= 1.0:
return 0.0
return float(np.clip((contrast - 1.0) / (BP_SCORE_CONTRAST_CAP - 1.0), 0.0, 1.0))
def add_bp_score_metrics(objects):
total_weight = (
BP_SCORE_COH_PEAK_WEIGHT +
BP_SCORE_COHERENCE_FACTOR_WEIGHT +
BP_SCORE_PROMINENCE_WEIGHT +
BP_SCORE_CONTRAST_WEIGHT
)
if total_weight <= 0:
total_weight = 1.0
for obj in objects:
coh_peak_score = float(np.clip(obj.get('peak', 0.0), 0.0, 1.0))
coherence_factor_score = float(np.clip(obj.get('coherence_factor_peak', 0.0), 0.0, 1.0))
prominence_score = float(np.clip(obj.get('prominence', 0.0), 0.0, 1.0))
contrast_score = _contrast_score_unit(obj.get('contrast', np.nan))
score_combined = (
BP_SCORE_COH_PEAK_WEIGHT * coh_peak_score +
BP_SCORE_COHERENCE_FACTOR_WEIGHT * coherence_factor_score +
BP_SCORE_PROMINENCE_WEIGHT * prominence_score +
BP_SCORE_CONTRAST_WEIGHT * contrast_score
) / total_weight
obj['score_old'] = coh_peak_score
obj['score_new'] = float(score_combined)
obj['score_coh_peak_part'] = coh_peak_score
obj['score_cf_part'] = coherence_factor_score
obj['score_prominence_part'] = prominence_score
obj['score_contrast_part'] = contrast_score
obj['score_selected'] = obj['score_new'] if BP_SCORE_MODE == 'combined' else obj['score_old']
return objects
def prepare_bp_objects_for_display(objects, remove_sidelobes=True):
if remove_sidelobes:
visible = [obj for obj in objects if not obj.get('sidelobe_candidate', False)]
else:
visible = list(objects)
if BP_SCORE_MODE == 'combined':
visible = sorted(visible, key=lambda obj: obj.get('score_new', 0.0), reverse=True)
else:
visible = sorted(visible, key=lambda obj: obj.get('score_old', obj.get('peak', 0.0)), reverse=True)
for rank, obj in enumerate(visible, start=1):
obj['display_index'] = rank
return visible
def bistatic_depth_signature(x_obj, z_obj):
signature = []
for i in range(N_tx):
for j in range(N_rx):
Rtx, Rrx = bistatic_ranges(i, j, x_obj, z_obj)
signature.append(0.5 * (Rtx + Rrx))
return np.array(signature, dtype=float)
def mark_sidelobe_candidates(objects):
for obj in objects:
obj['sidelobe_candidate'] = False
obj['sidelobe_parent'] = None
obj['sidelobe_range_rms_cm'] = np.nan
obj['sidelobe_dx_cm'] = np.nan
obj['sidelobe_dz_cm'] = np.nan
if not BP_SIDELOBE_DETECT:
return objects
signatures = [bistatic_depth_signature(obj['x'], obj['z']) for obj in objects]
for k, obj in enumerate(objects):
best_parent = None
best_rms_cm = np.inf
best_dx_cm = np.nan
best_dz_cm = np.nan
for p in range(k):
parent = objects[p]
rel_peak = obj['peak'] / (parent['peak'] + 1e-12)
dx_cm = abs(obj['x'] - parent['x']) * 100.0
dz_cm = abs(obj['z'] - parent['z']) * 100.0
rms_cm = float(np.sqrt(np.mean((signatures[k] - signatures[p])**2)) * 100.0)
is_candidate = (
rel_peak <= BP_SIDELOBE_MAX_REL_PEAK and
dx_cm >= BP_SIDELOBE_MIN_DX_CM and
dz_cm <= BP_SIDELOBE_MAX_DZ_CM and
rms_cm <= BP_SIDELOBE_RANGE_RMS_TOL_CM
)
if is_candidate and rms_cm < best_rms_cm:
best_parent = parent
best_rms_cm = rms_cm
best_dx_cm = dx_cm
best_dz_cm = dz_cm
if best_parent is not None:
obj['sidelobe_candidate'] = True
obj['sidelobe_parent'] = best_parent['index']
obj['sidelobe_range_rms_cm'] = best_rms_cm
obj['sidelobe_dx_cm'] = best_dx_cm
obj['sidelobe_dz_cm'] = best_dz_cm
return objects
if BP_SCORE_COMPUTE_INCOHERENT:
print('Расчет coherent + incoherent time-domain BP...', end=' ', flush=True)
bp_raw, bp_complex, bp_incoh_raw, bp_cf_map = backproject_coherent_and_incoherent(H_bp, T_bp, compensate=True)
bp_incoh_map_s = normalize_bp_map(bp_incoh_raw)
else:
print('Расчет coherent time-domain BP...', end=' ', flush=True)
bp_raw, bp_complex = backproject_coherent(H_bp, T_bp, compensate=True)
bp_incoh_raw = None
bp_incoh_map_s = None
bp_cf_map = None
bp_map_s = normalize_bp_map(bp_raw)
print('готово.')
print(f'BP score mode: {BP_SCORE_MODE} (peak=старый, combined=экспериментальный)')
if BP_VALIDATE_COMPENSATION:
print('Расчет coherent BP без compensation для валидации...', end=' ', flush=True)
bp_raw_nocomp, bp_complex_nocomp = backproject_coherent(H_bp, T_bp, compensate=False)
bp_map_nocomp_s = normalize_bp_map(bp_raw_nocomp)
print('готово.')
else:
bp_raw_nocomp = None
bp_complex_nocomp = None
bp_map_nocomp_s = None
bp_objects_all, bp_residual = find_bp_objects(bp_map_s)
bp_objects_all = add_phase_metrics(bp_objects_all, bp_complex)
bp_objects_all = add_local_prominence_metrics(bp_objects_all, bp_map_s)
bp_objects_all = add_incoherent_support_metrics(bp_objects_all, bp_incoh_map_s, bp_cf_map)
bp_objects_all = mark_sidelobe_candidates(bp_objects_all)
bp_objects_all = add_bp_score_metrics(bp_objects_all)
bp_map_display = bp_map_s.copy()
if BP_REMOVE_SIDELOBE_OBJECTS:
for obj in bp_objects_all:
if obj['sidelobe_candidate']:
bp_map_display[obj['region_mask']] = 0.0
bp_objects = prepare_bp_objects_for_display(bp_objects_all, remove_sidelobes=BP_REMOVE_SIDELOBE_OBJECTS)
bp_detected_object_count = len(bp_objects)
if bp_detected_object_count > BP_MAX_DETECTED_OBJECTS_TO_DRAW:
bp_objects_to_plot = []
else:
bp_objects_to_plot = bp_objects[:BP_DRAW_TOP_M_OBJECTS]
iz_max, ix_max = np.unravel_index(np.argmax(bp_map_s), bp_map_s.shape)
x_peak = x_grid_bp[ix_max]
z_peak = z_grid_bp[iz_max]
peak_value = bp_map_s[iz_max, ix_max]
main_obj = bp_objects[0] if bp_objects else None
print('\n' + '=' * 72)
print(' COHERENT TIME-DOMAIN BP: максимум карты и выбранный объект')
print('=' * 72)
print(f' argmax: x = {x_peak*100:+.1f} см, z = {z_peak*100:.1f} см, BP = {peak_value:.3f}')
if main_obj is not None:
print(f" selected: x = {main_obj['x']*100:+.1f} см, z = {main_obj['z']*100:.1f} см, "
f"score = {main_obj['score_selected']:.3f}, area = {main_obj['area_cm2']:.1f} см^2")
print('=' * 72)
print('\n' + '=' * 72)
n_sl_all = sum(obj['sidelobe_candidate'] for obj in bp_objects_all)
if BP_REMOVE_SIDELOBE_OBJECTS:
print(f' НАЙДЕННЫЕ ОБЪЕКТЫ НА COHERENT BP-КАРТЕ, max {MAX_OBJECTS} '
f'(SL скрыты: {n_sl_all})')
else:
print(f' НАЙДЕННЫЕ ОБЪЕКТЫ НА COHERENT BP-КАРТЕ, max {MAX_OBJECTS} '
f'(SL показаны: {n_sl_all})')
print('=' * 166)
print(f" {'#':<4} {'Det#':>5} {'Xc [см]':>10} {'Zc [см]':>10} {'Xmax [см]':>11} {'Zmax [см]':>11} "
f"{'Peak':>7} {'Incoh':>7} {'CF':>6} {'Prom':>7} {'Contr':>7} {'Score':>7} {'NewSc':>7} "
f"{'Area [см2]':>11} {'PhCoh':>7} {'SL?':>5} {'Parent':>6} {'RMSr [см]':>10}")
print('-' * 166)
for obj in bp_objects_to_plot:
sl_label = 'yes' if obj['sidelobe_candidate'] else 'no'
parent_label = '-' if obj['sidelobe_parent'] is None else str(obj['sidelobe_parent'])
rms_label = '-' if np.isnan(obj['sidelobe_range_rms_cm']) else f"{obj['sidelobe_range_rms_cm']:.1f}"
print(f" {obj['display_index']:<4} {obj['index']:>5} {obj['x']*100:>+10.1f} {obj['z']*100:>10.1f} "
f"{obj['x_peak']*100:>+11.1f} {obj['z_peak']*100:>11.1f} "
f"{obj['peak']:>7.3f} {obj['incoh_peak']:>7.3f} {obj['coherence_factor_peak']:>6.3f} "
f"{obj['prominence']:>7.3f} {obj['contrast']:>7.2f} "
f"{obj['score_selected']:>7.3f} {obj['score_new']:>7.3f} "
f"{obj['area_cm2']:>11.1f} {obj['phase_coherence']:>7.3f} "
f"{sl_label:>5} {parent_label:>6} {rms_label:>10}")
print('=' * 166)
# ══════════════════════════════════════════════════════
# 4. ТОЛЬКО BP-КАРТА
# ══════════════════════════════════════════════════════
if BP_ONLY_MAP_OUTPUT:
_bp_stdout_redirect.__exit__(None, None, None)
fig, ax = plt.subplots(figsize=(12, 7))
im = ax.imshow(
bp_map_display,
extent=[x_grid_bp[0]*100, x_grid_bp[-1]*100, z_grid_bp[-1]*100, z_grid_bp[0]*100],
aspect='auto',
cmap='jet',
vmin=0.25,
vmax=0.95,
)
plt.colorbar(im, ax=ax, label='Нормированная |coherent BP|')
ax.plot(x_tx * 100, z_tx * 100, 'r^', ms=12, label='Tx', zorder=5)
ax.plot(x_rx * 100, z_rx * 100, 'bv', ms=12, label='Rx', zorder=5)
for obj in bp_objects_to_plot:
ax.contour(
x_grid_bp * 100,
z_grid_bp * 100,
obj['region_mask'].astype(float),
levels=[0.5],
colors='white',
linewidths=0.9,
alpha=0.75,
)
ax.contour(
x_grid_bp * 100,
z_grid_bp * 100,
obj['center_mask'].astype(float),
levels=[0.5],
colors='cyan',
linewidths=0.8,
alpha=0.85,
)
if obj['sidelobe_candidate']:
ax.plot(obj['x'] * 100, obj['z'] * 100, 'x', color='yellow', ms=9,
mew=2.0, zorder=7)
label_text = f"{obj['index']} SL"
text_color = 'yellow'
else:
ax.plot(obj['x'] * 100, obj['z'] * 100, 'wo', ms=7,
markeredgecolor='k', mew=0.8, zorder=6)
label_text = str(obj.get('display_index', obj['index']))
text_color = 'white'
ax.text(obj['x'] * 100 + 3, obj['z'] * 100, label_text,
color=text_color, fontsize=9, weight='bold', zorder=7)
if len(bp_objects_to_plot) > 0:
ax.plot([], [], 'wo', ms=7, markeredgecolor='k', mew=0.8, label='Центроид области')
if not BP_REMOVE_SIDELOBE_OBJECTS:
ax.plot([], [], 'x', color='yellow', ms=9, mew=2.0, label='Sidelobe candidate')
ax.axhline(MIN_DEPTH * 100, color='white', lw=1.0, ls='--', alpha=0.75)
ax.set_xlabel('X [см]')
ax.set_ylabel('Глубина Z [см]')
map_title = f'Time-domain coherent BackProjection | score={BP_SCORE_MODE}, shown={len(bp_objects_to_plot)}/{bp_detected_object_count}'
if BP_REMOVE_SIDELOBE_OBJECTS:
map_title += ' (SL области скрыты на карте)'
ax.set_title(map_title)
ax.set_xlim(x_grid_bp[0] * 100, x_grid_bp[-1] * 100)
ax.set_ylim(z_grid_bp[-1] * 100, 0)
ax.legend(loc='lower right', fontsize=9)
ax.grid(alpha=0.22)
ax.invert_yaxis()
plt.tight_layout()
plt.show()
Reference in New Issue View Git Blame Copy Permalink