F1-steering-angle-model / utils /model_handler.py
daniel-saed's picture
Update utils/model_handler.py
1d44e42 verified
Raw
History Blame Contribute Delete
11.1 kB
import numpy as np
import pandas as pd
from typing import List, Dict
from PIL import Image
import cv2
import onnxruntime as ort
from utils.helper import BASE_DIR
from pathlib import Path
def denormalize_angles(normalized_angles):
"""
Convierte ángulos normalizados [-1,1] a grados [-180,180]
"""
return (normalized_angles + 1) / 2 * (180 - (-180)) + (-180)
def preprocess_image_exactly_like_pytorch(image_input):
"""
Preprocesa una imagen de OpenCV (como adjusted_edges)
para usarla con modelos ONNX.
Args:
image_input: Array NumPy de OpenCV (imagen de bordes, binaria, etc.)
Returns:
Array NumPy listo para inferencia con ONNX
"""
# Verificar que la entrada no sea None
if image_input is None:
raise ValueError("Received None as image input")
# Asegurar que la imagen es un array NumPy
if not isinstance(image_input, np.ndarray):
raise TypeError(f"Expected NumPy array, got {type(image_input)}")
# Verificar que la imagen tiene dimensiones válidas
if len(image_input.shape) < 2:
raise ValueError(f"Invalid image shape: {image_input.shape}")
# Copia para no modificar la original
img_copy = image_input.copy()
# Si es una imagen de bordes o binaria, normalmente tiene valores 0 y 255
# o 0 y 1. Asegurarse de que está en el rango [0, 255]
if img_copy.dtype != np.uint8:
if np.max(img_copy) <= 1.0:
# Si está en rango [0, 1], convertir a [0, 255]
img_copy = (img_copy * 255).astype(np.uint8)
else:
# De otro modo, simplemente convertir a uint8
img_copy = img_copy.astype(np.uint8)
# Para imágenes de bordes o binarias, asegurar que tenemos valores claros
# (si todos los valores son muy bajos, puede que no se vea nada)
if np.mean(img_copy) < 10 and np.max(img_copy) > 0:
# Estirar el contraste para mejor visualización
img_copy = cv2.normalize(img_copy, None, 0, 255, cv2.NORM_MINMAX)
# Asegurar que la imagen es de un solo canal (escala de grises)
if len(img_copy.shape) == 3:
if img_copy.shape[2] == 3:
# Convertir imagen BGR a escala de grises
img_copy = cv2.cvtColor(img_copy, cv2.COLOR_BGR2GRAY)
else:
# Tomar solo el primer canal
img_copy = img_copy[:, :, 0]
try:
# Convertir de NumPy array a PIL Image
img_pil = Image.fromarray(img_copy)
# Redimensionar con PIL
img_resized = img_pil.resize((224, 224), Image.BILINEAR)
# Convertir a numpy array
img_np = np.array(img_resized, dtype=np.float32)
# Normalizar de [0,255] a [0,1]
img_np = img_np / 255.0
# Normalizar con mean=0.5, std=0.5 (como en PyTorch)
img_np = (img_np - 0.5) / 0.5
# Reformatear para ONNX [batch_size, channels, height, width]
img_np = np.expand_dims(img_np, axis=0) # Añadir dimensión de canal
img_np = np.expand_dims(img_np, axis=0) # Añadir dimensión de batch
return img_np
except Exception as e:
print(f"Error processing image: {e}")
print(f"Image shape: {image_input.shape}, dtype: {image_input.dtype}")
print(f"Min value: {np.min(image_input)}, Max value: {np.max(image_input)}")
raise
def correct_outlier_angles(df, window_size=5, std_threshold=3.0, max_diff_threshold=80.0):
angles = df['steering_angle'].values
corrected_angles = angles.copy()
for i in range(len(angles)):
if i < window_size // 2 or i >= len(angles) - window_size // 2: # Evitar bordes
continue
# Definir ventana local
start_idx = max(0, i - window_size // 2)
end_idx = min(len(angles), i + window_size // 2 + 1)
window = angles[start_idx:end_idx]
# Calcular estadísticas locales incluyendo el valor actual
curr_angle = angles[i]
local_mean = np.mean(window)
local_std = np.std(window) if len(window) > 1 else 0
# Calcular distancia angular mínima considerando el rango cíclico (-180° a 180°)
def angular_distance(a, b):
diff = abs(a - b)
return min(diff, 360 - diff) if diff > 180 else diff
diff_from_mean = angular_distance(curr_angle, local_mean)
# Detectar outlier
is_outlier = (diff_from_mean > std_threshold * local_std) or (diff_from_mean > max_diff_threshold)
if is_outlier:
print(i, curr_angle, local_mean, local_std, diff_from_mean)
# Excluir el valor actual del cálculo del promedio de corrección
corrected_window = np.delete(window, i - start_idx)
if len(corrected_window) > 0:
corrected_mean = np.mean(corrected_window)
# Ajustar el ángulo corregido al rango cíclico más cercano
diff_to_corrected = angular_distance(curr_angle, corrected_mean)
if diff_to_corrected > 180:
corrected_angles[i] = corrected_mean - 360 if corrected_mean > 0 else corrected_mean + 360
else:
corrected_angles[i] = corrected_mean
# Crear nuevo DataFrame
corrected_df = df.copy()
corrected_df['steering_angle'] = corrected_angles
return corrected_df
class ModelHandler:
def __init__(self):
# Placeholder for actual model loading
self.current_model = None
self.current_model_name = None
self.fps = None
self.available_models = {
"F1 Steering Angle Detection": Path(BASE_DIR) / "models" / "f1-steering-angle-model.onnx",
"Track Position Analysis": "position_model",
"Driver Behavior Analysis": "behavior_model"
}
def _load_model_if_needed(self, model_name: str):
"""Load the model only if it's not already loaded or if it's different"""
if self.current_model is None or self.current_model_name != model_name:
print(f"Loading model: {model_name}") # Debugging info
self.current_model = ort.InferenceSession(self.available_models[model_name])
self.current_model_name = model_name
def process_frames(self, frames: List[np.ndarray], model_name: str) -> Dict:
"""Process frames through selected model with efficient batch processing"""
if not frames:
return []
# Load model only once
self._load_model_if_needed(model_name)
# Get input name once
input_name = self.current_model.get_inputs()[0].name
results = []
# Define optimal batch size - ajusta según tu hardware
BATCH_SIZE = 16
index = 0
# Process frames in batches
for batch_start in range(0, len(frames), BATCH_SIZE):
# Get current batch
batch_end = min(batch_start + BATCH_SIZE, len(frames))
current_batch = frames[batch_start:batch_end]
batch_inputs = []
# Pre-process all frames in the current batch
for frame in current_batch:
try:
# Procesar imagen pero mantener en formato que permita agrupación
#cv2.imwrite(r"img_test/"+str(index)+".jpg", frame)
index= index+1
processed_input = preprocess_image_exactly_like_pytorch(frame)
batch_inputs.append(processed_input)
except Exception as e:
print(f"Error preprocessing frame: {e}")
# Usar un tensor vacío del mismo tamaño como reemplazo
empty_tensor = np.zeros((1, 1, 224, 224), dtype=np.float32)
batch_inputs.append(empty_tensor)
try:
# Combinar todos los inputs pre-procesados en un solo lote grande
# Cada input tiene forma [1, 1, 224, 224], los concatenamos en la dimensión 0
batched_input = np.vstack(batch_inputs)
# Ejecutar inferencia sobre todo el lote a la vez
ort_inputs = {input_name: batched_input}
ort_outputs = self.current_model.run(None, ort_inputs)
# Procesar resultados por lotes
for i in range(len(current_batch)):
frame_idx = batch_start + i +1
predicted_angle_normalized = ort_outputs[0][i][0]
angle = denormalize_angles(predicted_angle_normalized)
confidence = np.random.uniform(0.7, 0.99)
results.append({
'frame_number': frame_idx,
'steering_angle': angle,
})
except Exception as e:
print(f"Error in batch processing: {e}")
# Si falla el procesamiento por lotes, volver a procesar individualmente
for i, frame in enumerate(current_batch):
frame_idx = batch_start + i +1
try:
input_data = preprocess_image_exactly_like_pytorch(frame)
ort_inputs = {input_name: input_data}
ort_outputs = self.current_model.run(None, ort_inputs)
predicted_angle_normalized = ort_outputs[0][0][0]
angle = denormalize_angles(predicted_angle_normalized)
confidence = np.random.uniform(0.7, 0.99)
results.append({
'frame_number': frame_idx,
'steering_angle': angle
})
except Exception as sub_e:
print(f"Error processing individual frame {frame_idx}: {sub_e}")
# Añadir un resultado con valores predeterminados
results.append({
'frame_number': frame_idx,
'steering_angle': 0.0
})
return results
def export_results(self, results: Dict) -> pd.DataFrame:
"""Convert results to pandas DataFrame for export"""
df = pd.DataFrame(results)
df['time'] = round(df['frame_number'] / self.fps,3)
df = correct_outlier_angles(df, window_size=3, std_threshold=100, max_diff_threshold=15.0)
df = correct_outlier_angles(df, window_size=3, std_threshold=100, max_diff_threshold=15.0)
df = correct_outlier_angles(df, window_size=3, std_threshold=100, max_diff_threshold=15.0)
df = correct_outlier_angles(df, window_size=3, std_threshold=100, max_diff_threshold=15.0)
return df