File size: 11,318 Bytes
58174b0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 | """
Synthetic data generation for training and testing the FSD model.
Generates realistic simulated scenarios including:
- Camera images (synthetic patterns representing road scenes)
- Ultrasonic distance readings
- Ground truth labels for all perception/planning/control tasks
"""
import torch
import numpy as np
from typing import Dict, Optional, Tuple
import math
from .config import VehicleConfig
class FSDDataGenerator:
"""
Generates synthetic training data for the FSD model.
Can produce:
- Simulated camera images
- Ultrasonic distance readings with noise
- Ground truth detection heatmaps
- Ground truth segmentation maps
- Ground truth occupancy grids
- Ground truth waypoints and control commands
"""
def __init__(
self,
vehicle_config: Optional[VehicleConfig] = None,
bev_size: int = 200,
image_size: Tuple[int, int] = (480, 640), # H, W
):
if vehicle_config is None:
vehicle_config = VehicleConfig()
self.config = vehicle_config
self.sensor_config = vehicle_config.sensor_config
self.bev_size = bev_size
self.image_size = image_size
def generate_batch(
self,
batch_size: int = 4,
scenario: str = "urban",
device: str = "cpu",
) -> Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor]]:
"""
Generate a batch of synthetic data.
Args:
batch_size: Number of samples
scenario: "urban", "highway", "parking", "intersection"
device: torch device
Returns:
inputs: Dict of model inputs
targets: Dict of ground truth labels
"""
B = batch_size
N_cam = self.sensor_config.num_cameras
N_us = self.sensor_config.num_ultrasonics
H, W = self.image_size
# ── Generate Camera Data ──
camera_images = self._generate_camera_images(B, N_cam, H, W, scenario)
camera_intrinsics = self._generate_intrinsics(B, N_cam)
camera_extrinsics = self._generate_extrinsics(B, N_cam)
# ── Generate Ultrasonic Data ──
us_distances, us_placements = self._generate_ultrasonic_data(B, N_us, scenario)
# ── Generate Ego State ──
ego_state = self._generate_ego_state(B, scenario)
# ── Generate Navigation Command ──
nav_command = torch.randint(0, 10, (B,))
# ── Generate Ground Truth ──
targets = self._generate_targets(B, scenario)
inputs = {
"camera_images": camera_images.to(device),
"camera_intrinsics": camera_intrinsics.to(device),
"camera_extrinsics": camera_extrinsics.to(device),
"ultrasonic_distances": us_distances.to(device),
"ultrasonic_placements": us_placements.to(device),
"ego_state": ego_state.to(device),
"nav_command": nav_command.to(device),
}
targets = {k: v.to(device) for k, v in targets.items()}
return inputs, targets
def _generate_camera_images(
self, B: int, N: int, H: int, W: int, scenario: str
) -> torch.Tensor:
"""Generate synthetic camera images with road-like patterns."""
images = torch.zeros(B, N, 3, H, W)
for b in range(B):
for n in range(N):
# Sky region (top half)
sky_color = torch.tensor([0.5, 0.7, 0.9]) + torch.randn(3) * 0.05
images[b, n, :, :H//3, :] = sky_color.view(3, 1, 1)
# Road region (bottom half)
road_gray = 0.3 + torch.randn(1) * 0.05
images[b, n, :, H//3:, :] = road_gray
# Lane lines (white stripes)
lane_y = torch.arange(H//3, H)
for lane_x in [W//4, W//2, 3*W//4]:
x_start = max(0, lane_x - 2)
x_end = min(W, lane_x + 2)
images[b, n, :, H//3:, x_start:x_end] = 0.9
# Add some random "objects" (colored rectangles)
if scenario in ["urban", "intersection"]:
num_objects = np.random.randint(1, 5)
for _ in range(num_objects):
obj_h = np.random.randint(10, 40)
obj_w = np.random.randint(10, 30)
obj_y = np.random.randint(H//4, H - obj_h)
obj_x = np.random.randint(0, W - obj_w)
color = torch.rand(3)
images[b, n, :, obj_y:obj_y+obj_h, obj_x:obj_x+obj_w] = color.view(3, 1, 1)
# Add noise
images[b, n] += torch.randn_like(images[b, n]) * 0.02
return images.clamp(0, 1)
def _generate_intrinsics(self, B: int, N: int) -> torch.Tensor:
"""Generate camera intrinsic matrices from config."""
K = torch.zeros(B, N, 3, 3)
for i, cam in enumerate(self.sensor_config.cameras):
if i >= N:
break
K[:, i, 0, 0] = cam.fx
K[:, i, 1, 1] = cam.fy
K[:, i, 0, 2] = cam.cx
K[:, i, 1, 2] = cam.cy
K[:, i, 2, 2] = 1.0
return K
def _generate_extrinsics(self, B: int, N: int) -> torch.Tensor:
"""Generate camera extrinsic matrices from config."""
T = torch.zeros(B, N, 4, 4)
for i, cam in enumerate(self.sensor_config.cameras):
if i >= N:
break
T_np = cam.placement.to_transform_matrix()
T[:, i] = torch.from_numpy(T_np).float()
return T
def _generate_ultrasonic_data(
self, B: int, N: int, scenario: str
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Generate ultrasonic distance readings and placements."""
placements = torch.zeros(B, N, 6)
for i, us in enumerate(self.sensor_config.ultrasonics):
if i >= N:
break
p = us.placement
placements[:, i] = torch.tensor([p.x, p.y, p.z, p.yaw, p.pitch, p.roll])
# Generate realistic distance readings based on scenario
if scenario == "parking":
base_dist = torch.rand(B, N, 1) * 2.0 + 0.5 # 0.5-2.5m close range
elif scenario == "urban":
base_dist = torch.rand(B, N, 1) * 3.0 + 1.0 # 1-4m
elif scenario == "highway":
base_dist = torch.rand(B, N, 1) * 4.0 + 2.0 # 2-6m (clamped later)
else: # intersection
base_dist = torch.rand(B, N, 1) * 3.5 + 0.5
# Add realistic noise
noise = torch.randn_like(base_dist) * 0.01
distances = (base_dist + noise).clamp(0.02, 5.0)
return distances, placements
def _generate_ego_state(self, B: int, scenario: str) -> torch.Tensor:
"""Generate ego vehicle state [speed, accel, steer, yaw_rate, x, y]."""
states = torch.zeros(B, 6)
max_speed = self.config.max_speed_ms
if scenario == "parking":
states[:, 0] = torch.rand(B) * 2.0 # speed: 0-2 m/s
states[:, 1] = torch.randn(B) * 0.5 # accel
states[:, 2] = torch.randn(B) * 0.3 # steer
elif scenario == "highway":
states[:, 0] = torch.rand(B) * max_speed * 0.3 + max_speed * 0.7 # 70-100% max
states[:, 1] = torch.randn(B) * 0.3
states[:, 2] = torch.randn(B) * 0.05 # minimal steering
elif scenario == "intersection":
states[:, 0] = torch.rand(B) * max_speed * 0.5 # 0-50% max
states[:, 1] = torch.randn(B) * 1.0
states[:, 2] = torch.randn(B) * 0.2
else: # urban
states[:, 0] = torch.rand(B) * max_speed * 0.7 # 0-70% max
states[:, 1] = torch.randn(B) * 0.5
states[:, 2] = torch.randn(B) * 0.15
states[:, 3] = torch.randn(B) * 0.1 # yaw rate
states[:, 4] = torch.randn(B) * 10 # x position
states[:, 5] = torch.randn(B) * 5 # y position
return states
def _generate_targets(self, B: int, scenario: str) -> Dict[str, torch.Tensor]:
"""Generate all ground truth labels."""
bev = self.bev_size
targets = {}
# Object detection heatmap (10 classes)
heatmap = torch.zeros(B, 10, bev, bev)
for b in range(B):
num_obj = np.random.randint(2, 8)
for _ in range(num_obj):
cls = np.random.randint(0, 10)
cx, cy = np.random.randint(20, bev-20), np.random.randint(20, bev-20)
sigma = np.random.uniform(2, 6)
y, x = torch.meshgrid(torch.arange(bev), torch.arange(bev), indexing='ij')
gaussian = torch.exp(-((x - cx)**2 + (y - cy)**2) / (2 * sigma**2))
heatmap[b, cls] = torch.max(heatmap[b, cls], gaussian)
targets["gt_heatmap"] = heatmap
# Segmentation (7 classes)
seg = torch.zeros(B, bev, bev, dtype=torch.long)
seg[:, :, :] = 0 # background
seg[:, bev//4:3*bev//4, :] = 1 # drivable
for b in range(B):
# Lane lines
for lane_x in [bev//4, bev//2, 3*bev//4]:
seg[b, :, max(0,lane_x-1):min(bev,lane_x+1)] = 2
targets["gt_segmentation"] = seg
# Occupancy grid
occ = torch.zeros(B, 1, bev, bev)
occ[:, :, :bev//4, :] = 1.0 # obstacles at edges
occ[:, :, 3*bev//4:, :] = 1.0
targets["gt_occupancy"] = occ
# Behavior labels
if scenario == "parking":
targets["gt_behavior"] = torch.full((B,), 7, dtype=torch.long) # park
elif scenario == "highway":
targets["gt_behavior"] = torch.full((B,), 0, dtype=torch.long) # keep lane
else:
targets["gt_behavior"] = torch.randint(0, 5, (B,))
# Waypoints (20 waypoints, each with x, y, heading, speed)
wp = torch.zeros(B, 20, 4)
for t in range(20):
wp[:, t, 0] = t * 0.5 # x: forward 0.5m per step
wp[:, t, 1] = torch.randn(B) * 0.1 # y: slight lateral variation
wp[:, t, 2] = torch.randn(B) * 0.02 # heading: nearly straight
wp[:, t, 3] = self.config.max_speed_ms * 0.7 # target speed
targets["gt_waypoints"] = wp
# Control commands
targets["gt_steering"] = torch.randn(B) * 5.0 # degrees
targets["gt_throttle"] = torch.rand(B) * 0.5 + 0.2
targets["gt_brake"] = torch.zeros(B)
return targets
def __len__(self):
return 1000 # virtual dataset size
def __getitem__(self, idx):
"""Dataset-style access for DataLoader compatibility."""
inputs, targets = self.generate_batch(batch_size=1)
# Squeeze batch dim
inputs = {k: v.squeeze(0) for k, v in inputs.items()}
targets = {k: v.squeeze(0) for k, v in targets.items()}
return inputs, targets
|