File size: 16,934 Bytes
cd793e6 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 | """
Multi-Modal Sensor Fusion Module
Inspired by BEVFusion and GaussianFusion architectures.
Fuses camera images and ultrasonic sensor data into a unified
Bird's Eye View (BEV) representation.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import List, Optional, Dict, Tuple
from .config import SensorConfig, CameraSensorConfig, UltrasonicSensorConfig
class CameraBackbone(nn.Module):
"""
Lightweight CNN backbone for camera feature extraction.
Extracts multi-scale features from each camera image.
Architecture inspired by EfficientNet-lite / ResNet-18 style blocks.
"""
def __init__(self, in_channels: int = 3, base_channels: int = 64):
super().__init__()
self.base_channels = base_channels
# Stage 1: Initial convolution
self.stage1 = nn.Sequential(
nn.Conv2d(in_channels, base_channels, 7, stride=2, padding=3, bias=False),
nn.BatchNorm2d(base_channels),
nn.ReLU(inplace=True),
nn.MaxPool2d(3, stride=2, padding=1),
)
# Stage 2: Feature extraction blocks
self.stage2 = self._make_stage(base_channels, base_channels * 2, num_blocks=2, stride=2)
# Stage 3
self.stage3 = self._make_stage(base_channels * 2, base_channels * 4, num_blocks=2, stride=2)
# Stage 4: Deepest features
self.stage4 = self._make_stage(base_channels * 4, base_channels * 8, num_blocks=2, stride=2)
# Feature Pyramid Network (FPN) for multi-scale fusion
self.fpn_lateral4 = nn.Conv2d(base_channels * 8, base_channels * 4, 1)
self.fpn_lateral3 = nn.Conv2d(base_channels * 4, base_channels * 4, 1)
self.fpn_lateral2 = nn.Conv2d(base_channels * 2, base_channels * 4, 1)
self.fpn_output4 = nn.Conv2d(base_channels * 4, base_channels * 4, 3, padding=1)
self.fpn_output3 = nn.Conv2d(base_channels * 4, base_channels * 4, 3, padding=1)
self.fpn_output2 = nn.Conv2d(base_channels * 4, base_channels * 4, 3, padding=1)
def _make_stage(self, in_channels, out_channels, num_blocks, stride):
layers = []
layers.append(ResBlock(in_channels, out_channels, stride))
for _ in range(1, num_blocks):
layers.append(ResBlock(out_channels, out_channels, 1))
return nn.Sequential(*layers)
def forward(self, x: torch.Tensor) -> Dict[str, torch.Tensor]:
"""
Args:
x: (B, C, H, W) camera image tensor
Returns:
Dict with multi-scale features
"""
c1 = self.stage1(x) # (B, 64, H/4, W/4)
c2 = self.stage2(c1) # (B, 128, H/8, W/8)
c3 = self.stage3(c2) # (B, 256, H/16, W/16)
c4 = self.stage4(c3) # (B, 512, H/32, W/32)
# FPN top-down pathway
p4 = self.fpn_lateral4(c4)
p3 = self.fpn_lateral3(c3) + F.interpolate(p4, size=c3.shape[2:], mode='bilinear', align_corners=False)
p2 = self.fpn_lateral2(c2) + F.interpolate(p3, size=c2.shape[2:], mode='bilinear', align_corners=False)
p4 = self.fpn_output4(p4)
p3 = self.fpn_output3(p3)
p2 = self.fpn_output2(p2)
return {"p2": p2, "p3": p3, "p4": p4}
class ResBlock(nn.Module):
"""Residual block with optional downsampling."""
def __init__(self, in_channels, out_channels, stride=1):
super().__init__()
self.conv1 = nn.Conv2d(in_channels, out_channels, 3, stride=stride, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(out_channels)
self.conv2 = nn.Conv2d(out_channels, out_channels, 3, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(out_channels)
self.shortcut = nn.Sequential()
if stride != 1 or in_channels != out_channels:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels, out_channels, 1, stride=stride, bias=False),
nn.BatchNorm2d(out_channels)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out = out + self.shortcut(x)
return F.relu(out)
class UltrasonicEncoder(nn.Module):
"""
Encodes ultrasonic sensor readings into a spatial feature representation.
Each ultrasonic sensor provides a distance reading that is mapped to a
spatial cone in BEV space.
"""
def __init__(self, num_sensors: int, hidden_dim: int = 128, bev_size: int = 200):
super().__init__()
self.num_sensors = num_sensors
self.hidden_dim = hidden_dim
self.bev_size = bev_size
# Per-sensor distance encoding
self.distance_encoder = nn.Sequential(
nn.Linear(1, 32),
nn.ReLU(),
nn.Linear(32, 64),
nn.ReLU(),
)
# Sensor placement encoding (x, y, z, yaw, pitch, roll)
self.placement_encoder = nn.Sequential(
nn.Linear(6, 32),
nn.ReLU(),
nn.Linear(32, 64),
nn.ReLU(),
)
# Combined sensor feature
self.sensor_fusion = nn.Sequential(
nn.Linear(128, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim),
)
# Project all sensor features to BEV grid
self.bev_projection = nn.Sequential(
nn.Linear(num_sensors * hidden_dim, 512),
nn.ReLU(),
nn.Linear(512, hidden_dim * (bev_size // 10) * (bev_size // 10)),
)
# Upsample to full BEV resolution
self.bev_upsample = nn.Sequential(
nn.ConvTranspose2d(hidden_dim, hidden_dim // 2, 4, stride=2, padding=1),
nn.BatchNorm2d(hidden_dim // 2),
nn.ReLU(),
nn.ConvTranspose2d(hidden_dim // 2, hidden_dim // 4, 4, stride=2, padding=1),
nn.BatchNorm2d(hidden_dim // 4),
nn.ReLU(),
nn.Conv2d(hidden_dim // 4, hidden_dim // 4, 3, padding=1),
nn.BatchNorm2d(hidden_dim // 4),
nn.ReLU(),
)
def forward(self, distances: torch.Tensor, placements: torch.Tensor) -> torch.Tensor:
"""
Args:
distances: (B, num_sensors, 1) - distance readings per sensor
placements: (B, num_sensors, 6) - sensor positions (x,y,z,yaw,pitch,roll)
Returns:
bev_features: (B, hidden_dim//4, bev_size//2~, bev_size//2~) BEV feature map
"""
B = distances.shape[0]
# Encode each sensor's distance
dist_feat = self.distance_encoder(distances) # (B, N, 64)
# Encode each sensor's position
place_feat = self.placement_encoder(placements) # (B, N, 64)
# Combine distance + placement
combined = torch.cat([dist_feat, place_feat], dim=-1) # (B, N, 128)
sensor_feat = self.sensor_fusion(combined) # (B, N, hidden_dim)
# Flatten all sensors and project to BEV
flat = sensor_feat.reshape(B, -1) # (B, N * hidden_dim)
bev_flat = self.bev_projection(flat) # (B, hidden_dim * small_h * small_w)
small_size = self.bev_size // 10
bev = bev_flat.reshape(B, self.hidden_dim, small_size, small_size)
# Upsample to larger BEV resolution
bev = self.bev_upsample(bev)
return bev
class ViewTransformer(nn.Module):
"""
Transforms camera perspective features into BEV space.
Uses Lift-Splat-Shoot (LSS) approach: predict depth distribution
per pixel, then scatter features into 3D space and collapse to BEV.
"""
def __init__(
self,
in_channels: int = 256,
num_depth_bins: int = 64,
depth_min: float = 1.0,
depth_max: float = 50.0,
bev_size: int = 200,
bev_resolution: float = 0.25, # meters per pixel
):
super().__init__()
self.in_channels = in_channels
self.num_depth_bins = num_depth_bins
self.bev_size = bev_size
self.bev_resolution = bev_resolution
# Depth distribution prediction
self.depth_net = nn.Sequential(
nn.Conv2d(in_channels, in_channels, 3, padding=1),
nn.BatchNorm2d(in_channels),
nn.ReLU(),
nn.Conv2d(in_channels, num_depth_bins, 1),
)
# Feature compression for BEV
self.feature_net = nn.Sequential(
nn.Conv2d(in_channels, in_channels // 2, 1),
nn.BatchNorm2d(in_channels // 2),
nn.ReLU(),
)
# Depth bins
self.register_buffer(
'depth_bins',
torch.linspace(depth_min, depth_max, num_depth_bins)
)
# BEV encoder after scattering
self.bev_encoder = nn.Sequential(
nn.Conv2d(in_channels // 2, in_channels // 2, 3, padding=1),
nn.BatchNorm2d(in_channels // 2),
nn.ReLU(),
nn.Conv2d(in_channels // 2, in_channels // 2, 3, padding=1),
nn.BatchNorm2d(in_channels // 2),
nn.ReLU(),
)
def forward(
self,
camera_features: torch.Tensor,
intrinsics: torch.Tensor,
extrinsics: torch.Tensor,
) -> torch.Tensor:
"""
Args:
camera_features: (B, N_cams, C, H, W) multi-camera features
intrinsics: (B, N_cams, 3, 3) camera intrinsic matrices
extrinsics: (B, N_cams, 4, 4) camera-to-ego transformation matrices
Returns:
bev: (B, C//2, bev_size, bev_size) BEV feature map
"""
B, N, C, H, W = camera_features.shape
# Reshape for batch processing
features = camera_features.reshape(B * N, C, H, W)
# Predict depth distribution
depth_logits = self.depth_net(features) # (B*N, D, H, W)
depth_probs = F.softmax(depth_logits, dim=1) # (B*N, D, H, W)
# Compress features
feat = self.feature_net(features) # (B*N, C//2, H, W)
C_out = feat.shape[1]
# Outer product: depth_probs * features -> volume
# (B*N, C_out, D, H, W)
feat_expanded = feat.unsqueeze(2) # (B*N, C_out, 1, H, W)
depth_expanded = depth_probs.unsqueeze(1) # (B*N, 1, D, H, W)
volume = feat_expanded * depth_expanded # (B*N, C_out, D, H, W)
# Simplified BEV pooling: average pool over depth and spatial dims
# In full implementation, would do proper 3D-to-BEV projection
volume = volume.reshape(B, N, C_out, self.num_depth_bins, H, W)
# Pool over depth dimension
bev_per_cam = volume.mean(dim=3) # (B, N, C_out, H, W)
# Adaptive pool each camera view to BEV size
bev_per_cam = bev_per_cam.reshape(B * N, C_out, H, W)
bev_per_cam = F.adaptive_avg_pool2d(bev_per_cam, (self.bev_size, self.bev_size))
bev_per_cam = bev_per_cam.reshape(B, N, C_out, self.bev_size, self.bev_size)
# Fuse all camera BEV views (mean fusion)
bev = bev_per_cam.mean(dim=1) # (B, C_out, bev_size, bev_size)
# Refine BEV features
bev = self.bev_encoder(bev)
return bev
class MultiModalSensorFusion(nn.Module):
"""
Main sensor fusion module that combines:
1. Multi-camera visual features (via CNN backbone + View Transformer → BEV)
2. Ultrasonic proximity features (via encoder → BEV)
Output: Unified BEV representation for downstream perception/planning.
Fully configurable for any number/placement of sensors.
"""
def __init__(
self,
sensor_config: SensorConfig,
bev_size: int = 200,
bev_resolution: float = 0.25,
camera_channels: int = 3,
backbone_base: int = 64,
bev_feature_dim: int = 256,
):
super().__init__()
self.sensor_config = sensor_config
self.bev_size = bev_size
self.bev_resolution = bev_resolution
self.bev_feature_dim = bev_feature_dim
num_cameras = sensor_config.num_cameras
num_ultrasonics = sensor_config.num_ultrasonics
# Camera processing pipeline
if num_cameras > 0:
self.camera_backbone = CameraBackbone(camera_channels, backbone_base)
self.view_transformer = ViewTransformer(
in_channels=backbone_base * 4, # FPN output channels
bev_size=bev_size,
bev_resolution=bev_resolution,
)
camera_bev_channels = backbone_base * 2 # output of view transformer
else:
self.camera_backbone = None
self.view_transformer = None
camera_bev_channels = 0
# Ultrasonic processing pipeline
if num_ultrasonics > 0:
self.ultrasonic_encoder = UltrasonicEncoder(
num_sensors=num_ultrasonics,
hidden_dim=128,
bev_size=bev_size,
)
# Get output size of ultrasonic encoder
us_bev_channels = 32 # hidden_dim // 4
else:
self.ultrasonic_encoder = None
us_bev_channels = 0
# Adaptive fusion of different sensor modalities
total_bev_channels = camera_bev_channels + us_bev_channels
self.fusion_conv = nn.Sequential(
nn.Conv2d(total_bev_channels, bev_feature_dim, 3, padding=1),
nn.BatchNorm2d(bev_feature_dim),
nn.ReLU(),
nn.Conv2d(bev_feature_dim, bev_feature_dim, 3, padding=1),
nn.BatchNorm2d(bev_feature_dim),
nn.ReLU(),
)
# Channel attention for adaptive sensor weighting
self.channel_attention = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Flatten(),
nn.Linear(bev_feature_dim, bev_feature_dim // 4),
nn.ReLU(),
nn.Linear(bev_feature_dim // 4, bev_feature_dim),
nn.Sigmoid(),
)
# Final BEV refinement with residual
self.bev_refine = nn.Sequential(
nn.Conv2d(bev_feature_dim, bev_feature_dim, 3, padding=1),
nn.BatchNorm2d(bev_feature_dim),
nn.ReLU(),
nn.Conv2d(bev_feature_dim, bev_feature_dim, 3, padding=1),
nn.BatchNorm2d(bev_feature_dim),
)
def forward(
self,
camera_images: Optional[torch.Tensor] = None,
camera_intrinsics: Optional[torch.Tensor] = None,
camera_extrinsics: Optional[torch.Tensor] = None,
ultrasonic_distances: Optional[torch.Tensor] = None,
ultrasonic_placements: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Args:
camera_images: (B, N_cams, 3, H, W)
camera_intrinsics: (B, N_cams, 3, 3)
camera_extrinsics: (B, N_cams, 4, 4)
ultrasonic_distances: (B, N_us, 1)
ultrasonic_placements: (B, N_us, 6)
Returns:
bev_features: (B, bev_feature_dim, bev_size, bev_size)
"""
bev_parts = []
# Process cameras
if self.camera_backbone is not None and camera_images is not None:
B, N, C, H, W = camera_images.shape
# Extract features for each camera
imgs = camera_images.reshape(B * N, C, H, W)
multi_scale = self.camera_backbone(imgs)
# Use p2 (highest resolution FPN output) for view transformation
cam_feat = multi_scale["p2"]
_, Cf, Hf, Wf = cam_feat.shape
cam_feat = cam_feat.reshape(B, N, Cf, Hf, Wf)
cam_bev = self.view_transformer(
cam_feat, camera_intrinsics, camera_extrinsics
)
bev_parts.append(cam_bev)
# Process ultrasonics
if self.ultrasonic_encoder is not None and ultrasonic_distances is not None:
us_bev = self.ultrasonic_encoder(ultrasonic_distances, ultrasonic_placements)
# Resize to match BEV size
us_bev = F.adaptive_avg_pool2d(us_bev, (self.bev_size, self.bev_size))
bev_parts.append(us_bev)
if len(bev_parts) == 0:
raise ValueError("No sensor data provided!")
# Concatenate all BEV features
bev_concat = torch.cat(bev_parts, dim=1)
# Fuse modalities
bev = self.fusion_conv(bev_concat)
# Channel attention
attn = self.channel_attention(bev).unsqueeze(-1).unsqueeze(-1)
bev = bev * attn
# Residual refinement
bev = bev + self.bev_refine(bev)
bev = F.relu(bev)
return bev
|