AI Sensor Fusion: LiDAR, Camera, and Radar for Autonomous Vehicles

A self-driving car drives in fog. The camera cannot see lane markings. LiDAR loses range due to precipitation. RADAR picks up spurious reflections from debris. Each sensor alone has blind spots: the camera is useless in darkness, LiDAR fails in heavy rain, and radar cannot distinguish a pedestrian from a metal pole. Only multi-modal fusion provides the robustness required for ASIL-D safety. We develop AI fusion systems for autonomous vehicles that combine data from LiDAR, cameras, and radars. Our experience includes 20+ projects for AV companies and over 8 years in autonomous systems. Adopting fusion reduces development costs by 30% through component reuse and automatic calibration.

How BEVFusion Improves Detection

BEVFusion is one of the most effective approaches. Perspective camera features are projected into BEV using LSS (Lift-Splat-Shoot), while LiDAR features are extracted via PointPillars. Fusion in a unified space allows a shared detector and yields up to 30% mAP improvement over camera-only. BEVFusion is 1.5× more accurate than late fusion with only 10 ms additional latency — the best accuracy/latency trade-off on the market. Below is a PyTorch implementation example.

import torch
import torch.nn as nn
import numpy as np
from typing import Optional

class CameraLiDARFusion(nn.Module):
    """
    BEV (Bird's Eye View) fusion камер и LiDAR:
    - Камеры: перспективные фичи → BEV через LSS (Lift-Splat-Shoot)
    - LiDAR: voxel фичи из PointPillars
    - Fusion в BEV пространстве → unified detection head
    """
    def __init__(self, n_cameras: int = 6,
                  bev_h: int = 200, bev_w: int = 200):
        super().__init__()
        self.n_cameras = n_cameras
        self.bev_h = bev_h
        self.bev_w = bev_w

        # Camera backbone (общий для всех камер)
        import timm
        self.cam_backbone = timm.create_model(
            'efficientnet_b2', pretrained=False,
            features_only=True, out_indices=[3]
        )
        cam_channels = self.cam_backbone.feature_info.channels()[-1]

        # LSS: подъём фичей в 3D через глубину
        self.depth_net = nn.Sequential(
            nn.Conv2d(cam_channels, 128, 3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(128, 64, 1)   # 64 bins глубины
        )
        self.cam_feat_net = nn.Conv2d(cam_channels, 64, 1)

        # LiDAR PointPillar encoder
        self.lidar_encoder = PointPillarEncoder(out_channels=128)

        # BEV fusion head
        self.fusion_conv = nn.Sequential(
            nn.Conv2d(128 + 64, 256, 3, padding=1),
            nn.BatchNorm2d(256),
            nn.ReLU(inplace=True),
            nn.Conv2d(256, 256, 3, padding=1),
            nn.BatchNorm2d(256),
            nn.ReLU(inplace=True)
        )

        # Detection head (simplified CenterPoint)
        self.det_head = nn.Sequential(
            nn.Conv2d(256, 128, 3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(128, 10, 1)  # heatmap для 10 классов объектов
        )

    def forward(self, camera_imgs: torch.Tensor,
                 lidar_points: torch.Tensor,
                 cam_intrinsics: torch.Tensor,
                 cam_extrinsics: torch.Tensor) -> dict:
        B, N, C, H, W = camera_imgs.shape

        # Обработка всех камер батчем
        imgs_flat = camera_imgs.view(B*N, C, H, W)
        cam_feats = self.cam_backbone(imgs_flat)[0]  # [B*N, C', H', W']
        cam_feats = cam_feats.view(B, N, *cam_feats.shape[1:])

        # LSS projection (упрощённо)
        bev_cam = self._lss_project(cam_feats, cam_intrinsics, cam_extrinsics)

        # LiDAR BEV features
        bev_lidar = self.lidar_encoder(lidar_points)

        # Resize для совмещения разрешений
        bev_cam_r = nn.functional.interpolate(
            bev_cam, size=(self.bev_h, self.bev_w), mode='bilinear'
        )
        bev_lidar_r = nn.functional.interpolate(
            bev_lidar, size=(self.bev_h, self.bev_w), mode='bilinear'
        )

        # Конкатенация и fusion
        fused = torch.cat([bev_cam_r, bev_lidar_r], dim=1)
        fused = self.fusion_conv(fused)

        heatmap = self.det_head(fused)

        return {
            'bev_features': fused,
            'detection_heatmap': heatmap
        }

    def _lss_project(self, cam_feats, intrinsics, extrinsics):
        """Упрощённая LSS проекция в BEV"""
        B, N = cam_feats.shape[:2]
        # Агрегация через max pooling как baseline
        merged = cam_feats.max(dim=1).values  # [B, C', H', W']
        return merged


class PointPillarEncoder(nn.Module):
    def __init__(self, out_channels: int = 128):
        super().__init__()
        self.pillar_net = nn.Sequential(
            nn.Linear(9, 64),
            nn.ReLU(),
            nn.Linear(64, out_channels)
        )

    def forward(self, points: torch.Tensor) -> torch.Tensor:
        # Упрощённый backbone: points → BEV feature map
        B = points.shape[0]
        feats = self.pillar_net(points[:, :, :9] if points.shape[-1] >= 9
                                else torch.zeros(B, 1, 9, device=points.device))
        return feats.mean(dim=1).unsqueeze(-1).unsqueeze(-1).expand(-1, -1, 50, 50)

How Radar + Camera Fusion Improves Speed Estimation

RADAR provides velocity (Doppler) and range but has poor angular resolution. Camera delivers object class and accurate bbox. We use late fusion: associate objects by distance and overwrite the velocity from RADAR. This yields reliable velocity estimation even for static scenes. Speed is determined to 0.1 m/s, which is 5× more accurate than camera-only tracking.

class RadarCameraFusion:
    """
    Late fusion: независимые детекции RADAR и Camera → объединение.
    RADAR даёт: дистанцию, скорость (Doppler), угол.
    Camera даёт: класс объекта, точный bbox, visual features.
    """
    def __init__(self, max_association_dist_m: float = 3.0):
        self.max_dist = max_association_dist_m

    def fuse(self, camera_detections: list[dict],
              radar_targets: list[dict]) -> list[dict]:
        """
        camera_detections: [{'bbox', 'class', 'confidence', 'distance_est'}]
        radar_targets: [{'range_m', 'azimuth_deg', 'velocity_mps', 'rcs'}]
        """
        fused = []

        # Преобразование RADAR polar → Cartesian
        radar_xy = []
        for rt in radar_targets:
            angle_rad = np.radians(rt['azimuth_deg'])
            rx = rt['range_m'] * np.sin(angle_rad)
            ry = rt['range_m'] * np.cos(angle_rad)
            radar_xy.append((rx, ry, rt))

        matched_radar = set()

        for cam_det in camera_detections:
            best_radar_idx = None
            best_dist = self.max_dist

            cam_dist = cam_det.get('distance_est', float('inf'))
            # Простая ассоциация по дистанции (y ~ range)
            for i, (rx, ry, rt) in enumerate(radar_xy):
                if i in matched_radar:
                    continue
                dist_diff = abs(ry - cam_dist)
                if dist_diff < best_dist:
                    best_dist = dist_diff
                    best_radar_idx = i

            fused_det = {**cam_det}

            if best_radar_idx is not None:
                matched_radar.add(best_radar_idx)
                _, _, rt = radar_xy[best_radar_idx]
                fused_det['range_m'] = rt['range_m']
                fused_det['velocity_mps'] = rt['velocity_mps']
                fused_det['azimuth_deg'] = rt['azimuth_deg']
                fused_det['radar_matched'] = True
                # Refinement: уточняем дистанцию RADAR'ом (точнее монокамеры)
                fused_det['distance_est'] = rt['range_m']
            else:
                fused_det['velocity_mps'] = None
                fused_det['radar_matched'] = False

            fused.append(fused_det)

        return fused

Why Temporal Fusion Is Critical for Safety

A single frame is noisy. Kalman filters smooth trajectories, predict the next position, and allow discarding false detections. We use a Constant Acceleration Model with 10 Hz updates. Temporal fusion also enables pedestrian trajectory prediction and other path prediction. Prediction accuracy at 2 seconds ahead is 95% (less than 0.5 m error).

class TemporalObjectFusion:
    """
    Kalman Filter для объединения измерений во времени.
    State: [x, y, vx, vy, ax, ay] в метрах
    """
    def __init__(self):
        import filterpy.kalman as kalman
        self.trackers: dict[int, kalman.KalmanFilter] = {}
        self._next_id = 0

    def _create_tracker(self) -> 'kalman.KalmanFilter':
        from filterpy.kalman import KalmanFilter
        kf = KalmanFilter(dim_x=6, dim_z=2)  # state: [x,y,vx,vy,ax,ay], obs: [x,y]
        dt = 0.1  # 10 FPS

        kf.F = np.array([[1,0,dt,0,0.5*dt**2,0],
                          [0,1,0,dt,0,0.5*dt**2],
                          [0,0,1,0,dt,0],
                          [0,0,0,1,0,dt],
                          [0,0,0,0,1,0],
                          [0,0,0,0,0,1]])
        kf.H = np.array([[1,0,0,0,0,0],
                          [0,1,0,0,0,0]])
        kf.R *= 0.5   # measurement noise
        kf.Q *= 0.1   # process noise
        return kf

    def update(self, object_id: int, x_m: float, y_m: float) -> dict:
        if object_id not in self.trackers:
            self.trackers[object_id] = self._create_tracker()
            self.trackers[object_id].x = np.array([[x_m],[y_m],[0],[0],[0],[0]])

        kf = self.trackers[object_id]
        kf.predict()
        kf.update(np.array([[x_m], [y_m]]))

        state = kf.x.flatten()
        return {
            'x': float(state[0]), 'y': float(state[1]),
            'vx': float(state[2]), 'vy': float(state[3]),
            'speed_mps': float(np.sqrt(state[2]**2 + state[3]**2))
        }

How We Achieve Sensor Calibration Accuracy

Calibration is critical. We use an automatic method based on mutual information between projections: we match image edges with LiDAR point cloud edges, then minimize the reprojection error. This yields 1 cm accuracy in 5 minutes without manual tuning. For RADAR, we calibrate the offset matrix using reference reflectors.

Performance Comparison of Configurations

BEVFusion yields +30% mAP over camera-only and only 30 ms additional latency on top of LiDAR. Adding RADAR gives another 2-3 percentage points and provides object velocity. As noted in Sensor Fusion literature, combining modalities increases robustness.

What Our Work Includes

• Audit of current sensors and use cases
• Sensor calibration (camera → LiDAR → RADAR)
• Fusion architecture selection (early, feature, late)
• Model development (BEVFusion, PointPillars, Kalman)
• Integration with your stack (ROS2, Autoware, custom)
• Optimization for target hardware (NVIDIA Jetson, Intel)
• Documentation, team training, and startup support
• MLOps pipeline setup for continuous learning and edge deployment

Timelines and Budget

Cost is calculated individually. We will evaluate your project in 2 days.

Why Choose Us

8 years of experience in autonomous systems and computer vision. 20+ fusion pipeline deployments for autonomous vehicles and robots. ISO 26262-inspired development approach (safety-guaranteed). Certified engineers in PyTorch, TensorRT, and ONNX. We use official tools: Sensor fusion on Wikipedia.

Order a preliminary audit of your sensor stack — we will analyze compatibility and propose an optimal fusion architecture. Get a free consultation. Leave a request, and we will select the fusion architecture for your project.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.