Custom AI ADAS Module Development: FCW, LDW, BSM

Ever had the autopilot brake too late on the highway, or mistake a pedestrian for a tree? We solve these problems with custom neural networks on embedded controllers. Over the last 7 years we've delivered 15+ ADAS functions—from AEB to blind spot monitoring—that work in real conditions: rain, snow, poor lane markings. Below is the concrete stack and architecture.

Each function is a separate computer vision problem with tight latency and accuracy requirements. For AEB, the delay must not exceed 30 ms—at 100 km/h, the car covers an extra 2.8 meters every 30 ms.

We use proven models: YOLOv8n for detection, CLRNet for lanes, Depth Anything for monocular depth. All wrapped into TensorRT or ONNX Runtime for inference on NVIDIA Orin and Qualcomm Snapdragon.

The main challenge is balancing accuracy and speed, especially when running 4–5 functions concurrently. On one project we cut AEB latency from 45 ms to 22 ms by replacing the backbone with EfficientNet-lite and switching to TensorRT INT8.

Key Functions and Implementation

import cv2
import numpy as np
from ultralytics import YOLO
import torch

class ADASSystem:
    def __init__(self, config: dict):
        self.lane_detector = self._load_lane_model(config)
        self.object_detector = YOLO(config['object_model'])  # YOLOv8n для скорости
        self.depth_estimator = self._load_depth_model(config)

        self.camera_matrix = np.array(config['camera_intrinsics'])
        self.focal_length = self.camera_matrix[0, 0]
        self.baseline = config.get('stereo_baseline', None)

        # Пороги для предупреждений
        self.ttc_warning = 2.5   # секунды — предупреждение
        self.ttc_critical = 1.5  # секунды — AEB
        self.lane_offset_threshold = 0.3  # метра

    def lane_departure_warning(self, frame: np.ndarray,
                                vehicle_speed: float) -> dict:
        """
        Детекция полосы: классика — UFLD (Ultra-Fast Lane Detection)
        или CLRNet для сложных условий (пересечения, плохая разметка).
        """
        lanes = self.lane_detector(frame)
        if len(lanes) < 2:
            return {'warning': False, 'reason': 'no_lanes'}

        # Центр автомобиля относительно полосы
        frame_center = frame.shape[1] // 2
        lane_center = (lanes[0][-1][0] + lanes[1][-1][0]) // 2
        offset_px = frame_center - lane_center

        # Перевод пикселей в метры через гомографию
        offset_m = offset_px * (3.5 / abs(lanes[1][-1][0] - lanes[0][-1][0]))

        warning = abs(offset_m) > self.lane_offset_threshold
        return {
            'warning': warning,
            'offset_meters': offset_m,
            'lane_width': abs(lanes[1][-1][0] - lanes[0][-1][0])
        }

    def collision_warning(self, frame: np.ndarray,
                           ego_speed: float) -> dict:
        detections = self.object_detector(frame, conf=0.5,
                                           classes=[0, 2, 3, 5, 7])
        depth_map = self.depth_estimator(frame)

        warnings = []
        for box in detections[0].boxes:
            x1, y1, x2, y2 = map(int, box.xyxy[0])
            cx = (x1 + x2) // 2

            # Дистанция из depth map (стерео или монокулярная)
            roi_depth = depth_map[y1:y2, x1:x2]
            distance = float(np.percentile(roi_depth, 10))  # ближняя часть объекта

            # TTC при текущей скорости
            if distance > 0 and ego_speed > 0:
                ttc = distance / ego_speed  # упрощённо, без учёта скорости объекта
            else:
                ttc = float('inf')

            if ttc < self.ttc_critical:
                action = 'AEB'
            elif ttc < self.ttc_warning:
                action = 'WARNING'
            else:
                continue

            warnings.append({
                'class': self.object_detector.model.names[int(box.cls)],
                'distance_m': distance,
                'ttc_sec': ttc,
                'action': action
            })

        return {'warnings': sorted(warnings, key=lambda x: x['ttc_sec'])}

How AEB Works

When an object is detected at less than critical distance (TTC < 1.5 s), the system activates braking. We combine stereo vision and monocular depth to estimate distance. In practice, at 60 km/h, distance accuracy is ±1.5 m at 20 m range.

Why Latency Matters More Than Accuracy

At 100 km/h, the car travels 27.8 m per second. A 100 ms system delay means 2.78 m of blind travel. Therefore for ADAS:

YOLOv8n with TensorRT FP16 on NVIDIA Orin: 3–5 ms per frame. On Qualcomm SA8295P (Snapdragon Ride): 8–12 ms via QNN SDK.

Source: ISO 26262, AEB response time requirements

Model Comparison: YOLOv8n vs YOLOv8s

For AEB, YOLOv8n is better: its latency on TensorRT INT8 is 5–8 ms on NVIDIA Orin, 2.5 times faster than YOLOv8s. Meanwhile, mAP (0.5) drops from 0.52 to 0.48—negligible in practice. For BSW we use YOLOv8s (latency up to 100 ms).

Monocular Depth Estimation

If no stereo camera is available, we use MonoDepth2 or DPT (Dense Prediction Transformer). Accuracy is lower than stereo but sufficient for warnings:

from transformers import AutoImageProcessor, AutoModelForDepthEstimation

class MonocularDepth:
    def __init__(self):
        self.processor = AutoImageProcessor.from_pretrained(
            "LiheYoung/depth-anything-large-hf"
        )
        self.model = AutoModelForDepthEstimation.from_pretrained(
            "LiheYoung/depth-anything-large-hf"
        )

    @torch.no_grad()
    def estimate(self, image: np.ndarray) -> np.ndarray:
        inputs = self.processor(images=image, return_tensors="pt")
        outputs = self.model(**inputs)
        depth = outputs.predicted_depth.squeeze().numpy()
        # Масштабируем в метры через калибровочный коэффициент
        return depth

Depth Anything v2 Large gives AbsRel = 0.076 on KITTI—sufficient for distance estimation within ±10% at 10–30 m.

How We Implement the ADAS System

1. Analyze requirements and select sensors (cameras, radars, lidars).
2. Collect and annotate data for target scenarios (urban, highway, night).
3. Train models with fine-tuning and quantization (INT8/FP16).
4. Deploy on target hardware (NVIDIA Orin, Qualcomm Snapdragon, TI TDA4).
5. Validate offline on datasets and online in real conditions.
6. Document and hand over to the client.

What Is Included in the Project

• System architecture and sensor selection (cameras, radars, lidars)
• Datasets: collection, annotation, augmentation
• Model training: fine-tuning, quantization (INT8/FP16), pruning
• Deployment on target hardware (NVIDIA Orin, Qualcomm Snapdragon, TI TDA4)
• Validation: offline (on datasets) and online (real scenarios)
• Documentation: functional spec, test plan, CI/CD pipeline
• Client team training

Certification and Standards

ADAS systems for production vehicles must comply with:

• ISO 26262 (Functional Safety, ASIL-B/C for AEB)
• ISO/SAE 21434 (Cybersecurity)
• UNECE R79/R130 (regulatory requirements for LDW and AEB)

For in-plant or carport applications (not public road), requirements are softer—we use automotive grade without full ISO 26262 certification.

Timelines and Investment

Development of a single module (e.g., AEB) requires a certain investment; contact us for a tailored estimate. Order a pilot project for 1–2 functions — turnaround 6–8 weeks.

Our Experience

Over 7 years in embedded vision, 15+ ADAS projects (from prototypes to pre-series). We use automotive-grade approaches: ROS 2, DDS, SafeRTOS.

Why Choose Us

We don't just train a model—we take it to production on your specific hardware, accounting for thermal profiles and energy budget.

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.