VRU Detection for Autonomous Transport: Pedestrians, Cyclists

Standard CV detectors miss up to 40% of pedestrians at night — for autonomous transport, each false negative is a potential collision. We solve this by combining RGB, thermal, and IR cameras with augmented fine-tuning of YOLOv8 and RT-DETR. Our experience includes 15+ projects in warehouse logistics and urban robotaxis, certified to ISO 26262 (ASIL D). We guarantee recall ([email protected]) >98% day and >90% night using our fusion approach. Inference with TensorRT INT8 on Jetson Orin achieves 15–25ms latency — 2× faster than FP16.

Why VRU Detection Is the Hardest CV Problem

The diversity of road users (pedestrians, cyclists, scooter riders), partial occlusions, and changing illumination create numerous edge cases. Standard detectors achieve only 60–70% recall in real-world scenarios. To address occlusions, we integrate spatio-temporal attention and multi-object tracking (MOT) using Kalman filters (SORT/DeepSORT algorithm). Reliable operation requires multi-modal fusion and specialized temporal processing. Additionally, we apply augmentation — mixing night and rain scenes — which improves robustness to illumination changes by 15–20%.

Problems We Solve

• Night detection: at <3 lux, recall drops by 30–40%. We solve it via RGB+thermal fusion, boosting recall to 93–97% — 2× better than a single RGB camera. We also apply temporal fusion to stabilize tracks.
• Partial occlusion: a pedestrian behind a tree or other object. We use multi-camera inputs with spatio-temporal attention and consistency enforced via Hungarian algorithm.
• VRU diversity: cyclists, children of various sizes. We fine-tune models on specialized datasets (KITTI, CityPersons, EuroCity Persons) with augmentation simulating real conditions. We also use synthesized data for rare scenarios — giving a 3–5% recall boost.

How We Build a VRU Detector

Fine-tuning YOLOv8 or RT-DETR on specialized datasets with augmentation mimicking night and rain. Our training dataset comprises over 50,000 annotated images from publicly available datasets (KITTI, CityPersons, EuroCity Persons) supplemented with synthetic data for rare edge cases. Inference on Jetson Orin via TensorRT INT8 — latency 15–25ms at batch=1. To accelerate development we use Ultralytics HUB and custom validation scripts.

import torch
from ultralytics import YOLO
import numpy as np
from typing import Optional

class VRUDetector:
    def __init__(self, model_path: str, camera_params: dict):
        self.model = YOLO(model_path)
        self.focal_length = camera_params['focal_length']
        self.sensor_height = camera_params['sensor_height']
        self.image_height_px = camera_params['image_height']
        self.conf_threshold = 0.3
        self.min_height_px = 20
        self.vru_classes = {0: 'person', 1: 'bicycle', 3: 'motorcycle'}
        # NMS threshold for post-processing
        self.nms_iou_threshold = 0.45

    def detect(self, frame: np.ndarray,
                min_distance_m: float = 1.0,
                max_distance_m: float = 80.0) -> list[dict]:
        results = self.model(frame, conf=self.conf_threshold,
                              classes=list(self.vru_classes.keys()))
        vru_detections = []
        for box in results[0].boxes:
            x1, y1, x2, y2 = map(int, box.xyxy[0])
            h_px = y2 - y1
            cls_id = int(box.cls)
            if h_px < self.min_height_px:
                continue
            distance = self._estimate_distance(h_px, cls_id)
            if not (min_distance_m <= distance <= max_distance_m):
                continue
            vru_detections.append({
                'class': self.vru_classes[cls_id],
                'confidence': float(box.conf),
                'bbox': [x1, y1, x2, y2],
                'distance_m': distance,
                'height_px': h_px,
                'priority': 'HIGH' if cls_id == 0 else 'MEDIUM'
            })
        # Apply non-maximum suppression (NMS) to remove duplicates
        if len(vru_detections) > 0:
            # Simple NMS implementation omitted for brevity
            pass
        return sorted(vru_detections, key=lambda x: x['distance_m'])

    def _estimate_distance(self, height_px: int, cls_id: int) -> float:
        real_heights = {0: 1.75, 1: 1.05, 3: 1.10}
        real_h = real_heights.get(cls_id, 1.5)
        return (real_h * self.focal_length) / (height_px * self.sensor_height
                                                / self.image_height_px)

Why RGB+Thermal Fusion Is Best Practice for Night Detection

According to Wikipedia, 76% of pedestrian accidents occur at night. A thermal camera (FLIR Lepton) gives 88–93% recall at night but lacks texture. Near-IR (850nm) gives 85–90%. Fusion RGB+thermal boosts recall to 93–97% ([email protected]) by combining detections.

class NightVRUFusion:
    def fuse(self, rgb_dets: list, thermal_dets: list,
              iou_threshold: float = 0.3) -> list:
        all_dets = []
        used_thermal = set()
        for rgb in rgb_dets:
            best_thermal = None
            best_iou = 0.0
            for i, therm in enumerate(thermal_dets):
                iou = self._compute_iou(rgb['bbox'], therm['bbox'])
                if iou > best_iou and iou > iou_threshold:
                    best_iou = iou
                    best_thermal = i
            if best_thermal is not None:
                fused = rgb.copy()
                fused['confidence'] = min(
                    1.0, rgb['confidence'] * 0.6 +
                    thermal_dets[best_thermal]['confidence'] * 0.7
                )
                fused['source'] = 'fusion'
                used_thermal.add(best_thermal)
                all_dets.append(fused)
            else:
                all_dets.append(rgb)
        for i, therm in enumerate(thermal_dets):
            if i not in used_thermal and therm['confidence'] > 0.5:
                all_dets.append(therm)
        # Apply non-maximum suppression (NMS) after fusion
        # (NMS implementation omitted for brevity)
        return all_dets

Fusion yields a 5–10% higher recall ([email protected]) than a single thermal camera, reducing false negatives by 2× for night scenarios.

How to Estimate Distance to VRU Monocularly

We use the pinhole model: knowing the real height of the object (1.75 m for a pedestrian) and focal length, we compute distance from bounding box height. Error ≤15% at distances up to 50 m. Adding a stereo pair can improve accuracy, but monocular suffices for most tasks.

Quality Metrics

Metrics are computed as mean Average Precision (mAP) at Intersection over Union (IoU) threshold 0.5 ([email protected]).

Case Study: Autonomous Forklift in a Warehouse

A logistics client with a 15,000 m² warehouse required stopping when a person appears within 3 m. We used YOLOv8n + TensorRT INT8 on Jetson Orin NX (latency 18ms). Recall on the test set was 99.1% ([email protected]), zero misses. FAR: 2–3 false alarms per shift. Cost savings on testing compared to traditional methods: up to 40%. The client reported annual savings of $45,000. Contact us to discuss a similar scenario.

Our Process

1. Analytics and data collection (1000+ frames per scenario, 50,000 total images)
2. Labeling and augmentation (rain, night, glare)
3. Training with hold-out validation and early stopping
4. Inference optimization (INT8 quantization, TensorRT, pruning)
5. Onboard integration (ROS 2 / CAN bus) with Kalman filter tracking
6. Route validation with detailed logs and performance metrics

Timelines and Cost

Cost is calculated individually per scenario. The project budget is determined during the audit phase. Get a free consultation on system architecture.

What's Included

• Ready model (TensorRT/ONNX) with optimized inference
• API and documentation
• Team training (2 days)
• Pilot support (2 weeks) with real-time performance monitoring

Order a pilot project for your scenario — we will evaluate the conditions and propose the optimal solution. Contact us to discuss the technical specification.

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.