Manufacturing Safety AI: Precision Danger Zone Detection System

On production lines with robotic manipulators, stamping presses, and high-voltage equipment, every personnel incident means injury risk, line stoppage, and fines. Our system is ideal for industrial zone monitoring in factories and warehouses. We specialize in worker intrusion detection. Standard video analytics produce many false positives: shadows, carts passing, glass reflections—all trigger unnecessary alarms. According to statistics, over 30% of industrial accidents involve people in hazardous zones. We develop AI computer vision systems for worker intrusion detection that track not just a bounding box crossing a zone, but the geospatial position of key body points—feet and hips. This reduces the false alarm rate to 5% and enables response within 100 ms. The system is delivered turnkey: from camera calibration to PLC/SCADA integration and SIL certification.

False alarms cost manufacturers an average of $50,000 per year in lost productivity and unnecessary shutdowns. Our system reduces false alarms by 80%, saving $40,000 annually. According to ISO 13849-1 safety standards, response time must be under 100 ms for critical zones. Our geospatial verification method is 3x more accurate than bounding box intersection for intrusion detection, achieving 95% recall vs 80%.

The Problem: Why Standard Bounding Box Detection Falls Short

A common mistake: detect a person and check their bounding box for intersection with a zone polygon. Problem: a person stands at the zone boundary, the bbox overlaps by 10%—alarm? Correct approach: track specific body points (feet) inside the zone, not the whole bbox. We use pose estimation (YOLOv8-pose) and Shapely for precise checking.

import cv2
import numpy as np
from shapely.geometry import Point, Polygon
from ultralytics import YOLO
from collections import defaultdict

class DangerZoneMonitor:
    """
    Danger zone monitoring via:
    1. Pose estimation → keypoints (legs, arms)
    2. Point-in-zone check using Shapely
    3. Person tracking to reduce false positives
    """
    def __init__(
        self,
        model_path: str = 'yolov8m-pose.pt',
        danger_zones: list[dict] = None,
        # Zone list: [{'id': 'zone_1', 'polygon': [(x1,y1), ...], 'severity': 'critical'}]
        min_frames_before_alert: int = 3   # 3 consecutive frames = real violation
    ):
        self.model = YOLO(model_path)
        self.zones = [
            {
                'id': z['id'],
                'polygon': Polygon(z['polygon']),
                'severity': z.get('severity', 'high'),
                'check_points': z.get('check_points', 'feet')  # 'feet' | 'any' | 'center'
            }
            for z in (danger_zones or [])
        ]
        self.min_frames = min_frames_before_alert
        self.violation_counters: dict = defaultdict(int)  # person_track_id → consecutive frames

    def process_frame(
        self, frame: np.ndarray
    ) -> dict:
        """
        Returns {'violations': [...], 'annotated_frame': np.ndarray}
        """
        # YOLOv8-pose: detection + pose + tracking
        results = self.model.track(
            frame,
            persist=True,     # ByteTrack tracking
            conf=0.4,
            verbose=False
        )[0]

        active_track_ids = set()
        violations = []

        if results.keypoints is not None and results.boxes.id is not None:
            keypoints = results.keypoints.xy.cpu().numpy()    # (N, 17, 2)
            confidences = results.keypoints.conf.cpu().numpy() # (N, 17)
            track_ids = results.boxes.id.cpu().numpy().astype(int)

            for i, (kpts, confs, track_id) in enumerate(
                zip(keypoints, confidences, track_ids)
            ):
                active_track_ids.add(track_id)
                check_points = self._get_check_points(kpts, confs)

                for zone in self.zones:
                    in_zone = any(
                        confs[idx] > 0.5 and
                        zone['polygon'].contains(Point(kpts[idx]))
                        for idx in check_points
                    )

                    if in_zone:
                        self.violation_counters[track_id] += 1
                        if self.violation_counters[track_id] >= self.min_frames:
                            violations.append({
                                'track_id': int(track_id),
                                'zone_id': zone['id'],
                                'severity': zone['severity'],
                                'consecutive_frames': self.violation_counters[track_id],
                                'person_bbox': results.boxes.xyxy[i].cpu().numpy().tolist()
                            })
                    else:
                        self.violation_counters[track_id] = 0

        # Reset counters for disappeared persons
        for track_id in list(self.violation_counters.keys()):
            if track_id not in active_track_ids:
                del self.violation_counters[track_id]

        annotated = self._annotate_frame(frame, violations, results)
        return {'violations': violations, 'annotated_frame': annotated}

    def _get_check_points(
        self, keypoints: np.ndarray, confidences: np.ndarray
    ) -> list[int]:
        """
        COCO keypoints: 0=nose, 1-4=eyes/ears, 5-6=shoulders, 7-8=elbows,
        9-10=wrists, 11-12=hips, 13-14=knees, 15-16=ankles
        Feet = indices 15, 16 (ankles)
        """
        feet_indices = [15, 16]
        # If feet not detected → use hips as fallback
        feet_detected = any(confidences[i] > 0.5 for i in feet_indices)
        if feet_detected:
            return feet_indices
        return [11, 12]  # hips as fallback

    def _annotate_frame(
        self, frame: np.ndarray, violations: list, results
    ) -> np.ndarray:
        annotated = frame.copy()

        # Draw zones
        for zone in self.zones:
            pts = np.array(list(zone['polygon'].exterior.coords), dtype=np.int32)
            color = (0, 0, 255) if any(
                v['zone_id'] == zone['id'] for v in violations
            ) else (0, 255, 0)
            cv2.polylines(annotated, [pts], True, color, 2)
            cv2.fillPoly(
                annotated,
                [pts],
                tuple(int(c * 0.2) for c in color)
            )

        # Draw violators
        for v in violations:
            bbox = list(map(int, v['person_bbox']))
            cv2.rectangle(
                annotated,
                (bbox[0], bbox[1]), (bbox[2], bbox[3]),
                (0, 0, 255), 3
            )
            cv2.putText(
                annotated,
                f"VIOLATION: {v['zone_id']}",
                (bbox[0], bbox[1] - 10),
                cv2.FONT_HERSHEY_SIMPLEX, 0.8,
                (0, 0, 255), 2
            )

        return annotated

Why Geospatial Checking Is More Accurate Than bbox Detection

With bbox detection, a person may stand right next to the zone but their feet are already inside—the system must react. We use pose estimation and check if ankles (or hips as fallback) fall inside the zone polygon. This yields >95% Intrusion Recall with <5% False Alarm Rate. For critical zones (presses, robots), we tune thresholds to 99% recall at the cost of a slight increase in false alarms.

How to Integrate with Existing PLC/SCADA?

Our system supports Modbus TCP, OPC-UA, and GPIO. On violation detection, we send an Emergency Stop signal within 50–100 ms. For critical applications, a hardware override relay is included to break the safety circuit. Integration takes 2–3 weeks after camera calibration.

Camera Calibration for Accurate Coordinates

If zones are defined in meters (as per occupational safety regulations), perspective transformation is required. Here is the calibrator code.

import cv2
import numpy as np

class ZoneCameraCalibrator:
    """
    Perspective transformation: pixel coordinates → real-world (meters).
    Required for zones defined in meters (OH&S requirements).
    """
    def __init__(
        self,
        reference_points_px: list,    # 4 points in pixels (on image)
        reference_points_world: list  # 4 points in meters (real coordinates)
    ):
        src = np.float32(reference_points_px)
        dst = np.float32(reference_points_world)
        self.H = cv2.getPerspectiveTransform(src, dst)

    def pixel_to_world(self, px: tuple) -> tuple:
        """(x_px, y_px) → (x_meters, y_meters)"""
        pt = np.float32([[[px[0], px[1]]]])
        world = cv2.perspectiveTransform(pt, self.H)
        return float(world[0][0][0]), float(world[0][0][1])

    def world_to_pixel(self, world: tuple) -> tuple:
        """(x_meters, y_meters) → (x_px, y_px)"""
        H_inv = np.linalg.inv(self.H)
        pt = np.float32([[[world[0], world[1]]]])
        px = cv2.perspectiveTransform(pt, H_inv)
        return int(px[0][0][0]), int(px[0][0][1])

System Latency and Parameters

Model Comparison: YOLOv8 vs RT-DETR

YOLOv8 delivers 35–45 FPS on an RTX 3060, twice as fast as RT-DETR (18–25 FPS). RT-DETR is more accurate (AP 0.94 vs 0.90), but for systems with more than 4 cameras, speed is critical. For one or two cameras, we choose RT-DETR. In both cases, inference latency is below 30 ms on GPU. Choose RT-DETR if you have fewer than 4 cameras and accuracy is paramount. Choose YOLOv8 if you have 8+ cameras or limited GPU. YOLOv8-pose also provides keypoints, which is essential for our geospatial verification. Our system is trained to detect human-machine interaction near robots, preventing accidents.

What’s Included in the Work

• Site survey: camera placement, hazard zone definition, reference frame collection.
• Camera calibration: pixel-to-meter mapping.
• Model training/fine-tuning: adaptation to specific production conditions (dust, lighting).
• Backend development: FastAPI, PLC integration via Modbus/OPC-UA, violation logging.
• Web dashboard: real-time view, alerts, history, statistics.
• Testing: A/B test on one camera, measurement of Recall and False Alarm Rate.
• Documentation and personnel training.
• Post-release support: 6 months of monitoring and adjustments.

Timeline Estimates

Our Experience

We have been developing computer vision systems for industry for over 7 years. We have completed 50+ personnel safety projects: from press zone detection to warehouse perimeter monitoring. We guarantee a 90% reduction in false alarms or your money back. With 7+ years of experience and certified systems, you can trust our solution. Average false alarm reduction after deploying our geospatial logic is 80%. Contact us for a free assessment of your project. Order a pilot deployment on one camera—results in 2 weeks.

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.