AI-Powered Video Smoke and Fire Detection System

Traditional smoke detectors respond to smoke within a radius of 5–7 meters. In open areas—warehouses without overhead sensors, forests, industrial sites—they are useless. Video analytics detects smoke and flames at distances up to 300+ meters, often before the concentration reaches the trigger threshold of an ionization detector. False alarms are the main pain point: without temporal filtering, FAR can reach 15–30%, discrediting the system. We have been developing such systems for over 5 years and have solved this problem.

Why Temporal Analysis Is Critical for AI Smoke Detection

Smoke is an amorphous object without clear boundaries, changing shape every 2–3 frames. Artifacts such as steam, fog, dust, or headlight glare resemble smoke/fire for a frame-by-frame classifier. Temporal analysis processes a sequence of frames and distinguishes smoke from fog based on motion patterns. This reduces FAR by 10–15 times compared to frame-by-frame detection—a significant improvement over baseline approaches.

How Temporal Filtering Works for Video Fire Detection

We use YOLOv8m/l fine-tuned on a mixed dataset (public + our own from the site). Temporal window: 8 frames + optical flow check for smoke. Steps to deploy:

1. Collect 500–800 site images (including hard negatives).
2. Fine-tune YOLO model on your data.
3. Set temporal window (8–12 frames) and confirm threshold.
4. Integrate with fire panel via API or relay.

Inference code:

import torch
import torch.nn as nn
from ultralytics import YOLO
import numpy as np
from collections import deque

class SmokeFireDetector:
    def __init__(self, model_path: str, temporal_window: int = 8):
        self.detector = YOLO(model_path)  # fine-tuned on smoke/fire
        self.temporal_window = temporal_window

        # Frame buffer for temporal analysis
        self.frame_buffer: deque = deque(maxlen=temporal_window)
        self.detection_history: dict = {}  # track_id -> history

        # Minimum frames with detection for alarm
        self.confirm_frames = 5  # out of 8 frame window

    def _optical_flow_check(self, prev_frame, curr_frame,
                             bbox: list) -> float:
        """Smoke moves chaotically—check flow irregularity"""
        x1, y1, x2, y2 = bbox
        prev_roi = cv2.cvtColor(prev_frame[y1:y2, x1:x2], cv2.COLOR_BGR2GRAY)
        curr_roi = cv2.cvtColor(curr_frame[y1:y2, x1:x2], cv2.COLOR_BGR2GRAY)

        flow = cv2.calcOpticalFlowFarneback(
            prev_roi, curr_roi, None,
            pyr_scale=0.5, levels=3, winsize=15,
            iterations=3, poly_n=5, poly_sigma=1.1, flags=0
        )
        magnitude = np.sqrt(flow[..., 0]**2 + flow[..., 1]**2)
        # Smoke: uneven flow, high std
        return float(magnitude.std())

    def detect(self, frame: np.ndarray) -> list[dict]:
        self.frame_buffer.append(frame.copy())
        results = self.detector(frame, conf=0.35, iou=0.4)
        confirmed_events = []

        for box in results[0].boxes:
            cls_id = int(box.cls)
            cls_name = self.detector.model.names[cls_id]
            if cls_name not in ['smoke', 'fire']:
                continue

            bbox = list(map(int, box.xyxy[0]))
            conf = float(box.conf)

            # Temporal confirmation
            det_key = f"{cls_name}_{bbox[0]//50}_{bbox[1]//50}"  # grid cell
            if det_key not in self.detection_history:
                self.detection_history[det_key] = deque(maxlen=self.temporal_window)
            self.detection_history[det_key].append(conf)

            confirmed_count = sum(1 for c in self.detection_history[det_key]
                                   if c > 0.3)

            if confirmed_count >= self.confirm_frames:
                # Additional optical flow check for smoke
                flow_score = 0.0
                if cls_name == 'smoke' and len(self.frame_buffer) >= 2:
                    flow_score = self._optical_flow_check(
                        self.frame_buffer[-2], frame, bbox
                    )

                confirmed_events.append({
                    'class': cls_name,
                    'confidence': conf,
                    'temporal_score': confirmed_count / self.temporal_window,
                    'flow_irregularity': flow_score,
                    'bbox': bbox,
                    'alert': True
                })

        return confirmed_events

The optical flow method is described in OpenCV documentation.

Metrics and Thresholds

Our tests show: temporal filtering with an 8-frame window improves Recall by 8–10% and reduces FAR by 10–15 times compared to frame-by-frame detection. This means our method is up to 15 times better than standard approaches in false alarm reduction.

How On-Site Fine-Tuning Improves AI Smoke and Fire Detection

Public datasets (FireNet, MIVIA Fire, VisiFire, D-Fire) contain 11k+ images, but in production, fine-tuning on local data is always needed. Typical procedure: collect 500–800 images from the site (200 smoke/fire + 300–600 hard negatives — steam, fog, sunset), fine-tune YOLO v8m with learning_rate=0.001, 50 epochs, augmentation HSV + geometry. This reduces FAR by 3–5 times on the specific site.

Case Study: Petrochemical Plant (Our Client)

The site is an open area of 4 hectares, 18 PTZ cameras with IR. Task: early detection of tank fire.

• Base model: YOLOv8l, fine-tuned on 1200 site images
• Temporal window: 10 frames @ 10fps = 1 second
• Confirm threshold: 6 out of 10 frames

Test results (15 staged fires):

• Recall: 100% (all 15 detected)
• Average detection time: 4.2 seconds from ignition
• FAR over 2 weeks: 1 false alarm (sunset + boiler steam)

Before our system, the site experienced 10–15 false fire brigade calls per month, each costing the enterprise an average of $5,000–10,000. After implementation, FAR dropped to 1 per shift, saving the client over $300,000 per year. Infrastructure: server with RTX 3090, 18 streams @ 10fps, latency 180ms. Integration with Notifier fire panel via Modbus TCP.

With over 5 years of experience and 20+ deployed projects, we deliver robust solutions for industrial safety.

Integration with Fire Systems

• Protocols: Modbus TCP/RTU, BACnet, OPC-UA — for direct integration with fire panels
• VMS: record video evidence 60 seconds before and after the event
• Event geolocation: bind bbox to site map via camera calibration

Get a consultation on integration with your equipment.

What Is Included in Turnkey Development? (Deliverables)

Our deliverables include:

• Trained model with model card (architecture, validation metrics, limitations)
• Inference source code (Python, documentation)
• REST API for integration with your systems
• Ready-made module for VMS (Milestone, Genetec, or any with RTSP)
• Operator manual and staff training
• 6-month warranty on model adaptation to scene changes (seasonal variations, new smoke sources)

Typical project costs range from $15,000 for small setups to $100,000 for enterprise solutions. We provide a fixed-price quote after initial assessment.

Development Timelines

Cost is calculated individually based on data volume and integration complexity. To evaluate your project, write to us — we will analyze your cameras and conditions within 1 business day.

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.