Adaptive Confidence-Based Fire and Smoke Detection: A Two-Stage YOLO Framework With Selective Segmentation Validation for Real-Time Surveillance Systems

Bekbol Sultankulov, Marat Nurtas, Daniker Chepashev, Elmira Zhilkibayeva

IEEE Access, 2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.1109/access.2026.3741661
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Automated fire and smoke detection in industrial surveillance systems represents a safety-critical computer vision challenge where both missed detections and false alarms carry significant operational consequences. Single-stage YOLO-based detectors achieve competitive inference speeds but generate false positive detections at operationally disruptive rates under adverse industrial conditions, including partial occlusion, low illumination, smoke-induced visual degradation, steam emissions, and reflective surfaces characteristic of petrochemical and manufacturing environments. This paper proposes AC-YOLO, an adaptive confidence-based two-stage detection and validation framework that integrates a fine-tuned YOLOv9c primary detector with selective segmentation-based validation to suppress false positives while preserving real-time throughput. The core contribution is a three-tier confidence-routing mechanism that partitions detections into directly accepted high-confidence outputs, empirically selected intermediate-confidence detections subjected to pixel-level segmentation validation, and discarded low-confidence outputs. The intermediate confidence interval of 0.50-0.85 was identified through ablation as the stratum containing 78.3% of all false positive detections generated by the single-stage baseline. Two segmentation validation backends are evaluated within the framework: YOLO26-seg, optimized for real-time deployment, and SAM 3 (Segment Anything Model 3, Meta AI), a foundation model supporting concept-level promptable segmentation. Multi-criteria validation applies pixel occupancy ratio, shape compactness, and mask confidence scoring under a conjunctive decision rule, with each criterion contributing an independently measurable reduction in false positive rate. The Stage 2 validator is trained, and the framework evaluated, on a heterogeneous dataset of 35,412 annotated images spanning public sources and four active Kazakhstani industrial facilities. Experimental evaluation on a 3,542-image test set demonstrates that AC-YOLO with the YOLO26-seg backend achieves 88.4% precision and 74.1% recall at 26.8 FPS, reducing the false positive rate from 18.7% to 10.2% relative to the single-stage YOLOv9c baseline while maintaining throughput above the 25 FPS real-time threshold. Comparative evaluation of YOLO26-seg and SAM 3 reveals a well-defined accuracy-latency trade-off: SAM 3 achieves a lower false positive rate (7.8%) at 4.1 FPS, establishing it as appropriate for offline validation, while YOLO26-seg satisfies both accuracy and real-time constraints for live deployment. Production deployment at all four facilities over a 30-day period yielded a false alarm rate of 19.0%, representing a 50.6% relative reduction over the single-stage baseline. All fire events confirmed through operator-verified alarm logs were successfully detected during the monitored period, although this metric cannot rule out events missed by both the system and human operators.