Source-linked AI summary
Vision-enabled detection of safety helmet compliance in construction zones
Tri Nhut Do*, Ba Loc Pham
TL;DR
Construction sites need effective monitoring of helmet wearing; this paper combines YOLO-based face and helmet recognition with coordinate compatibility checks to identify workers without helmets. The system demonstrated processing at 20 to 60 frames per second, and its training metrics improved as loss values decreased.
Problem
The paper seeks to improve monitoring of helmet wearing on construction sites to help minimise workers’ head injuries.
Source: §1 · p. 2Method
Results
20 to 60 frames per second: the system demonstrated real-time processing, while training metrics increased and loss values decreased.
Source: §3.1.2 · pp. 6–7, §4.3 · p. 9Takeaways & Limitations
Depending on configuration, the alert module can direct detection information to site supervisors, safety officers, or emergency response teams.
Source: §3 · p. 4The study uses two output classes: helmet and no-helmet.
Source: §3.1.1 · p. 5
Abstract
from arXiv · showhide
In the rapidly evolving field of construction management, worker safety remains a top priority. This paper introduces an innovative vision-based system for real-time detection of helmet compliance, specifically designed for construction sites, utilizing advanced computer vision techniques and machine learning algorithms within the YOLO (you only look once) framework. Our system leverages high-resolution video feeds from strategically positioned cameras to monitor adherence to safety regulations regarding helmet usage. By employing deep learning methodologies, the system effectively identifies individuals not wearing helmets, thereby significantly mitigating the risk of head injuries among workers. Our training and validation results revealed an impressive precision exceeding 97% at mAP@0.5 for both helmeted and non-helmeted individuals. Furthermore, our experiments demonstrate exceptional detection accuracy, demonstrating the system's resilience under varying lighting conditions and diverse worker movements. The consistent decrease in loss and improvement in metrics throughout training validates the effectiveness of the YOLOv8 model in enhancing recognition performance. The implications of this research extend beyond mere regulatory compliance, opening avenues for innovative applications in occupational safety management. This study highlights the critical role of technology in protecting lives and lays the groundwork for future advancements in smart construction environments.
1. Introduction
The paper proposes a vision-based system to detect construction workers without helmets, with real-time monitoring as its aim. It uses YOLO-based face and helmet recognition and checks whether their detected coordinates are compatible.
1. Introduction
- Zhang et al. [3] explored PPE compliance detection on construction sites, but their video-analysis framework lacked real-time processing.The paper presents real-time processing as important in dynamic environments.Source: §1 · p. 1
- Qian et al. [4] implemented real-time helmet detection, while variable lighting and occlusions remained challenges for their method.Source: §1 · p. 2
- The proposed system aims to support real-time alerts and diverse-condition detection, alongside enforcement, education, and incentives for helmet use.Source: §1 · p. 2
- A face detection that cannot form a compatible coordinate pair with a helmet detection is classified as a person without a helmet.The system uses separate YOLO-based face and helmet recognition algorithms.Source: §1 · p. 2
- YOLO extracts image features with a CNN, predicts bounding boxes and class probabilities, then applies non-maximum suppression to remove overlaps.Source: §2 · p. 2
2. Background review of recognition algorithms
The background review outlines the stages of face and helmet recognition algorithms and places them among a range of construction-site safety monitoring approaches.
2. Background review of recognition algorithms
- Face detection proceeds through CNN feature extraction, detection with non-maximum suppression, and post-processing using landmarks, alignment, and recognition.Source: §2 · pp. 2–3
- Helmet detection uses CNN feature extraction and object detection, followed by post-processing techniques including classification, pose estimation, and tracking.Source: §2 · p. 3
- Construction-site detection approaches range from manual checks to advanced automated systems.Source: §2 · p. 3
3. System design
The system uses video-based YOLO processing to identify workers without helmets and route detections to alerts and supervisors. Its YOLOv8 evaluation reports an F1 score of 0.94 at a confidence threshold of 0.444 and precision of 1.00 at 0.911, alongside improving metrics and decreasing losses during training.
3. System design
- The camera-to-alert workflow processes live video with YOLO-based detection and can notify supervisors or safety officers when non-helmet wearing is detected.The output supervisor block is triggered when non-helmet wearing is confirmed.Source: §3 · pp. 3–4
3.1. Helmet and non-helmet wearing detection processing design
- The proposed YOLOv8 design detects two output classes, helmet and no_helmet, and produces object locations and class labels.Its described architecture includes a feature-extracting backbone, a feature-aggregation neck, and multi-scale detection levels.Source: §3.1 · p. 4
3.1.1. Data capturing and labelling
- The dataset is split into 6,210 training images and 2,079 test images, with helmet and no_helmet as the two labels.Images are labelled with MakeSense.ai, duplicates removed, and resized so at least one side is 640 pixels.Source: §3.1.1 · p. 5
- The evaluation framework uses accuracy, precision, recall, F1 score, and mAP derived from confusion-matrix counts.The F1 score is described as the harmonic mean of precision and recall.Source: §3 · p. 3, §3.1.1 · p. 5, §3.1.2 · p. 5
3.1.2. Training and validating
- F1 score reached 0.94 at a confidence threshold of 0.444, while precision reached 1.00 at 0.911.The thresholds were selected from the F1-confidence and precision-confidence curves, respectively.Source: Figure 7 · p. 6, Figure 8 · p. 6, §3.1.2 · pp. 6–7
- Decreasing training and validation losses and increasing precision, recall, and mAP indicate improving YOLOv8 performance over training.The reported loss trends cover box, classification, and distribution focal losses.Source: §3.1.2 · pp. 6–7, Figure 10 · p. 7
4. Experimental results and evaluation
The evaluation tested the system at active outdoor construction sites, across worker densities, and in a video-processing trial. Reported results cover detection accuracy, processing speed, and YOLOv8 training trends.
4.1. Experimental scenarios and results
- Tests spanned active outdoor civil, building, and electrical construction sites, with worker groups performing tasks such as lifting, assembling, and operating machinery.The scenarios varied the mix of helmeted and non-helmeted workers.Source: §4.1 · p. 7, Figure 11 · p. 7, Figure 12 · p. 8
4.2. Experimental evaluation across varying image densities
- Across 500 images, helmet-less detection reached 99.8% true positive rate below 10 people, 99.5% accuracy at 10–20, and 99% accuracy above 20 (maximum 30).Performance may degrade beyond 30 people, a density not tested; the passage also reports 12% edge-case errors.Source: Figure 11 · p. 7, §4.2 · p. 8, Figure 12 · p. 8
4.3. Real-time experimental results
- With four helmets per frame, total processing took around 15 ms, equivalent to 4 ms per helmet; the system demonstrated 20–60 FPS real-time capability.The passage notes that speed decreases when more helmets appear in a frame.Source: §4.3 · p. 9, Figure 15 · p. 9
4.4. Experimental evaluation
- Over 10 epochs, precision rose from 0.91 to 0.94, recall from 0.88 to 0.94, mAP50 from 0.94 to 0.97, and mAP50-95 from 0.62 to 0.68 as losses declined.The evaluation describes the small train–validation loss gap as evidence of minimal overfitting.Source: §4.3 · p. 9, §4.4 · p. 10
- YOLOv8 achieved mAP50 0.98, above YOLOv7 (0.95), EfficientDet-D2 (0.94), and DETR (0.93), while its mAP50-95 (0.68) trailed Faster R-CNN (0.70).Its inference speed was 50–60 FPS, compared with Faster R-CNN at 5–10 FPS and DETR at 10–15 FPS.Source: §4.4 · p. 10, Table 1 · p. 10