Streamlined Safety Helmet Detection: An Enhanced YOLOv5 Approach

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Oct 17 2023

In a paper published in the journal PLOS ONE, researchers tackled the challenge of designing a robust safety helmet detection method for engineering projects while considering practical concerns, such as cost control for hardware facilities. They introduced an optimization approach for the BottleneckCSP (Cross-Stage Partial) structure within the YOLOv5 (You Only Look Once version 5) backbone network. This optimization significantly reduced model complexity without altering network input and output sizes.

The paper proposed an upsampling feature enhancement module to enhance semantic information and mitigate information loss due to upsampling. Moreover, they introduced self-attention mechanisms, utilizing both a channel attention module and a location attention module, to adaptively fuse adjacent shallow feature maps and upsampled feature maps, thereby generating new feature maps with improved semantics and precise location information. Compared to existing approaches, the method achieved faster inference speeds, reaching an impressive 416 frames per second (FPS) under the same compute capability, and demonstrated superior performance with a mean Average Precision (mAP) of 94.2%.

Background

Construction sites pose various safety hazards, including the risk of accidents such as falling objects or heavy items injuring people. One standard safety measure is safety helmets, but many workers avoid wearing them due to discomfort, endangering their lives. Helmet-wearing supervision is a manual process, requiring significant human intervention, and is susceptible to errors. To address this, machine vision application for automated safety helmet detection is essential, reducing accidents and ensuring worker safety.

With advancements in artificial intelligence, machine vision, especially deep learning using Convolutional Neural Networks (CNNs), has become a viable approach for this purpose. Traditional methods relying on handcrafted features have limitations in natural construction environments due to lighting, angles, and obstructions. Recent CNN-based algorithms, such as YOLO and SSD, offer faster and more practical helmet detection solutions. Nonetheless, these algorithms often require substantial computational resources and highlight the importance of developing a lightweight helmet detection model suitable for construction sites with limited computing power.

Related Work

Researchers have categorized past works in safety helmet detection methods into two primary types: those based on traditional machine learning and those rooted in deep learning. Traditional machine learning methods, although have made progress, have certain limitations. For instance, they involve complex steps like feature extraction and classification. One method used K-nearest neighbor classifiers to identify whether motorcyclists were wearing helmets. An alternative method entailed combining color-based features with support vector machines to categorize helmet colors.

On the other hand, recent advances in deep learning have introduced more efficient methods like YOLOv5, widely used for helmet detection due to its balance between accuracy and speed. For instance, researchers have developed improved versions of YOLO to enhance the model's performance, introducing features such as adaptive anchor box adjustment, feature weighting, and attention mechanisms to achieve more robust and accurate helmet detection. These advancements represent a shift towards faster and more effective helmet detection methods.

Proposed Method

YOLOv5 is a one-stage object detection method that simplifies the detection process by treating it as a regression problem. Unlike two-stage methods, YOLOv5 directly regresses the position and category of objects in the entire image without first extracting candidate boxes. This approach speeds up inference and simplifies the model. Mosaic data augmentation is employed during YOLOv5 training to improve the model's ability to detect small targets by randomly selecting and splicing four images after scaling and color perturbation. The network utilizes the Focus and Cross Stage Progression Network (CSPnet) structures to enhance feature extraction and inference speed.

Moreover, YOLOv5 incorporates a Generalized Intersection over Union Loss   (GIOU_Loss) to address issues related to non-overlapping bounding boxes, resulting in improved detection accuracy and speed. In pursuit of a balance between speed and accuracy, the study optimized the YOLOv5 Backbone module, introduced an upsampling module to enhance features, and incorporated attention mechanisms to eliminate redundant information. The network architecture of this optimized SHDet method based on YOLOv5 showcases the Lightweight Backbone module, an Upsample Module for feature enhancement, a Channel Attention mechanism to remove redundancy, and a Positional Attention mechanism to eliminate spatial redundancy.

The Lightweight Backbone module focuses on reducing computational complexity by modifying the BottleneckCSP structure. Additionally, an Upsampling module combines sub-pixel convolution and layer-by-layer parallel dilated convolution to enhance feature maps' resolution and context information, increasing sensitivity to small objects. The Channel Attention mechanism minimizes redundant information introduced during feature fusion, while the Positional Attention module highlights the importance of object spatial regions using a self-attention mechanism. These modifications aim to strike a balance between detection speed and accuracy in safety helmet detection.

Results and Conclusion

The study conducted experiments to verify the effectiveness of the proposed lightweight helmet detection network. They used their self-curated dataset and the publicly accessible Safety Helmet Wearing Dataset (SHWD) for evaluation. The results showed that the proposed SHDet method outperformed other approaches in terms of both performance and speed on their dataset while maintaining its superiority on the SHWD dataset compared to recent safety helmet detection methods. This evidence underlines the feasibility and practicality of the SHDet method.

In summary, this paper simplified the YOLOv5 backbone network to enable the proposed network to run on mobile devices efficiently. The integrated upsampling, channel attention, and positional attention modules significantly improved the safety helmet detection performance. While the SHDet method demonstrated top-tier performance and speed compared to recent algorithms, there is still room for further optimization. Future work will focus on incorporating a self-attention mechanism from Transformer into the YOLOv5 framework to enhance the model's global context feature extraction capability.