Deep Learning for Real-time Safety Helmet Detection

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Apr 1 2024

In a paper published in the journal Scientific Reports, researchers proposed employing deep learning techniques for real-time detection of safety helmet usage in the construction industry.

They enhanced the you only look once version 5 small (YOLOv5s) network by introducing a bidirectional feature pyramid network (BiFPN) bidirectional feature pyramid network, refined the post-processing method to soft non-maximum suppression (Soft-NMS), and improved the loss function for faster convergence. Their novel model, BiFEL-YOLOv5s, exhibited an increase in average precision and a boost in recall rate while maintaining adequate detection speed, making it suitable for real-time helmet detection in various construction scenarios.

Related Work

Past work in construction safety has progressed from traditional methods to advanced deep-learning techniques for safety helmet detection. Traditional approaches relied on color and shape features, utilizing methods like support vector machines (SVM) and histogram of orientation gradient (HOG). However, these methods struggled with complex environments and needed more generalization.

Recent advancements have shifted towards deep learning, with one-stage algorithms such as YOLOv5 and improvements like YOLOv4-based models showing promise for real-time detection. Furthermore, researchers have proposed enhancements such as multi-scale feature extraction, channel attention mechanisms, and new loss functions to improve accuracy and speed. These developments aim to address the challenges posed by complex construction environments and enhance the efficiency of safety helmet monitoring.

Enhancements for Safety Helmet Detection

The proposed methods encompass several critical enhancements to improve safety helmet detection in construction environments. Initially, the selection of the YOLOv5 model is motivated by its balance between accuracy, speed, and resource efficiency, making it suitable for real-world application.

In particular, YOLOv5s, the minor variant within the YOLOv5 family, is chosen due to its minimal resource consumption and commendable detection accuracy compared to other versions. This decision is crucial given the constraints of equipment performance and budget considerations typical in construction settings.

Concerning the loss function utilized in the YOLOv5 model, traditional IoU loss is employed, but its limitations, especially regarding aspect ratio consideration, prompt the introduction of the focal-enhanced intersection over union (EIoU) loss. This novel loss function addresses the complete IoU (CIoU) loss deficiency by separately considering width and height during regression, thus enhancing the model's performance. Additionally, incorporating focal loss mitigates issues associated with low-quality samples, ensuring better model optimization and performance.

Furthermore, attention mechanisms play a pivotal role in improving detection performance, particularly for small objects like safety helmets. Three attention mechanisms, namely squeeze-and-excitation networks (SeNet), convolutional block attention module (CBAM), and coordinate attention (CA), are compared for their effectiveness in safety helmet detection. SeNet emphasizes global channel relationships, while CBAM integrates channel and spatial attention.

Conversely, CA attention incorporates location information into channel attention, offering a comprehensive approach. Researchers assessed each mechanism for their computational complexity and ability to capture spatial correlations, with CBAM standing out due to its sequential attention structure and comprehensive feature extraction capabilities.

Moreover, NMS is a crucial post-processing step in object detection that eliminates redundant bounding boxes. However, the traditional NMS method has limitations in determining IoU thresholds and may lead to false suppression of overlapping objects. To address this, researchers propose Soft-NMS as an improvement, wherein instead of directly deleting overlapping boxes, it reduces their confidence scores, ensuring better retention of overlapping objects.

Finally, researchers enhance the feature pyramid network in YOLOv5 by introducing BiFPN, aiming to improve multi-scale fusion efficiency. The model learns the importance of different input features by simplifying the feature fusion network and adjusting weight values for feature fusion, enhancing overall performance. These proposed methods collectively contribute to advancing safety helmet detection in construction environments, addressing the challenges posed by complex conditions, and improving the efficiency of safety monitoring systems.

Helmet Detection Algorithm Evaluation

The study leverages the safe helmet-wearing dataset (SHWD) dataset, comprising 3241 images from an open repository, to evaluate safety helmet detection algorithms. With standardized annotations indicating helmet presence and individual labels, the dataset provides a diverse benchmark for algorithm assessment across varied construction scenarios.

Experimentation employs an 11th Gen Intel(R) Core (TM) i7-11800H CPU and an NVIDIA GeForce RTX 3050 Ti Laptop GPU, which stands for NVIDIA's Graphics Processing Unit (GPU), utilizing PyTorch for model training. Evaluation metrics encompass precision, recall, average precision (AP), mean average precision (mAP), and frames per second (FPS), ensuring a comprehensive assessment of detection accuracy and speed.

Experimental findings underscore the efficacy of the proposed BiFEL-YOLOv5s model, which integrates the SE attention mechanism, BiFPN network, Focal-EIoU Loss, and Soft-NMS. This novel approach significantly enhances detection accuracy and recall rate while effectively addressing occlusion challenges while maintaining detection speed. Comparative analysis with existing algorithms demonstrates superior performance in precision, speed, and mitigation of misdetection and leakage issues, positioning BiFEL-YOLOv5s as a robust solution for safety helmet detection in construction environments.

Conclusion

In conclusion, this paper explored safety helmet object detection in construction environments using deep learning techniques, mainly focusing on the YOLO algorithm. Through experimentation and analysis, YOLOv5s emerged as the optimal base network, enhanced further with SE attention mechanism, Focal-EIoU loss, and Soft-NMS. The proposed BiFEL-YOLOv5s model demonstrated significant improvements in accuracy and real-time processing, making it a viable solution for safety helmet detection in complex work scenarios.