Research on YOLOv5n-based lightweight hardhatwearing detection algorithm
Wanbo Luo, Ahmad Ihsan Mohd Yassin, Khairul Khaizi Mohd Shariff, Rajeswari Raju · 2024
Hardhats are essential protective gear for construction workers to ensure their safety. Real-time detection of hardhat-wearing is crucial due to the high frequency of casualties caused by weak safety awareness. Traditionally, data is obtained through manual inspection of construction sites or video footage. However, this method is labor-intensive and costly. As researchers explore neural networks, hardhat-wearing detection methods based on deep learning can ensure hardhat-wearing compliance. However, deep learning-based detection models typically require numerous parameters and computations, making them unsuitable for running on embedded devices with limited resources. This paper replaced the You Only Look Once v5 (YOLOv5n) backbone with a further lightweight backbone to reduce parameters and floating-point operations (FLOPs) and improve detection speed. Experiments were conducted to determine the most suitable lightweight backbone to replace the YOLOv5n backbone. The options considered were the backbone of GhostNet_050, LCNet_050, ShuffleNetv2_050, or MobileNetv3_small_050. The experimental results indicated that the YOLOv5n-LC-050 model reduced parameters and FLOPs by 32.66% and 43.9%, respectively, compared to the YOLOv5n model. However, the model's Mean Average Precision at 50% Intersection over Union (mAP50) only decreased by two percentage points and reached 86.8%, which achieved a good balance between accuracy and speed.