Edge End Safety Helmet Wearing Detection Based on Detection Head Optimization

Safety helmet wearing detection is an crucial means in safeguarding the lives of construction workers, but it often requires edge computing due to unstable networks and privacy concerns. In resource-constrained scenarios, the detection head of a model often accounts for an excessive amount of computation. Therefore, this study adopts RepConv, GhostConv, PConv, and GroupConv to reconstruct the shared parameters of the small model’s detection head and explores the performance of different models and the detection speed on Jetson Xavier NX. Among them, the detection head improved by RepConv increases the mean average precision (mAP) and F1 score of safety helmet wear detection by 0.6% and 0.72%, respectively, without increasing computation; the detection head improved by PConv increases the frames per second (FPS) of the model by 20.10% and reduces computation by 1/3. Overall, there is still great potential for improving the detection head.