Helmet Detection Algorithm Based on Improved YOLOv7

Abstract

To address the issues of small target leakage, misdetection, and poor detection effect in helmet-wearing target detection, a small target helmet-wearing detection based on YOLOv7 is designed. Firstly, the space-to-depth (SPD) block is introduced to replace the convolution of the first down-sampling, and all subsequent Max Pooling (MP) structures are replaced with SPD structures, which can better capture the information of different depths in the graphs, and improve the performance of detecting small target helmets. Secondly, the use of ContentAware ReAssembly of Features (CARAFE) for the up-sampling section increases the receptive field, adapts to different types of feature maps, and improves the feature fusion capability; Then, SPPFCSPC is obtained by changing the connection method of SPPCSPC, and the spatial pyramid pooling module GhostSPPFCSPC is built by fusing a Ghost module and an SPPFCSPC module, in which the SPPCSPC module consists of two submodules, the spatial pyramid pooling (SPP) module and the connected spatial pyramid convolution (CSPC) module, which decreases the number of parameters and boosts the network performance in the meantime; Thirdly, the design uses SIoU to optimize the bounding box return loss function, which achieves faster convergence in the training phase, better performance in inference, and improved model accuracy and robustness. To demonstrate that the improved algorithm has better results than the original model with mAP boosted by 5.8%, accuracy boosted by 0.8%, parameter count decreased by 7.08%, and FPS up to 65 f/s.