Research on helmet detection model for electric vehicle riders based on improved YOLOv5

Existing helmet-mounted detection methods for cycling electric vehicles in network models have many parameter quantities and calculation quantities. In order to further improve the above parameters, the study puts forward a light test method for helmet on the basis of the improvement of YOLOv5 to form a lightweight and high-precision detection model. First of all, the DSConv module is instead of the Conv in the C3 module, which reduces parameters and calculated amount, thereby further improve the lightweight model. Secondly, adding the SE attention mechanism to filter unrelated information and solve information overload, so that allow the structure to focus more on vital information. Then replace the neck structure in YOLOv5 with the improved Effcient-RepGFPN structure to complete the feature fusion process. Meanwhile, this method uses EIoU loss to obtain more accurate box positioning and improve the precision of bounding box regression. After the network is improved, comparative experiments are carried out with various algorithms. Finally, the ablation experiment of the improved method is carried out on the constructed dataset to continue testing the validity about test model. Through the experiment to prove that this method makes mAP increased by 1.68%, the quantity of parameters is reduced to 68.77% of the unadorned algorithm. Meanwhile, reduce the amount of calculation to 52.53% and reduce the inference time.