In order to solve the problems of small target, intensive, poor accuracy, slow detection speed and difficult application in helmet wearing detection task, this paper proposed a EfficientNetV2-SSD algorithm based on SSD network. Aiming at the problem of multiple SSD network parameters, the improved lightweight network EfficientnetV2 is used to replace the feature extraction network in the SSD to reduce network parameters and improve detection speed. For small targets that are difficult to detect, top-down and bottom-up FPN pyramid structures are used to enrich the information of all prediction feature layers to the maximum extent and improve the detection accuracy of small targets. Aiming at the characteristics of the detected target such as helmet, the size and proportion of the prior frame are redesigned to improve the accuracy of small target detection, accelerate the convergence speed of the network and reduce the network volume. The experimental results show that EfficientNetV2-SSD improved the average accuracy of helmet wearing detection by 7.01% and the network volume was reduced by 75% compared with the SSD network, with better practicability.