In order to solve the problem of low accuracy and poor adaptability of helmet detection, the author proposed an improved helmet detection method based on YOLOv3 network. Aiming at ensuring the accuracy of helmet detection and increasing the attention to the helmet in the picture, this method used attention mechanism to enhance the spatial information and semantic information extracted from the picture, and reduced the loss of image details. And then the author used deformable convolution to adapt to the variety of human posture, enhance the adaptability of the model to the target, and reduce a few of training samples. In the end, the author changes the size of output feature map and the shallow network features are integrated to improve the recognition rate of small targets such as human head. The self-made HELMET dataset is used to train and test the method, and the comparative experiments show that compared with other detection methods, this method can extract more target features, achieve higher mean average accuracy, and has better adaptability in practical application.