Aiming at the problems that the traditional detection methods of motorcycle helmet detection have low accuracy, poor generalization ability and large number of target detection network parameters, which are difficult to run on embedded devices, an improved MNXt-ECA-D-YOLOv2 target detection algorithm model of YOLOv2 is proposed. First, MobileNeXt network is introduced to replace original YOLOv2 backbone network, and a densely connected network structure is introduced into the sandglass block of MobileNeXt. At the same time, the effective channel attention mechanism is introduced into the network. And, different activation functions are applied at different depth network layers. Finally, DropBlock module is added before the network output convolutional layer. K-means clustering algorithm is adopted to redesign the anchor box size of self-made dataset. The experimental results show that compared with the original YOLOv2 under the same experimental conditions, the proposed method improves the AP50 metric by 3.53% and the model size reduced by 77.44%, and the detection speed increased by nearly 4 times. Comparison experiments demonstrate that the improved YOLOv2 has a higher average accuracy rate, a smaller model, and faster inference speed in CPU. Therefore, the proposed improved YOLOv2 model is valuable in practical applications.