With the growth of non-motorized vehicles, low-carbon and environmentally friendly travel has gradually become the mainstream of people all over the world. The wide distribution and small proportion of non-motorized vehicles in the image make it difficult for traffic regulators to detection them, which increases the work of urban managers. Focusing on the task of non-motorized vehicle detection, we improve the YOLOv5s. The symmetric revolution module is used to construct the spatial attention mechanism, so as to improve the robustness of the network model. We design a spatial attention module to reduce the parameters. The proposed dimension decoupling module decouples the dimensions of the width and height of the feature map respectively, so that the network can adapt to the width and height of non-motorized vehicle targets. Through multi-scale receptive field fusion module, residual connection is used to improve the detection ability of the network to small targets. Finally, above modules are integrated into the YOLOv5s. Experiments on real road traffic images show that the proposed model has achieved good results in non-motorized vehicle detection.