Abstract: The existing vehicle recognition system has a complex network structure and a large model memory footprint, making it difficult to deploy to edge devices with limited resources. To this end, a lightweight vehicle recognition algorithm called YOLOv4-LVR based on improved YOLOv4 is proposed. We replaced the backbone network and used depthwise separable convolution to make lightweight improvements to the network, and then improved network performance by adding attention mechanism, updating a priori box, and introducing Focal Loss. Experiments on the KITTI dataset show that proposed algorithm greatly reduces the computational and storage overhead of the network while maintaining high accuracy, and can be deployed on edge devices at a low cost to effectively identify target vehicles.