【摘要】 自动识别车牌号码是智慧交通中的重要内容，针对现有车牌识别算法计算量大，不满足微型化、实时性等需求，提出一种基于边缘设备和改进YOLOv5算法的车牌号码识别方法。首先，构建车牌数据集；其次，通过改进YOLOv5网络模型架构，并引入注意力机制，提升对车牌号码的检测能力，并与未改进的YOLOv5算法作性能对比；最后，将Intel Movidius NCS2与树莓派硬件设备结合，进行实时推理。实验结果表明，改进的YOLOv5算法在边缘设备上的实时画面推理速度最快达到3.316 ms,YOLOv5算法推理速度为5.772 ms，改进的YOLOv5算法与原算法相比，其推理速度平均提升了13.41%。本文提出的方法能在边缘设备上提高车牌检测速度，并达到较高的准确率。更多还原

【Abstract】 Automatic recognition of license plate numbers is an important content in smart transportation. In view of the large amount of calculation of existing license plate recognition algorithms, which cannot meet the requirements of miniaturization and real-time performance, a license plate number recognition based on edge devices and improved YOLOv5 algorithm is proposed.First, build a license plate data set; secondly, by improving the YOLOv5 network model architecture and introducing an attention mechanism, the detection ability of license plate numbers is improved, and the performance is compared with the unimproved YOLOv5 algorithm; finally, Intel Movidius NCS2 and Raspberry Pi are compared. The hardware devices are combined for real-time reasoning. The experimental results show that the real-time picture inference speed of the improved YOLOv5 algorithm on edge devices is up to 3.316ms, and the inference speed of the YOLOv5 algorithm is 5.772 ms. Compared with the original algorithm, the improved YOLOv5 algorithm has an average increase of 13.41% in the inference speed. The method can improve the speed of license plate detection on edge devices and achieve high accuracy.更多还原