【作者】 王伟；

【导师】 蔡晓军；

【作者基本信息】 山东大学 ， 计算机技术（专业学位）， 2023， 硕士

【摘要】 随着社会的发展,道路问题凸显的越来越严重,各国学者开始研究和尝试智能交通系统作为改善现有道路状况、提高交通运行效率的重要方向。而要实现智能交通系统,关键在于能够实时动态地掌握道路交通信息。车速作为重要的交通参数,决定着交通事故发生的严重性,更重要的是关系到行车安全。因此通过对车辆速度实施有效的监控,可以有效减少交通事故和降低道路拥堵。因此,实时监控车辆行驶速度是智能交通系统研究的基础和先决条件之一。对此,本文对基于FPGA的车速检测系统展开研究和优化设计。首先设计一款可编程SOC超高速摄像机对高速运行车辆拍摄,并实现对视频的预处理,获得连续不丢帧的图像信息,高速相机能够将高速运动物体完整获取并且可以进行慢放,对于高速运动物体的运动轨迹预测具有极为重要的意义。将预处理的图像进行目标检测训练和目标检测算法改进,主要改进包括几个方面:首先YOLOv5的Backbone网络用MobileV3模块进行替换,达到提升检测推理速度的目的,提高了目标追踪的精准度。其次,将两个损失函数进行线性结合,达到减弱在一些状态下GIoU退化为IoU的程度和加速收敛的目的。最后,在车牌识别阶段,先通过CRNN算法进行无分割车牌识别,再利用双向LSTM网络来实现车牌的识别。由实验结果可以看出,降低了模型权重,提升了目标追踪的精准度。通过比较普通视频中的车速和经过高速相机处理过的视频车速,经过预处理的视频车速精度明显提升。最后为了在智能交通中实现低功耗以及更好地部署在移动端,将算法移植到FPGA上进行加速,在Xinlinx的ZYNQMPSoC系列开发板ZCU104上进行FPGA加速。后期选用合适的嵌入式开发板和改进高速相机输出接口,实现相机和嵌入式设备直连,进一步缩小系统体积和提高便携性。更多还原

【Abstract】 With the development of society,road problems are becoming more and more serious,and scholars from various countries have begun to study and try intelligent transportation systems as an important direction to improve existing road conditions and improve traffic operation efficiency.The key to realizing an intelligent transportation system is to be able to grasp road traffic information dynamically in real time.As an important traffic parameter,vehicle speed determines traffic flow information and congestion,and more importantly,it is related to driving safety.Therefore,by monitoring the speed of vehicles in real time,it can not only improve people’s travel efficiency and quality of life,but also effectively prevent traffic accidents and road congestion caused by speeding.Therefore,real-time monitoring of vehicle driving speed is one of the foundations and prerequisites for intelligent transportation system research.In this paper,this paper studies and optimizes the design of FPGA-based vehicle speed detection system.Firstly,a programmable SOC ultra-high-speed camera is designed to shoot high-speed vehicles,and realize the pre-processing of the video,obtain continuous image information without losing frames,and the high-speed camera can completely capture the motion scene through a very high frame rate and play it back at a very slow speed,which is of great significance for the dynamic analysis and prediction of moving objects.The preprocessed images are trained for object detection and improved object detection algorithms,and the main improvements include several aspects:First,the main module of MobileV3 network is used to replace the DeepSORT re-recognition network and YOLOv5’s Backnone network,which improves the speed of detection inference and improves the accuracy of target tracking.Secondly,by combining the CIoU and GIoU loss functions,the degree of GIoU degradation to IoU and accelerated convergence in some states are reduced.Finally,in the license plate recognition stage,the CRNN algorithm is used to perform undivided license plate recognition,and then the two-way LSTM network is used to realize the license plate recognition.It can be seen from the experimental results that the weight of the model is reduced and the accuracy of target tracking is improved.By comparing the speed of the vehicle in the ordinary video with the speed of the video processed by the high-speed camera,the speed accuracy of the preprocessed video is significantly improved.Finally,in order to facilitate the practical application of low power consumption,flexibility and convenience in intelligent transportation,the algorithm is embedded,and the hardware acceleration project is realized on the ZCU104 ZYNQ MPSoC series development board of Xinlinx.In the later stage,a suitable embedded development board and an improve high-speed camera output interface are selected to realize direct connection between the camera and embedded devices,further reducing the system size and improving portability.更多还原