【作者】 张浩；

【导师】 王虎； 张瑞明；

【作者基本信息】 合肥工业大学 ， 机械工程（专业学位）， 2025， 硕士

【摘要】 近年来,我国道路交通网络快速发展,道路所涉及的范围越来越广,涵盖了城市的各个区域,交通事故发生的概率也越来越高,特别是在弯道路口处,据统计信息显示,此处发生重大交通事故的占比高达70%。本文针对弯道交通预警的需求,利用智能识别技术,在对弯道路况进行识别检测研究的基础上,基于嵌入式平台,开发了一种智能弯道交通预警系统。论文的主要工作如下:(1)根据弯道路况和实际的智能交通要求,基于智能识别技术和嵌入式平台,开发了智能识别的弯道预警系统,进行了总体方案的设计、硬件选型和系统的整体结构设计。(2)在对深度神经网络理论进行介绍的基础上,对YOLOv5的网络框架进行了详细论述,针对Yolov5算法的5.0与6.0版本进行特性参数对比,同时分析了处理器类型、浮点数精度、预训练模型对图片推理的影响,并对两种版本的图片推理性能进行对比分析。(3)在对Tensor RT加速原理进行简述的基础上,针对Yolov5算法的5.0与6.0版本使用Tensor RT进行加速同时分析处理器类型、浮点数精度、预训练模型对图片推理的影响,并与未使用Tensor RT加速前5.0与6.0版本对图片的推理进行对比分析。最终得出使用6.0版本在GPU处理器、FP＿16浮点数精度、n预训练模型并使用Tensor RT加速条件下推理性能最优。(4)设计了一种预警算法,详细论述了该预警算法的计算流程,对该算法通过实例分析,并将算法处理后的预警信息传输给LED屏。(5)分析5.0与6.0版本处理器类型、浮点数精度、预训练模型对视频推理的影响,选择出对视频推理性能最好的模型;搭建弯道预警系统的实验平台,并测试各项功能。实验结果表明,弯道预警系统使用GPU、FP＿16浮点数精度、n预训练模型并经过Tensor RT加速后的Yolov5-6.0版本识别算法下,结合预警算法,对路面的实时路况进行监控,能够判断弯道路况下人与车的运动状态,并根据人、车的运动状态进行实时的声光预警。更多还原

【Abstract】 In recent years,China’s road traffic has been continuously developing,covering increasingly broader areas across urban regions,while the probability of traffic accidents has also risen significantly,particularly at curved road sections.Statistical data indicates that such locations account for up to 70%of accidents.To address the demand for curved road traffic warnings,this paper leverages intelligent recognition technology to develop an intelligent curved road traffic warning system based on an embedded platform,building upon research on curved road condition detection.The main contributions of this thesis are as follows:(1)Based on curved road conditions and practical intelligent transportation requirements,an intelligent curved road warning system was developed using intelligent recognition technology and an embedded platform,including overall scheme design,hardware selection,and system architecture design.(2)After introducing deep neural network theory,the YOLOv5 network framework was discussed in detail.A comparative analysis of the characteristic parameters between YOLOv5 versions 5.0 and 6.0 was conducted,along with an examination of the impact of processor type,floating-point precision,and pre-trained models on image inference.The inference performance of the two versions was also compared.(3)Building upon Tensor RT acceleration principles,YOLOv5 versions 5.0 and6.0 were accelerated using Tensor RT.The influence of processor type,floating-point precision,and pre-trained models on image inference was analyzed,and a comparison was made with the inference performance before Tensor RT acceleration.The results demonstrated that version 6.0 achieved optimal inference performance under GPU processing,FP＿16 floating-point precision,pre-trained model"n,"and Tensor RT acceleration.(4)A warning algorithm was designed and analyzed through case studies.After algorithmic processing,the configured warning information was transmitted to an LED display.(5)The impact of processor type,floating-point precision,and pre-trained models on video inference was analyzed for versions 5.0 and 6.0,and the model with the best video inference performance was selected.An experimental platform for the curved road warning system was constructed,and its functionalities were tested.Experimental results show that under the YOLOv5-6.0 recognition algorithm—optimized with GPU processing,FP＿16 floating-point precision,pre-trained model"n,"and Tensor RT acceleration—combined with the warning algorithm,the curved road warning system can effectively monitor real-time road conditions,detect the movement states of pedestrians and vehicles on curved roads,and provide real-time audio-visual warnings based on their motion states.更多还原