【作者】 刘安；

【导师】 李盛； 马俊杰；

【作者基本信息】 武汉理工大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 随着城市交通的不断发展和车辆数量的急剧增加,高速公路管理面临日益严峻挑战。车型识别是提高高速公路管理效率和安全性的重要研究内容之一。传统的车型识别技术如视频图像、雷达、地磁感应器等大多容易受天气环境影响、存在监控盲区、灵敏度低等问题,尤其在恶劣天气下的高速公路中难以发挥作用。因此,亟需发展新的方法来实现全天候高速公路车型识别。光栅阵列具有抗干扰能力强、监测范围广、灵敏度高以及敷设简单等优势,为克服传统车型识别技术的种种限制提供了一种新的解决途径。本文基于光栅阵列传感技术提出了一种车辆检测与车型识别方法。通过光栅阵列传感光缆采集车辆通过时的高速公路振动信号,综合运用信号处理、模式识别、深度学习等方法对车辆振动信号进行检测并开展车型识别。为了满足高速公路系统轻量化部署、实时性监测需求,利用网络重构和通道剪枝方法对车型识别深度学习模型轻量化方法开展了研究。论文的主要工作如下:(1)高速公路光栅阵列车辆检测方法研究。首先阐述光栅阵列振动信号检测原理,并设计高速公路车辆监测传感网络布设方案。其次针对高速公路监测需求划分了四种车辆型号分类标准,设计了车辆振动信号采集实验,分析了四种典型类型车辆振动信号的时域频域特征,揭示了直接基于时频特性辨识车型的困难。最后针对传统双门限检测法在噪声环境中标定门限阈值困难的问题,提出结合FCM聚类算法和BIC判别准则的自适应双门限法检测并提取车辆振动信号。(2)基于光栅阵列传感网络的车型识别方法研究。首先针对车辆振动信号复杂,非线性模式,采用Efficient Net-b0网络结合双向长短时记忆网络组成复合网络模型对不同类型的车辆振动信号进行识别。其次利用光栅传感网络可以捕获车辆行驶时的空间分布和时序特征优势,提出基于连续多光栅传感数据特征融合与组合模型相结合的车型识别方法,最后通过实验对比了与其他传统识别方法的效果。(3)车型识别轻量化网络模型方法研究。针对复合网络识别车型时模型体积大、识别效率低的问题,首先采用Triplet轻量级注意力机制和Ghost轻量级卷积层进行网络模型轻量化重构,其次利用通道剪枝算法进一步压缩模型体积,提高识别效率及精度,并使模型具有高压缩比性能,最后开展对比消融实验,检验了提出的车型识别模型轻量化重构方法的性能优势。更多还原

【Abstract】 With the continuous development of urban transportation and the rapid increase in the number of vehicles,highway management is facing increasingly severe challenges.In order to enhance the efficiency and safety of highway management,vehicle type recognition technology has become a crucial research area.Traditional vehicle type recognition technologies such as video images,radar,and magnetic sensors are often susceptible to weather conditions,have monitoring blind spots,and exhibit low sensitivity.Adverse weather conditions,in particular,pose a high-risk period for highway traffic accidents.Therefore,there is an urgent need to develop new methods to achieve all-weather highway monitoring.The grating array possesses advantages such as strong anti-interference capability,wide monitoring range,high sensitivity,and simple deployment,providing a new solution approach to overcome various limitations of traditional vehicle recognition technologies.This thesis proposes a vehicle detection and classification method based on grating array sensing technology.By utilizing the grating array sensing optical fiber to collect highway vibration signals when vehicles pass through,a combination of signal processing,pattern recognition,deep learning,and other methods is employed to detect vehicle vibration signals and conduct vehicle classification.To meet the lightweight deployment and real-time monitoring requirements of highway systems,research is conducted on lightweight methods for vehicle classification deep learning models using network reconstruction and channel pruning techniques.The main contributions of the thesis are as follows:(1)Research and signal analysis of highway grating array vehicle detection method.Firstly,elucidate the principle of vibration signal detection of the grating array and design a scheme for deploying a high-speed highway vehicle monitoring sensor network.Secondly,addressing the difficulty of calibrating threshold values in noisy environments using traditional dual-threshold detection methods,an adaptive dual-threshold method combining FCM clustering algorithm and BIC discriminant criterion is proposed to detect and extract vehicle vibration signals.Finally,four types of vehicle classification standards are defined according to the monitoring requirements of highways,and experiments for collecting vehicle vibration signals are designed.The research focuses on the time-frequency domain characteristics of vibration signals of four types of vehicles and the challenges of vehicle classification.(2)Research on vehicle recognition method based on grating array sensing network.Firstly,considering the complexity and non-linear nature of vehicle vibration signals,an Efficient Net-b0 network combined with bidirectional long short-term memory(Bi LSTM)network is employed to form a composite network model for identifying different types of vehicle vibration signals.Secondly,taking advantage of the ability of the grating sensing network to capture spatial distribution and temporal characteristics of vehicle movement,a vehicle recognition method based on the fusion and combination of features extracted from continuous multi-grating sensing data is proposed.Finally,experimental comparisons are conducted to evaluate the effectiveness of the proposed method against other traditional recognition methods.(3)Research on lightweight network model methods for vehicle recognition.To address the issues of large model size and slow recognition speed in vehicle recognition composite network models,the Triplet lightweight attention mechanism and Ghost lightweight convolutional layers are employed for model lightweight reconstruction.Additionally,channel pruning algorithms are utilized to further compress the model size and improve recognition efficiency while maintaining high-precision recognition and high compression ratio performance of the vehicle recognition network.Comparative ablation experiments are conducted to examine the performance advantages of the vehicle recognition model after lightweight reconstruction.更多还原