WO2020114120A1 - 一种车辆信息识别方法、系统、存储器及处理器 - Google Patents

一种车辆信息识别方法、系统、存储器及处理器 Download PDF

Info

Publication number
WO2020114120A1
WO2020114120A1 PCT/CN2019/112504 CN2019112504W WO2020114120A1 WO 2020114120 A1 WO2020114120 A1 WO 2020114120A1 CN 2019112504 W CN2019112504 W CN 2019112504W WO 2020114120 A1 WO2020114120 A1 WO 2020114120A1
Authority
WO
WIPO (PCT)
Prior art keywords
vehicle information
data set
vehicle
image
information data
Prior art date
Application number
PCT/CN2019/112504
Other languages
English (en)
French (fr)
Inventor
刘若鹏
栾琳
曾梦萍
Original Assignee
深圳光启空间技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳光启空间技术有限公司 filed Critical 深圳光启空间技术有限公司
Publication of WO2020114120A1 publication Critical patent/WO2020114120A1/zh

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements

Definitions

  • the present invention relates to the field of information recognition technology, and in particular, to a vehicle information recognition method, system, memory, and processor.
  • the video surveillance vehicle information system as an important part of smart security and smart transportation in the Internet of Things application for urban public safety integrated management, is facing great challenges of in-depth application.
  • the main application bottleneck is how to quickly and efficiently present structured vehicle information in the video.
  • the video surveillance vehicle information system implemented by the traditional method uses the method of convolutional neural network, although it can extract different input features, including shallow features, deep features, local features, etc., and classify them according to the feature information. At present, it has achieved good results in image classification, but there is room for improvement in the real-time performance and accuracy of fine-grained classification. It is impossible to extract vehicle information in a unified manner, and it is necessary to analyze the fusion information multiple times.
  • the technical problem to be solved by the present invention is to provide a vehicle information recognition method, system, memory and processor, which can directly parse the information of the vehicle itself into each independent class on the basis of yolov3 to realize end-to-end input and output , So that the network output is the structured information of the vehicle, so that the vehicle information recognition detection process takes less time and the recognition efficiency is high.
  • an embodiment of the present invention provides a vehicle information recognition method, including: establishing a vehicle information data set; setting parameters based on the yolov3 network; training the vehicle information data set based on the yolov3 network, A multi-attribute network model that can identify the vehicle information is obtained; the obtained multi-attribute network model is used to perform multi-attribute prediction on the vehicle information of the image to be identified to identify the vehicle structured information in the image to be identified.
  • the establishment of the vehicle information data set includes: preprocessing the vehicle information data set; and performing data enhancement processing on the vehicle information data set.
  • the setting based on the yolov3 network parameters includes: setting the class parameters based on the yolov3 network to the number of categories defined by the vehicle information data set; and setting the filter parameters based on the yolov3 network to 3 ⁇ (class parameters +5).
  • the training of the vehicle information data set based on the yolov3 network to obtain a multi-attribute network model that can identify the vehicle information includes: classifying the vehicle information data set with a marking tool to generate .xml Format calibration file; convert the .xml format calibration file to a .txt file and generate a .txt file that stores the path of each image in the vehicle information data set; change according to the category setting in the vehicle information data set Network hyperparameters.
  • the use of the obtained multi-attribute network model to perform multi-attribute prediction of the vehicle information of the image to be recognized includes: inputting video resources or static images containing vehicle information into the trained network model, and outputting the Structured information of vehicle information.
  • the preprocessing of the vehicle information data set includes: counting the vehicle information data set and storing them in different folders according to different categories.
  • the data enhancement processing on the vehicle information data set includes: performing flip processing on the image of the vehicle information data set, renaming the corresponding flipped image and storing the same image in the vehicle information data set Under the folder.
  • the data enhancement processing of the vehicle information data set includes: performing mirror image processing on the image of the vehicle information data set, and renaming the corresponding mirror image image in the vehicle information data set as the same image Under the folder.
  • the angle of the turning process is -15°-15°.
  • the converting the .xml format calibration file into a corresponding .txt format file and generating a .txt file storing the path where each image in the vehicle information data set is located includes: reading the .xml format Calibration file to obtain the width and height of the image containing the vehicle dataset, the width and height of the vehicle included in the image of the vehicle dataset, and the width and height of the vehicle included in the image of the vehicle dataset One-step processing; calculate the coordinates of the vehicle center point contained in the image of the vehicle data set; save as a .txt text file according to the category serial number, center point coordinates, and width and height information of the vehicles contained in the image of the vehicle data set; Generate a .txt text file that stores the path of the sample vehicle information data set.
  • the .txt text file is saved in the label folder.
  • the normalizing the width and height of the vehicle included in the image of the vehicle data set refers to: converting the aspect ratio of the vehicle included in the image of the vehicle data set to 0 to 1 The value between.
  • an embodiment of the present invention provides a storage medium, the storage medium including a stored program, wherein the above-mentioned vehicle information recognition method is executed when the program is run.
  • an embodiment of the present invention provides a processor for running a program, wherein the above-mentioned vehicle information recognition method is executed when the program is running.
  • an embodiment of the present invention provides a vehicle information recognition system, including: a vehicle information data set setting module electrically connected, a yolov3 based network parameter setting module, a yolov3 based network training module, a yolov3 based network prediction module
  • the vehicle information data set setting module is used to establish a vehicle information data set
  • the yolov3 based network parameter setting module is used to set a yolov3 based network parameter
  • the yolov3 based network training module is used to perform the vehicle information data set Training to obtain a multi-attribute network model that can identify the vehicle information
  • the yolov3 based network prediction module is used to use the obtained multi-attribute network model to perform multi-attribute prediction on the vehicle information of the image to be identified to identify the pending Identify the vehicle structured information in the image.
  • the vehicle information data set setting module further includes a data preprocessing module and a data enhancement processing module, the data preprocessing module is used to preprocess the vehicle information data set; the data enhancement processing module is used to Data enhancement processing is performed on the vehicle information data set.
  • the yolov3 based network parameter setting module sets the class parameter based on the yolov3 network as the number of categories defined by the vehicle information data set; sets the filter parameter based on the yolov3 network parameter to 3 ⁇ (class parameter +5) .
  • the data preprocessing module counts the vehicle information data set and stores them in different folders according to different categories.
  • the data enhancement processing module performs flip processing on the image of the vehicle information data set, and renames the corresponding flipped image to be stored in the same folder of the image of the vehicle information data set.
  • the data enhancement processing module performs mirror image processing on the image of the vehicle information data set, and renames and saves the corresponding mirror image image in the same folder of the image of the vehicle information data set.
  • the above technical solution has the following advantages: on the basis of yolov3, setting network parameters that affect the performance of the network model makes the convergence speed faster and the network model more effective, and directly analyzes the information of the vehicle itself as Each independent class realizes end-to-end input and output, making the network output as the structured information of the vehicle, so that the vehicle information recognition and detection process takes less time and has higher recognition efficiency, and is generally applicable to the field of vehicle detection.
  • FIG. 1 is a flowchart of the vehicle information recognition method of the present invention.
  • FIG. 2 is a flowchart of a preferred embodiment of the vehicle information recognition method of the present invention.
  • FIG. 3 is a structural diagram of the vehicle information recognition system of the present invention.
  • Yolov3 mainly explains from three aspects, the input, structure and output of the network.
  • ResNet-53 adopts ResNet, which is a layer-hopping connection, and its performance is completely better than that of ResNet-152 and ResNet-101.
  • the reasons are: the difference of the basic unit of the network; the fewer the number of network layers, the fewer the parameters and the need for Less calculation.
  • the yolov3 network uses the front 52 layer of darknet-53 (there is no fully connected layer).
  • the yolov3 network is a fully convolutional network, which makes extensive use of residual jumper connections.
  • sampling was generally performed using max-pooling or average-pooling with a size of 2*2 and a stride of 2.
  • a convolution with a step size of 2 is used for downsampling.
  • upsampling and route operations are used in the network, and 3 tests are performed in a network structure.
  • a key point of the deep model is whether it can converge normally.
  • This residual structure can ensure that the network structure can still converge under deep conditions, and the model can be trained; the deeper the network, the expression The better the features, the better the classification + detection effect; 1*1 convolution in the residual, use network in The idea of network greatly reduces the number of channels for each convolution, on the one hand, it reduces the amount of parameters (the larger the number of parameters, the larger the saved model), on the other hand, it reduces the amount of calculation to a certain extent.
  • Yolov3 uses the sigmoid function for center coordinate prediction. This makes the output value between 0 and 1. Under normal circumstances, YOLO does not predict the exact coordinates of the center of the bounding box. It predicts: the offset associated with the upper left corner of the grid cell of the prediction target; and uses the cell size in the feature map to normalize.
  • the predictions bw and bh from yolov3 are normalized using the height and width of the image.
  • the predictions bx and by of the box are (0.3, 0.8), then 13 x 13
  • the actual width and height of the feature map are (13 x 0.3, 13 x 0.8).
  • FIG. 1 is a flowchart of the vehicle information recognition method of the present invention. As shown in FIG. 1, a vehicle information recognition method includes steps:
  • Collect a certain number of vehicle information samples establish a vehicle information data set, set network parameters based on yolov3, and train the collected vehicle information samples.
  • the training results lead to a multi-attribute network model that can identify vehicle attributes.
  • the structured information of the vehicle information of the image to be recognized such as car model, year, body color and other attributes, are identified at a time.
  • step S11 is a flowchart of a preferred embodiment of the vehicle information recognition method of the present invention.
  • establishing the vehicle information data set in step S11 includes: preprocessing the vehicle information data set; and performing data enhancement processing on the vehicle information data set.
  • the vehicle information data set can be a video from different vehicles or a collection of static images containing different vehicles.
  • the number of vehicle samples used for training is small, and the distribution is uneven. Therefore, the vehicle information data set needs to be subjected to data enhancement processing, which can increase the amount of vehicle sample data used for training, and the training results are more reliable.
  • Pre-process the vehicle information data set such as statistical vehicle information data set
  • Carry out data enhancement processing on the vehicle information data set for example, sequentially flip some images of the vehicle information data set, and rename the corresponding flipped images in the same folder of some images of the vehicle information data set.
  • the angle of the flip processing can be set to -15° ⁇ 15°. It is also possible to perform mirror image processing on some images of the vehicle information data set, and to rename and save the corresponding mirror image images in the same folder of some images of the vehicle information data set. For samples with a small number of images, the angle transformation is relatively more diverse, so as to achieve the effect of sample equalization and sample diversification.
  • the flip or mirror image processing is generally performed according to the number of test samples and the specific size of the image of the vehicle information data set.
  • the first step use the calibration tool labelImg to calibrate all samples in the folder vehicle information dataset image. Among them, all categories of vehicle structured information such as model, year, and body color are independently classified into one category, and the categories are independent of each other. After the sample is calibrated, a corresponding calibration file in .xml format is generated.
  • Step 2 Convert the generated .xml format calibration files into .txt files one by one.
  • the vehicle Normalize the width and height of the vehicle included in the image of the data set, such as setting the ratio of the width and height of the vehicle included in the image of the vehicle data set to between 0 and 1, and calculate the coordinates of the center point according to the vehicle data
  • the category number, center point coordinates, and width and height of the vehicles included in the set image are saved as .txt files. Converting .xml format file format to .txt format file, through software code.
  • Step 3 Generate a .txt text file with the path of all samples as input for network training.
  • the parameters of the detection layer based on the yolov3 network model need to be defined according to the number of categories input by the network. If the number of categories changes, the number of filters in the corresponding network detection layer also changes accordingly. In addition, the learning rate has a profound effect on the performance of the network model.
  • Step 1 The category parameter in the network is equal to the number of categories defined by all samples.
  • Step 3 According to the experimental summary, the learning rate directly affects the performance of the network model. After many experiments to adjust the super-parameter learning rate, for example, you can continue to try to learn the value of the value: 0.001, 0.0001, 0.00001, etc., and finally need to choose the appropriate value according to the training needs, so that the network training can quickly converge, and the network model performance is better .
  • the vehicle information recognition method of the present invention By adopting the vehicle information recognition method of the present invention, the information of the vehicle itself is directly parsed into each independent class on the basis of yolov3, and end-to-end input and output are realized, so that the network output is the structured information of the vehicle, so that the vehicle information is recognized
  • the detection process takes less time and the recognition efficiency is high.
  • An embodiment of the present invention further provides a storage medium, the storage medium includes a stored program, wherein the above-mentioned vehicle information identification method flow is executed when the above program runs.
  • the above storage medium may be set to store program code for performing the following process of the vehicle information recognition method:
  • the foregoing storage medium may include, but is not limited to: a U disk, a read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory (referred to as RAM), mobile hard disk, magnetic disk or optical disk and other media that can store program code.
  • ROM read-only memory
  • RAM random access memory
  • mobile hard disk magnetic disk or optical disk and other media that can store program code.
  • the program of the built-in vehicle information recognition method flow runs faster, thereby quickly and efficiently completing the recognition of all vehicle information.
  • An embodiment of the present invention further provides a processor for running a program, wherein, when the program is executed, the steps in the vehicle information recognition method described above are executed.
  • the above program is used to perform the following steps:
  • the program of the built-in vehicle information recognition method flow runs faster, thereby quickly and efficiently completing the recognition of all vehicle information.
  • a vehicle information recognition system includes: a vehicle information data set setting module that is electrically connected, a yolov3 network parameter setting module, a yolov3 network training module, and a yolov3 network prediction module;
  • the vehicle information data set setting module is used to establish the vehicle information data set
  • the yolov3 based network parameter setting module is used to set yolov3 based network parameters
  • the yolov3 network training module is used to train the vehicle information data set to obtain a multi-attribute network model that can identify vehicle information;
  • the yolov3 network prediction module is used to use the obtained multi-attribute network model to perform multi-attribute prediction on the vehicle information of the image to be recognized, so as to identify the vehicle structured information in the image to be recognized.
  • the vehicle information data set setting module also includes a data preprocessing module and a data enhancement processing module.
  • the data preprocessing module is used to preprocess the vehicle information data set, for example, to store them in different folders according to different categories.
  • the data enhancement processing module is used for performing data enhancement processing on the vehicle information data set, for example, performing an inversion process on the image of the vehicle information data set, and renaming the corresponding inverted image in the same folder as the picture of the vehicle information data set.
  • the angle of the flip process can be set to -15° to 15° as needed.
  • the data enhancement processing module can also perform mirror image processing on the image of the vehicle information data set, and rename and save the corresponding mirror image image in the same folder of the picture of the vehicle information data set. If the number of vehicle information data sets is n (n>1), half of the data needs to be flipped and the other half of the data needs to be mirrored, then the final number of vehicle data sets is 2n, which can be increased The number of training samples makes the subsequent network training more accurate.
  • yolov3 network parameter setting module Based on the yolov3 network parameter setting module, set the number of categories defined for all samples based on the class parameters in the yolov3 network; set the filter parameter based on the yolov3 network parameters to 3 ⁇ (class parameter +5); based on the results of multiple experiments, set the The learning rate parameter in the yolov3 network parameter.
  • a marking tool is used to classify the vehicle information data set to generate a .xml format calibration file; the .xml format calibration file is converted to the corresponding txt file, and generate a .txt text file that stores each image path in the vehicle information data set; change the network hyperparameter according to the category setting in the vehicle information data set.
  • the converting the .xml format calibration file into a corresponding txt file and generating a .txt text file storing each image path in the vehicle information data set includes: reading the .xml format calibration file to obtain the vehicle containing the vehicle The width and height of the image of the data set, the width and height of the vehicle included in the image of the vehicle data set; normalize the width and height of the vehicle included in the image of the vehicle data set; calculate the vehicle Vehicle center point coordinates included in the image of the data set; by category number, vehicle center point coordinates and the width and height information of the vehicle included in the image of the vehicle data set are saved as a .txt text file; a path with all samples is generated .Txt text file.
  • the .txt text file is saved in the label folder. Normalizing the width and height of the vehicle included in the image of the vehicle data set means that the aspect ratio of the vehicle included in the image of the vehicle data set is converted into a value between 0 and 1.
  • a large number of video resources or static images containing vehicle information are input into the trained network model, and the structured information of the vehicle information is output at once, such as the model, year, and body color.
  • Models like domestic ones: BAIC, BYD, Changhe, Great Wall, Dongfeng, Southeast, UFO, Futian, Fudi, Hafei, Aerospace, Hongqi, Huapu, Huaxiang, Ji'ao, Geely, Refine, Jiangling, Jinbei, Cheetah, South Automobile, Chery, Huizhong, Shenlong, Wanfeng, SAIC, Dawning, Shuanghuan, Wuling, Xiali, Jinlong, FAW, Iveco, Zhonghua, Zhongshun, ZTE, Li , Guan Zhi, Legend, Cheetah, Lu Feng, etc.; Germany: BMW, Audi, Porsche, Mercedes-Benz, Volkswagen; UK: Rolls-Royce, Bentley, Jaguar
  • Years such as: A-2010; B-2011; C-2012; D-2013; E-2014; F-2015; G-2016; H-2017; J-2018; K-2019; L-2020; M-2021 ; N-2022; P-2023; R-2024; S-2025; T-2026; V-2027; W-2028; X-2029; Y-2030; 1-2031; 2-2032; 3-2033; 4 -2034; 5-2035; 6-2036; 7-2037; 8-2038; 9-2039; A-2040, etc.
  • Body colors such as: A: White B: Gray C: Yellow D: Pink E: Red F: Purple G: Green H: Blue I: Brown J: Black Z: Others.
  • the structured information is: Chery, B, A, then the vehicle represented by the image is structured
  • the information is: The model is a domestically produced Chery, the year B refers to the vehicle was produced in 2011, and the body color A refers to the vehicle body color is white.
  • vehicle information such as model, year, body color, etc. are useful for deep learning-based methods, but they are basically processed separately, which is time-consuming and requires a larger amount of data for network training.
  • vehicle information recognition method, system, memory and processor of the present invention the vehicle video information is input through one-time processing without any fusion process, and then a variety of different attribute value sets of vehicle information are output at one time, such as vehicle model identification , Year identification, body color identification, etc.
  • vehicle information such as model, year, body color, etc.
  • one-time prediction output is the vehicle's structured information such as model, Information such as year and body color.
  • the invention takes less time, has high recognition efficiency, and can better meet the application requirements of actual scenarios.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

本发明提供了一种车辆信息识别方法、系统、存储器及处理器。其中所述车辆信息识别方法包括建立车辆信息数据集;基于yolov3网络对所述车辆信息数据集进行训练,得到可以识别所述车辆信息的多属性网络模型;运用得到的所述多属性网络模型来对待识别图像的车辆信息进行多属性预测,以识别所述待识别图像中的车辆结构化信息。在yolov3的基础上,设置影响网络模型性能的网络参数,使得收敛速度更快、网络模型的效果更优,直接将车辆本身的信息解析为每个独立的类,实现端到端的输入与输出,使得网络输出为每个车辆的结构化信息,从而车辆信息识别检测过程耗时少,识别效率高,普遍适用于车辆检测领域。

Description

一种车辆信息识别方法、系统、存储器及处理器 技术领域
本发明涉及信息识别技术领域,尤其涉及一种车辆信息识别方法、系统、存储器及处理器。
背景技术
目前,视频监控车辆信息系统作为面向城市公共安全综合管理的物联网应用中智慧安防和智慧交通的重要组成部分,面临着深度应用的巨大挑战。主要应用瓶颈是视频中车辆结构化信息如何快速高效提出。
技术问题
传统方法实现的视频监控车辆信息系统,应用卷积神经网络的方法,虽然能够提取输入的不同特征,其中包含浅层特征,深层特征,局部特征等,并根据特征信息进行分类。目前,它已经在图像分类上取得了良好的效果,但对于细粒度分类的实时性以及准确率还有改进空间,不能统一提取车辆信息,需要多次分析融合信息,过程繁琐,耗时长。
技术解决方案
本发明所要解决的技术问题是提供一种车辆信息识别方法、系统、存储器及处理器,能够在yolov3的基础上直接将车辆本身的信息解析为每个独立的类,实现端到端的输入与输出,使得网络输出为车辆的结构化信息,从而车辆信息识别检测过程耗时少,识别效率高。
为解决上述技术问题,一方面,本发明一实施例提供了一种车辆信息识别方法,包括:建立车辆信息数据集;设置基于yolov3网络参数;基于yolov3网络对所述车辆信息数据集进行训练,得到可以识别所述车辆信息的多属性网络模型;运用得到的所述多属性网络模型来对待识别图像的车辆信息进行多属性预测,以识别所述待识别图像中的车辆结构化信息。
优选地,所述建立车辆信息数据集包括:对所述车辆信息数据集进行预处理;对所述车辆信息数据集进行数据增强处理。
优选地,所述设置基于yolov3网络参数包括:设置基于yolov3网络的类参数为所述车辆信息数据集定义的类别数;设置基于yolov3网络的滤波器参数为3×(类参数+5)。
优选地,所述基于yolov3网络对所述车辆信息数据集进行训练,得到可以识别所述车辆信息的多属性网络模型包括:用打标工具对所述车辆信息数据集进行类别标定,生成.xml格式的标定文件;将.xml格式的标定文件转换为.txt文件,并生成存储所述车辆信息数据集中每张图像所在的路径的.txt文件;根据所述车辆信息数据集中的类别设定更改网络超参。
优选地,所述运用得到的所述多属性网络模型来对待识别图像的车辆信息进行多属性预测包括:将含有车辆信息的视频资源或者静态图像输入进训练好的网络模型,一次性输出所述车辆信息的结构化信息。
优选地,所述对所述车辆信息数据集进行预处理包括:统计所述车辆信息数据集,按不同的类别分别存放到不同的文件夹。
优选地,所述对所述车辆信息数据集进行数据增强处理包括:对所述车辆信息数据集的图像进行翻转处理,将对应的翻转图像另命名保存在所述车辆信息数据集的图像相同的文件夹下。
优选地,所述对所述车辆信息数据集进行数据增强处理包括:对所述车辆信息数据集的图像进行镜像处理,将对应的镜像图像另命名保存在所述车辆信息数据集的图像相同的文件夹下。
优选地,所述翻转处理的角度为-15°~15°。
优选地,所述将.xml格式的标定文件转换为相对应的.txt格式文件并生成存储所述所述车辆信息数据集中每张图像所在的路径的.txt文件包括:读取.xml格式的标定文件,得到包含所述车辆数据集的图像的宽高,所述车辆数据集的图像中包含的车辆的宽高及类别;将所述车辆数据集的图像中包含的车辆的宽高进行归一化处理;计算所述车辆数据集的图像中包含的车辆中心点坐标;按所述车辆数据集的图像中包含的车辆的类别序号、中心点坐标以及宽高信息存为.txt文本文件;生成存有所述车辆信息数据集样本的路径的.txt文本文件。
优选地,所述.txt文本文件保存在标签文件夹中。
优选地,所述将所述车辆数据集的图像中包含的车辆的宽高进行归一化处理指的是:将所述车辆数据集的图像中包含的车辆的宽高比转化为0 ~1之间的数值。
另一方面,本发明的一实施例提供了一种存储介质,所述存储介质包括存储的程序,其中,所述程序运行时执行上述的车辆信息识别方法。
另一方面,本发明的一实施例提供了一种处理器,所述处理器用于运行程序,其中,所述程序运行时执行上述的车辆信息识别方法。
另一方面,本发明的一实施例提供了一种车辆信息识别系统,包括:通过电连接的车辆信息数据集设置模块、基于yolov3网络参数设置模块、基于yolov3网络训练模块、基于yolov3网络预测模块;所述车辆信息数据集设置模块用于建立车辆信息数据集;所述基于yolov3网络参数设置模块用于设置基于yolov3网络参数;所述基于yolov3网络训练模块用于对所述车辆信息数据集进行训练,得到可以识别所述车辆信息的多属性网络模型;所述基于yolov3网络预测模块用于运用得到的所述多属性网络模型来对待识别图像的车辆信息进行多属性预测,以识别所述待识别图像中的车辆结构化信息。
优选地,所述车辆信息数据集设置模块还包括数据预处理模块和数据增强处理模块,所述数据预处理模块用于对所述车辆信息数据集进行预处理;所述数据增强处理模块用于对所述车辆信息数据集进行数据增强处理。
优选地,所述基于yolov3网络参数设置模块设置基于yolov3网络中的类参数为所述车辆信息数据集定义的类别数;设置基于yolov3网络参数中的滤波器参数为3×(类参数+5)。
优选地,所述数据预处理模块统计所述车辆信息数据集,按不同的类别分别存放到不同的文件夹。
优选地,所述数据增强处理模块对所述车辆信息数据集的图像进行翻转处理,将对应的翻转图像另命名保存在所述车辆信息数据集的图像相同的文件夹下。
优选地,所述数据增强处理模块对所述车辆信息数据集的图像进行镜像处理,将对应的镜像图像另命名保存在所述车辆信息数据集的图像相同的文件夹下。
有益效果
与现有技术相比,上述技术方案具有以下优点:在yolov3的基础上,设置影响网络模型性能的网络参数,使得收敛速度更快、网络模型的效果更优,直接将车辆本身的信息解析为每个独立的类,实现端到端的输入与输出,使得网络输出为车辆的结构化信息,从而车辆信息识别检测过程耗时少,识别效率高,普遍适用于车辆检测领域。
附图说明
为了更清楚地说明本发明实施例中的技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本发明的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动性的前提下,还可以根据这些附图获得其它的附图。
图1是本发明车辆信息识别方法流程图。
图2是本发明车辆信息识别方法一优选实施例流程图。
图3是本发明车辆信息识别系统结构图。
本发明的实施方式
下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其它实施例,都属于本发明保护的范围。
yolov3主要从三个方面来说明,网络的输入、结构、输出。
(1)网络输入:网络输入的大小一般为320*320,416*416,608*608。这个大小必须是32的整数倍数,yolov3有5次下采样,每次采样步长为2,所以网络的最大步幅(步幅指层的输入大小除以输出)为2^5=32。
(2)网络结构:一般先训练了一个darknet-53,训练这个主要是为了主要有两个目的:a.这个网路结构能在ImageNet有好的分类结果,从而说明这个网路能学习到好的特征(设计新的网络结构,这个相当于调整参数),b.为后续检测模型做初始化。darknet-53相对于ResNet-152和ResNet-101,darknet-53不仅在分类精度上差不多,计算速度还比ResNet-152和ResNet-101强多了,网络层数也比他们少。
Darknet-53采用了ResNet这种跳层连接方式,性能完全比ResNet-152和ResNet-101这两种深层网络好,原因是:网络的基本单元的差异;网络层数越少,参数少,需要的计算量少。
yolov3网路使用了darknet-53的前面的52层(没有全连接层)。yolov3这个网络是一个全卷积网络,大量使用残差的跳层连接。之前的工作中,采样一般都是使用size为2*2,步长(stride)为2的max-pooling或者average-pooling进行降采样。但在这个网络结构中,使用的是步长为2的卷积来进行降采样。同时,网络中使用了上采样、route操作,还在一个网络结构中进行3次检测。
 使用残差的结构的好处:深度模型一个关键的点就是能否正常收敛,残差这种结构能保证网络结构在很深的情况下,仍能收敛,模型能训练下去;网络越深,表达的特征越好,分类+检测的效果都会提升;残差中的1*1卷积,使用network in network的想法,大量的减少了每次卷积的channel,一方面减少了参数量(参数量越大,保存的模型越大),另一方面在一定程度上减少了计算量。
(3)网络输出:
a.首先先确定网络输出特征层的大小。比如输入为320*320时,则输出为320/32=10,因此输出为10*10大小的特征层(feature map),此时有10*10=100个cell;同理当输入为416*416时输出的特征层为13*13大小的特征层,13*13=169个cell;输入为608*608时,输出的feature map大小为19*19,cell有19*19=361个。进行每进行一次up-sample时,输出特征层扩大一倍。
b. Anchor box的确定。这个先验框不同于之前Faster-Rcnn和SSD那样人工设定,在yolov2和yolov3中,都采用了对图像中的object采用k-means聚类。
c. feature map中的每一个cell都会预测3个边界框(bounding box) ,每个bounding box都会预测三个东西:(1)每个框的位置(4个值,中心坐标tx和ty,框的高度bh和宽度bw),(2)一个objectness prediction ,(3)N个类别,coco数据集80类,voc20类。因此对于coco数据集,在网络输入为416*416时,网络的输出大小为13*13(3*(4+1+80))=43095
yolov3使用 sigmoid 函数进行中心坐标预测。这使得输出值在 0 和 1 之间。正常情况下,YOLO 不会预测边界框中心的确切坐标。它预测的是:与预测目标的网格单元左上角相关的偏移;并且使用feature map中的cell大小进行归一化。
当输入图像为416*416,如果中心的预测是 (0.4, 0.7),则第二个cell在 13 x 13 特征图上的相对坐标是 (1.4, 1.7),具体的位置x坐标还需要1.4乘以cell的宽,y坐标为1.7乘以cell的高。
Bounding box的宽度bw和高度bh
yolov3得出的预测 bw 和bh 使用图像的高和宽进行归一化,框的预测 bx 和 by 是 (0.3, 0.8),那么 13 x 13 特征图的实际宽和高是 (13 x 0.3, 13 x 0.8)。
d. 三次检测,每次对应的感受不同,32倍降采样的感受最大,适合检测大的目标,所以在输入为416*416时,每个cell的三个anchor box为(116 ,90); (156 ,198); (373 ,326)。16倍适合一般大小的物体,anchor box为(30,61); (62,45); (59,119)。8倍的感受最小,适合检测小目标,因此anchor box为(10,13); (16,30); (33,23)。所以当输入为416*416时,实际总共有(52*52+26*26+13*13)*3=10647个proposal box。
实施例一
图1是本发明车辆信息识别方法流程图。如图1所示,一种车辆信息识别方法,包括步骤:
S11、建立车辆信息数据集;
S12、设置基于yolov3网络参数;
S13、基于yolov3网络对车辆信息数据集进行训练,得到可以识别车辆信息的多属性网络模型;
S14、运用得到的多属性网络模型来对待识别图像的车辆信息进行多属性预测,以识别待识别图像中的车辆结构化信息。
采集一定数量的车辆信息样本,建立车辆信息数据集,设置基于yolov3的网络参数,对采集的车辆信息样本进行训练。训练结果得出可以识别车辆属性的多属性网络模型。对于待识别图像的车辆信息的结构化信息,如车型、年份、车身颜色等属性,进行一次性识别。
采用本发明车辆信息识别方法,通过网络训练,采用端对端的方式,可以一次性对大量的车辆信息进行一次性识别出各自车辆的结构化信息,识别速度快,效率高。
实施例二
图2是本发明车辆信息识别方法一优选实施例流程图。如图2所示,具体实施时,步骤S11建立车辆信息数据集包括:对车辆信息数据集进行预处理;对车辆信息数据集进行数据增强处理。车辆信息数据集可以是来自不同车辆的视频,也可以是包含不同车辆的静态图像集合。由于通常情况下,用来训练的车辆样本数量少,并且分布不均衡,所以需要对车辆信息数据集进行数据增强处理,这样可以增加用来训练的车辆样本数据量,训练的结果更可靠。
对车辆信息数据集进行预处理,如统计车辆信息数据集,按不同的类别分别存放到不同的文件夹。对车辆信息数据集进行数据增强处理,如对车辆信息数据集的某些图像进行依次翻转处理,将对应的翻转图像另命名保存在车辆信息数据集的某些图像相同的文件夹下。为了确保图像翻转的合理性,可以将翻转处理的角度设为-15°~15°。也可以对车辆信息数据集的某些图像进行镜像处理,将对应的镜像图像另命名保存在车辆信息数据集的某些图像相同的文件夹下。图像数量少的样本则角度变换相对更多样化,从而达到样本均衡以及样本多样化的效果。一般根据测试样本的数量和车辆信息数据集图像的具体大小来进行翻转或者镜像处理。
网络训练的输入过程:
由于本方法是基于监督学习的方法实现的,因此需要对样本制作相应的标签。
第一步:用标定工具labelImg对文件夹车辆信息数据集图像中所有样本进行标定。其中,将车辆结构化信息如车型,年份,以及车身颜色等所有类别都独立归属为一类,类别之间相互独立。标定好样本后,生成相对应的.xml格式的标定文件。
第二步:将生成的.xml格式的标定文件一一对应转换为.txt文件。通过读取.xml格式的标定文件中的信息,如车辆数据集的图像的宽高,车辆数据集的图像中包含的车辆的宽高以及车辆数据集的图像中包含的车辆的类别,将车辆数据集的图像中包含的车辆的宽高归一化处理,如将车辆数据集的图像中包含的车辆的宽高比例设置为0~1之间,同时计算其中心点坐标,并按照车辆数据集的图像中包含的车辆的类别序号,中心点坐标,以及宽高的形式存为.txt文件。将.xml格式文件格式转换为.txt格式文件,通过软件代码实现。
第三步:生成存有所有样本的路径的.txt文本文件,作为网络训练的输入。
网络参数设定:
基于yolov3网络模型中检测层的参数需要根据网络输入的类别数来定义。如果类别数改变,其对应的网络检测层的滤波器个数也相应更改。另外,学习速率对网络模型的性能影响深远。
第一步:网络中的种类参数等于所有样本定义的类别数。
第二步:网络参数中滤波器(filters)根据公式滤波器(filters)= 3×(类参数+5)计算得到。此设置方式是根据基于yolov3网络推荐的方式。
第三步:根据实验总结,学习速率直接影响网络模型的性能。经多次实验调整超参学习速率,例如可以不断尝试学习速率值可以为:0.001、0.0001、0.00001等,最终需要根据训练需要选择合适的数值,使得网络训练能够快速收敛,并且网络模型性能较优。
属性预测:
运用得到的多属性网络模型来对待识别图像的车辆信息进行多属性预测,将任意的车辆视频输入进多属性网络模型,输出的是需要预测的车辆的车型,年份,车身颜色等结构化信息。
通过采用本发明车辆信息识别方法,在yolov3的基础上直接将车辆本身的信息解析为每个独立的类,实现端到端的输入与输出,使得网络输出为车辆的结构化信息,从而车辆信息识别检测过程耗时少,识别效率高。
实施例三
本发明的实施例还提供了一种存储介质,该存储介质包括存储的程序,其中,上述程序运行时执行上述的车辆信息识别方法流程。
可选地,在本实施例中,上述存储介质可以被设置为存储用于执行以下车辆信息识别方法流程的程序代码:
S11、建立车辆信息数据集;
S12、设置基于yolov3网络参数;
S13、基于yolov3网络对车辆信息数据集进行训练,得到可以识别车辆信息的多属性网络模型;
S14、运用得到的多属性网络模型来对待识别图像的车辆信息进行多属性预测,以识别待识别图像中的车辆结构化信息。
可选地,在本实施例中,上述存储介质可以包括但不限于:U盘、只读存储器(Read-Only Memory,简称为ROM)、随机存取存储器(Random Access Memory,简称为RAM)、移动硬盘、磁碟或者光盘等各种可以存储程序代码的介质。
由此可见,通过采用本发明存储介质,内置的车辆信息识别方法流程的程序运行速度更快,从而快速高效完成所有车辆信息的识别。
实施例四
本发明的实施例还提供了一种处理器,该处理器用于运行程序,其中,该程序运行时执行上述的车辆信息识别方法中的步骤。
可选地,在本实施例中,上述程序用于执行以下步骤:
S11、建立车辆信息数据集;
S12、设置基于yolov3网络参数;
S13、基于yolov3网络对车辆信息数据集进行训练,得到可以识别车辆信息的多属性网络模型;
S14、运用得到的多属性网络模型来对待识别图像的车辆信息进行多属性预测,以识别待识别图像中的车辆结构化信息。
可选地,本实施例中的具体示例可以参考上述实施例及具体实施时所描述的示例,本实施例在此不再赘述。
由此可见,通过采用本发明处理器,内置的车辆信息识别方法流程的程序运行速度更快,从而快速高效完成所有车辆信息的识别。
实施例五
图3是本发明车辆信息识别系统结构图。如图3所示,一种车辆信息识别系统,包括:通过电连接的车辆信息数据集设置模块、基于yolov3网络参数设置模块、基于yolov3网络训练模块、基于yolov3网络预测模块;
车辆信息数据集设置模块用于建立车辆信息数据集;
基于yolov3网络参数设置模块用于设置基于yolov3网络参数;
基于yolov3网络训练模块用于对车辆信息数据集进行训练,得到可以识别车辆信息的多属性网络模型;
基于yolov3网络预测模块用于运用得到的多属性网络模型来对待识别图像的车辆信息进行多属性预测,以识别待识别图像中的车辆结构化信息。
具体实施时,车辆信息数据集设置模块还包括数据预处理模块和数据增强处理模块,数据预处理模块用于对车辆信息数据集进行预处理,例如按不同的类别分别存放到不同的文件夹。数据增强处理模块用于对车辆信息数据集进行数据增强处理,例如对车辆信息数据集的图像进行翻转处理,将对应的翻转图像另命名保存在车辆信息数据集的图片相同的文件夹下。考虑到数据增强处理的有效性,翻转处理的角度可以根据需要设为-15°~15°。数据增强处理模块还可以对车辆信息数据集的图像进行镜像处理,将对应的镜像图像另命名保存在车辆信息数据集的图片相同的文件夹下。假如车辆信息数据集个数为n(n>1),其中有一半的数据需要进行翻转处理,另一半的数据需要进行镜像处理,则最后形成的车辆数据集个数为2n个,这样可以增加进行训练的样本的个数,使得后续网络训练精度更高。
基于yolov3网络参数设置模块设置基于yolov3网络中的类参数为所有样本定义的类别数;设置基于yolov3网络参数中的滤波器参数为3×(类参数+5);根据多次实验结果,设置基于yolov3网络参数中的学习速率参数。
基于yolov3网络训练模块在对车辆信息数据集进行训练过程中,用打标工具对所述车辆信息数据集进行类别标定,生成.xml格式的标定文件;将.xml格式的标定文件转换为对应的txt文件,并生成存储所述车辆信息数据集中每张图像路径的.txt文本文件;根据所述车辆信息数据集中的类别设定更改网络超参。所述将.xml格式的标定文件转换为对应的txt文件,并生成存储所述车辆信息数据集中每张图像路径的.txt文本文件包括:读取.xml格式的标定文件,得到包含所述车辆数据集的图像的宽高,所述车辆数据集的图像中包含的车辆的宽高及类别;将所述车辆数据集的图像中包含的车辆的宽高进行归一化处理;计算所述车辆数据集的图像中包含的车辆中心点坐标;按类别序号、车辆中心点坐标及所述车辆数据集的图像中包含的车辆的宽高信息存为.txt文本文件;生成存有所有样本的路径的.txt文本文件。所述.txt文本文件保存在标签文件夹中。车辆数据集的图像中包含的车辆的宽高进行归一化处理指的是:将所述车辆数据集的图像中包含的车辆的宽高比转化为0 ~1之间的数值。
基于yolov3网络预测模块在预测中,将大量的含有车辆信息的视频资源或者静态图像输入进训练好的网络模型,一次性输出车辆信息的结构化信息,如车型、年份、车身颜色等。车型如国内的:北汽、比亚迪、昌河、长城、东风、东南、飞碟、福田、富迪、哈飞、航天、红旗、华普、华翔、吉澳、吉利、瑞风、江铃、金杯、猎豹、南汽、奇瑞、汇众、申龙、万丰、上汽、曙光、双环、五菱、夏利、金龙、一汽、依维柯、中华、中顺、中兴、力
Figure 5fdb
、观志、传奇、猎豹、陆丰等;德国:宝马、奥迪、保时捷、奔驰、大众;英国:劳斯来斯、本特利、美洲虎、摩根、眼镜蛇、莲花等;意大利:菲亚特、法拉利、兰博基尼、玛沙拉蒂、兰旗等;法国:雷诺、标志、雪铁龙等;瑞士:绅宝、沃尔沃等;美国:通用、克莱斯勒、福特等;日本:丰田、本田、五十铃、马自达、铃木、日产、富士、三菱等;韩国:现代、双龙、大宇等。年份如:A-2010;B-2011;C-2012;D-2013;E-2014;F-2015;G-2016;H-2017;J-2018;K-2019;L-2020;M-2021;N-2022;P-2023;R-2024;S-2025;T-2026;V-2027;W-2028;X-2029;Y-2030;1-2031;2-2032;3-2033;4-2034;5-2035; 6-2036;7-2037;8-2038;9-2039;A-2040等。车身颜色如:A:白 B:灰 C:黄 D:粉 E:红 F:紫 G:绿 H:蓝 I:棕 J:黑 Z:其他。如大量的含有车辆信息的视频资源或者静态图像输入进训练好的网络模型,其中有一张图像一次性输出车辆信息的结构化信息为:奇瑞、B、A,则该图像所代表的车辆结构化信息为:车型为国产奇瑞,年份B指的是该车辆生产于2011年,车身颜色A指的是该车辆车身颜色是白色。
现有设计中,对于车辆信息的识别,车辆信息如车型、年份、车身颜色等有用到基于深度学习的方法的,但是基本都是单独处理的,这样耗时多,网络训练的数据量更大。采用本发明的车辆信息识别方法、系统、存储器及处理器,通过一次性处理输入车辆视频信息,中间没有任何融合的过程,再一次性输出车辆信息的各种不同的属性值集合,如车型识别、年份识别、车身颜色识别等。基于yolov3网络模型,将车辆信息如车型、年份、车身颜色等都分为不同的类别,端到端训练网络模型,通过训练好的网络模型,一次性预测输出为车辆的结构化信息如车型、年份、车身颜色等信息。两者相比,本发明耗时少, 识别效率高,更能满足实际场景应用需求。
工业实用性
以上对本发明实施例进行了详细介绍,本文中应用了具体个例对本发明的原理及实施方式进行了阐述,以上实施例的说明只是用于帮助理解本发明的方法及其核心思想;同时,对于本领域的一般技术人员,依据本发明的思想,在具体实施方式及应用范围上均会有改变之处,综上所述,本说明书内容不应理解为对本发明的限制。

Claims (20)

  1. 一种车辆信息识别方法,其特征在于,包括:
    建立车辆信息数据集;
    设置基于yolov3网络参数;
    基于yolov3网络对所述车辆信息数据集进行训练,得到可以识别所述车辆信息的多属性网络模型;
    运用得到的所述多属性网络模型来对待识别图像的车辆信息进行多属性预测,以识别所述待识别图像中的车辆结构化信息。
  2. 根据权利要求1所述的车辆信息识别方法,其特征在于,所述建立车辆信息数据集包括:
    对所述车辆信息数据集进行预处理;
    对所述车辆信息数据集进行数据增强处理。
  3. 根据权利要求1所述的车辆信息识别方法,其特征在于,所述设置基于yolov3网络参数包括:
    设置基于yolov3网络的类参数为所述车辆信息数据集定义的类别数;
    设置基于yolov3网络的滤波器参数为3×(类参数+5)。
  4. 根据权利要求1所述的车辆信息识别方法,其特征在于,所述基于yolov3网络对所述车辆信息数据集进行训练,得到可以识别所述车辆信息的多属性网络模型包括:
    用打标工具对所述车辆信息数据集进行类别标定,生成.xml格式的标定文件;
    将.xml格式的标定文件转换为.txt文件,并生成存储所述车辆信息数据集中每张图像所在的路径的.txt文件;
    根据所述车辆信息数据集中的类别设定更改网络超参。
  5. 根据权利要求1所述的车辆信息识别方法,其特征在于,所述运用得到的所述多属性网络模型来对待识别图像的车辆信息进行多属性预测包括:将含有所述车辆信息的视频资源或者静态图像输入进训练好的网络模型,一次性输出所述车辆信息的结构化信息。
  6. 根据权利要求2所述的车辆信息识别方法,其特征在于,所述对所述车辆信息数据集进行预处理包括:
    统计所述车辆信息数据集,按不同的类别分别存放到不同的文件夹。
  7. 根据权利要求2所述的车辆信息识别方法,其特征在于,所述对所述车辆信息数据集进行数据增强处理包括:
    对所述车辆信息数据集的图像进行翻转处理,将对应的翻转图像另命名保存在所述车辆信息数据集的图像相同的文件夹下。
  8. 根据权利要求2所述的车辆信息识别方法,其特征在于,所述对所述车辆信息数据集进行数据增强处理包括:
    对所述车辆信息数据集的图像进行镜像处理,将对应的镜像图像另命名保存在所述车辆信息数据集的图像相同的文件夹下。
  9. 根据权利要求7所述的车辆信息识别方法,其特征在于,所述翻转处理的角度为-15°~15°。
  10. 根据权利要求4所述的车辆信息识别方法,其特征在于,所述将.xml格式的标定文件转换为.txt文件,并生成存储车辆信息数据集中每张图像所在的路径的.txt文件包括:
    读取.xml格式的标定文件,得到包含所述车辆数据集的图像的宽高,所述车辆数据集的图像中包含的车辆的宽高及类别;
    将所述车辆数据集的图像中包含的车辆的宽高进行归一化处理;
    计算所述车辆数据集的图像中包含的车辆中心点坐标;
    按所述车辆数据集的图像中包含的车辆的类别序号、中心点坐标以及宽高存为.txt文本文件;
    生成存有所述车辆信息数据集样本的路径的.txt文本文件。
  11. 根据权利要求10所述的车辆信息识别方法,其特征在于,所述.txt文本文件保存在标签文件夹中。
  12. 根据权利要求10所述的车辆信息识别方法,其特征在于,所述将所述车辆数据集的图像中包含的车辆的宽高进行归一化处理指的是:
    将所述车辆数据集的图像中包含的车辆的宽高比转化为0 ~1之间的数值。
  13. 一种存储介质,其特征在于,所述存储介质包括存储的程序,其中,所述程序运行时执行权利要求1至12中任一项所述的车辆信息识别方法。
  14. 一种处理器,其特征在于,所述处理器用于运行程序,其中,所述程序运行时执行权利要求1至12中任一项所述的车辆信息识别方法。
  15. 一种车辆信息识别系统,其特征在于,包括:通过电连接的车辆信息数据集设置模块、基于yolov3网络参数设置模块、基于yolov3网络训练模块、基于yolov3网络预测模块;
    所述车辆信息数据集设置模块用于建立车辆信息数据集;
    所述基于yolov3网络参数设置模块用于设置基于yolov3网络参数;
    所述基于yolov3网络训练模块用于对所述车辆信息数据集进行训练,得到可以识别所述车辆信息的多属性网络模型;
    所述基于yolov3网络预测模块用于运用得到的所述多属性网络模型来对待识别图像的车辆信息进行多属性预测,以识别所述待识别图像中的车辆结构化信息。
  16. 根据权利要求15所述的车辆信息识别系统,其特征在于,所述车辆信息数据集设置模块还包括数据预处理模块和数据增强处理模块,
    所述数据预处理模块用于对所述车辆信息数据集进行预处理;
    所述数据增强处理模块用于对所述车辆信息数据集进行数据增强处理。
  17. 根据权利要求15所述的车辆信息识别系统,其特征在于,所述基于yolov3网络参数设置模块设置基于yolov3网络中的类参数为所述车辆信息数据集定义的类别数;设置基于yolov3网络参数中的滤波器参数为3×(类参数+5)。
  18. 根据权利要求16所述的车辆信息识别系统,其特征在于,所述数据预处理模块统计所述车辆信息数据集,按不同的类别分别存放到不同的文件夹。
  19. 根据权利要求16所述的车辆信息识别系统,其特征在于,所述数据增强处理模块对所述车辆信息数据集的图像进行翻转处理,将对应的翻转图像另命名保存在所述车辆信息数据集的图像相同的文件夹下。
  20. 根据权利要求16所述的车辆信息识别系统,其特征在于,所述数据增强处理模块对所述车辆信息数据集的图像进行镜像处理,将对应的镜像图像另命名保存在所述车辆信息数据集的图像相同的文件夹下。 
PCT/CN2019/112504 2018-12-07 2019-10-22 一种车辆信息识别方法、系统、存储器及处理器 WO2020114120A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811500432.3A CN111291779A (zh) 2018-12-07 2018-12-07 一种车辆信息识别方法、系统、存储器及处理器
CN201811500432.3 2018-12-07

Publications (1)

Publication Number Publication Date
WO2020114120A1 true WO2020114120A1 (zh) 2020-06-11

Family

ID=70974906

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/112504 WO2020114120A1 (zh) 2018-12-07 2019-10-22 一种车辆信息识别方法、系统、存储器及处理器

Country Status (2)

Country Link
CN (1) CN111291779A (zh)
WO (1) WO2020114120A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112906497A (zh) * 2021-01-29 2021-06-04 中国海洋大学 一种基于嵌入式的安全帽检测方法和设备
CN115984274A (zh) * 2023-03-20 2023-04-18 菲特(天津)检测技术有限公司 一种车辆外观检测模型、构建方法及检测方法
CN115984840A (zh) * 2022-12-20 2023-04-18 航天科技控股集团股份有限公司 一种保障系统用油料车的身份信息识别系统及方法

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112347693B (zh) * 2020-10-26 2023-12-22 上海感探号信息科技有限公司 一种车辆运行动态镜像模拟方法、装置和系统
CN113191426A (zh) * 2021-04-28 2021-07-30 深圳市捷顺科技实业股份有限公司 一种车辆识别模型创建方法、车辆识别方法及相关组件

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106599869A (zh) * 2016-12-22 2017-04-26 安徽大学 一种基于多任务卷积神经网络的车辆属性识别方法
CN107578007A (zh) * 2017-09-01 2018-01-12 杭州电子科技大学 一种基于多特征融合的深度学习人脸识别方法
CN108764148A (zh) * 2018-05-30 2018-11-06 东北大学 基于监控视频的多区域实时动作检测方法
CN108875754A (zh) * 2018-05-07 2018-11-23 华侨大学 一种基于多深度特征融合网络的车辆再识别方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106874840B (zh) * 2016-12-30 2019-10-22 东软集团股份有限公司 车辆信息识别方法及装置
CN107092922B (zh) * 2017-03-13 2018-08-31 平安科技(深圳)有限公司 车损识别方法及服务器
CN108537117B (zh) * 2018-03-06 2022-03-11 哈尔滨思派科技有限公司 一种基于深度学习的乘客检测方法和系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106599869A (zh) * 2016-12-22 2017-04-26 安徽大学 一种基于多任务卷积神经网络的车辆属性识别方法
CN107578007A (zh) * 2017-09-01 2018-01-12 杭州电子科技大学 一种基于多特征融合的深度学习人脸识别方法
CN108875754A (zh) * 2018-05-07 2018-11-23 华侨大学 一种基于多深度特征融合网络的车辆再识别方法
CN108764148A (zh) * 2018-05-30 2018-11-06 东北大学 基于监控视频的多区域实时动作检测方法

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112906497A (zh) * 2021-01-29 2021-06-04 中国海洋大学 一种基于嵌入式的安全帽检测方法和设备
CN115984840A (zh) * 2022-12-20 2023-04-18 航天科技控股集团股份有限公司 一种保障系统用油料车的身份信息识别系统及方法
CN115984274A (zh) * 2023-03-20 2023-04-18 菲特(天津)检测技术有限公司 一种车辆外观检测模型、构建方法及检测方法

Also Published As

Publication number Publication date
CN111291779A (zh) 2020-06-16

Similar Documents

Publication Publication Date Title
WO2020114120A1 (zh) 一种车辆信息识别方法、系统、存储器及处理器
CN109598279B (zh) 基于自编码对抗生成网络的零样本学习方法
Rao et al. Distracted driving recognition method based on deep convolutional neural network
Zhao et al. Adaptive logit adjustment loss for long-tailed visual recognition
Rani et al. Object detection and recognition using contour based edge detection and fast R-CNN
Wang et al. Multifocus image fusion using convolutional neural networks in the discrete wavelet transform domain
CN105809121A (zh) 多特征协同的交通标志检测与识别方法
CN113255793B (zh) 一种基于对比学习的细粒度舰船识别方法
CN108830254B (zh) 一种基于数据均衡策略和密集注意网络的细粒度车型检测与识别方法
CN114998695A (zh) 一种提高图像识别速度的方法及系统
Xu et al. BANet: A balanced atrous net improved from SSD for autonomous driving in smart transportation
CN110059541A (zh) 一种驾驶中的手机使用行为检测方法及装置
CN110909656B (zh) 一种雷达与摄像机融合的行人检测方法和系统
Kim et al. Detection of centerline crossing in abnormal driving using CapsNet
Wang et al. Material-aware Cross-channel Interaction Attention (MCIA) for occluded prohibited item detection
CN115588126A (zh) 一种融合GAM、CARAFE和SnIoU的车辆目标检测方法
Li et al. An end-to-end multi-task learning model for drivable road detection via edge refinement and geometric deformation
Zhang et al. Multi-scale vehicle logo detector
CN106570514A (zh) 一种基于词袋模型和支持向量机的汽车轮毂分类方法
Chen et al. Road object detection using a disparity-based fusion model
Wang et al. YOLO-ERF: lightweight object detector for UAV aerial images
CN115359442A (zh) 基于部件表征学习和个性化属性结构的车辆重识别方法
CN112131996B (zh) 基于通道分离卷积的路侧图像多尺度行人快速检测方法
Deng et al. A coarse to fine framework for recognizing and locating multiple diatoms with highly complex backgrounds in forensic investigation
CN114328174A (zh) 一种基于对抗学习的多视图软件缺陷预测方法及系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19893812

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19893812

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 19893812

Country of ref document: EP

Kind code of ref document: A1