WO2021238019A1 - 基于Ghost卷积特征融合神经网络实时车流量检测系统及方法 - Google Patents

基于Ghost卷积特征融合神经网络实时车流量检测系统及方法 Download PDF

Info

Publication number
WO2021238019A1
WO2021238019A1 PCT/CN2020/120742 CN2020120742W WO2021238019A1 WO 2021238019 A1 WO2021238019 A1 WO 2021238019A1 CN 2020120742 W CN2020120742 W CN 2020120742W WO 2021238019 A1 WO2021238019 A1 WO 2021238019A1
Authority
WO
WIPO (PCT)
Prior art keywords
layer
ghost
training set
network
real
Prior art date
Application number
PCT/CN2020/120742
Other languages
English (en)
French (fr)
Inventor
张莉
于厚舜
屈蕴茜
王邦军
孙涌
Original Assignee
苏州大学
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 苏州大学 filed Critical 苏州大学
Publication of WO2021238019A1 publication Critical patent/WO2021238019A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20212—Image combination
    • G06T2207/20221—Image fusion; Image merging
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30204—Marker
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30242—Counting objects in image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30248—Vehicle exterior or interior
    • G06T2207/30252—Vehicle exterior; Vicinity of vehicle

Definitions

  • the invention relates to the technical field of vehicle flow detection, in particular to a real-time vehicle flow detection system and method based on ghost convolution feature fusion neural network.
  • the existing vehicle flow detection methods can be roughly classified into three categories:
  • the first category is to bury an induction coil in the ground. When a vehicle passes through the induction coil area, the induction coil device is subjected to pressure to send out a pulse signal to determine whether there is a vehicle.
  • Commonly used methods include loop coil detection method, geomagnetic detection method, electromagnetic detection method and so on. This type of method requires more damage to the ground during installation and maintenance, and has a higher installation cost. At the same time, it also causes inconvenience to the travel of the vehicle.
  • the second category is the suspension detection method. When the vehicle enters the detection range, the detector emits microwaves, ultrasonic waves or infrared rays to determine whether there is a vehicle.
  • This type of detection method is easy to install and will not cause damage to the road surface. However, when two vehicles enter the detection range side by side, only one vehicle can be counted. The detection accuracy is affected and the detection range is small.
  • the third category is to shoot traffic video through a camera, and use computer vision technology and image processing technology to process the video. Compared with the first two types of traditional methods, the video detection method has obvious advantages: the installation and maintenance of the device are convenient, the road conditions of the entire road can be photographed, the detection range is relatively wide, and the acquired traffic information is relatively comprehensive.
  • Target detection and classification based on deep learning has entered a new stage.
  • convolutional neural networks have strong generalization, can overcome the difficulty of changing the appearance of vehicles, and can be constructed with feature descriptions driven by adaptive training data, which has greater flexibility and comprehensive capabilities .
  • most video-based vehicle flow detection methods adopt the target detection scheme.
  • the advantage of the target detection scheme is that the algorithm can not only count the number of vehicles, but also give the position of the vehicle in the image.
  • the slow speed of methods based on target detection is that these methods consume most of the computing power on vehicle positioning and type recognition. In fact, this positioning and classification for the detection of traffic flow is unnecessary, and the counting method based on target detection The recognition rate will be greatly reduced in a scene with heavy traffic and heavy vehicle overlap and occlusion.
  • Shi et al. proposed a lightweight convolutional neural network (Compact Convolutional Neural Network, C-CNN) for real-time detection of crowd density.
  • This network has low hardware performance requirements and can also be used for vehicle flow counting. It is also a technology in the field of target counting.
  • the C-CNN network structure is relatively simple and as a lightweight network itself does not use model compression technology, so the detected The error is large and good performance cannot be obtained.
  • the technical problem to be solved by the present invention is to overcome the problem of large error after detection in the prior art and failure to obtain good performance, so as to provide a ghost-based convolution feature with small error after detection and good performance. Fusion neural network real-time vehicle flow detection system and method.
  • a real-time vehicle flow detection system based on ghost convolution feature fusion neural network of the present invention includes: a data preprocessing module for extracting frames from the captured video to obtain the first training set, The target images in the first training set are labeled to form a second training set, and then a Gaussian filter is used to generate a density map of the target image to form a third training set; the network training module trains the network through the third training set Obtain a network model, where the network includes multiple ghost convolutional layers; a target information prediction module, which extracts frames from the captured video as a test image, inputs the test image into the network model for prediction, and obtains The target information of the test image.
  • the network includes a first layer, a second layer, a third layer, a fourth layer, and a fifth layer, wherein the first layer is a three-column convolutional layer, and the three The sizes of the convolution kernels of the column convolution layers are different, the second layer is a convolution layer with five layers of convolution kernels all the same size, and the fifth layer is a feature fusion layer with two layers of convolution kernels having the same size.
  • the feature maps obtained by the three-column convolution of the first layer are concatenated and then processed by the maximum pooling layer.
  • the feature maps obtained by convolution of the third layer and the fourth layer are processed by the maximum pooling layer.
  • the network model includes a loss function.
  • the Gaussian filter after the Gaussian filter is used to generate the density map of the target image, it further includes the step of normalizing all the images.
  • the network model further includes an activation function.
  • the present invention also provides a real-time vehicle flow detection method based on ghost convolution feature fusion neural network, which includes the following steps: Step S1: extract frames from the captured video to obtain a first training set, and compare the first training set to the first training set. Mark the target image to form a second training set, and then use a Gaussian filter to generate a density map of the target image to form a third training set; step S2: train the network through the third training set to obtain a network model, wherein The network includes multiple ghost convolutional layers; step S3: extract frames from the captured video as test images, input the test images into the network model for prediction, and obtain target information of the test images.
  • the method for labeling the target images in the first training set is: using a labeling tool to label the target images in the first training set.
  • a predicted density map of the test image is obtained, and the target information of the test image can be obtained by summing the predicted density map.
  • the target position information that has been marked in the image is used, and the Gaussian filter is used to generate the density map of the target image.
  • the Gaussian filter is used to generate the density map of the target image.
  • normalize all images in the network training module, use the processed images and the generated density map to train the network; in the target information prediction module, given an image, use the trained network to predict the image Target information.
  • the network of the present invention includes multiple ghost convolutional layers, so that the network can reduce the amount of parameters while ensuring the original performance, reduce the consumption of hardware resources and speed up the running speed; not only the method is simple, but the error after detection is small , Can get good performance.
  • Figure 1 is a flow chart of the real-time vehicle flow detection system based on ghost convolution feature fusion neural network of the present invention
  • Figure 2 is a schematic diagram of the network model of the present invention.
  • Figure 3 is a comparison of the target count results of the present invention on the TRANCOS data set
  • Figure 4 is a comparison of the parameter values of the target count results of the present invention on the TRANCOS data set
  • Figure 5 is a flowchart of a real-time vehicle flow detection method based on ghost convolution feature fusion neural network.
  • this embodiment provides a real-time vehicle flow detection system based on ghost convolutional feature fusion neural network, including: a data preprocessing module 10 for extracting frames from the captured video to obtain the first A training set, marking the target images in the first training set to form a second training set, and then using a Gaussian filter to generate a density map of the target image to form a third training set; the network training module 20 passes the third training set The training set trains the network to obtain a network model, where the network includes multiple ghost convolutional layers; the target information prediction module 30 extracts frames from the captured video as a test image, and inputs the test image to the network Prediction is made in the model, and the target information of the test image is obtained.
  • a data preprocessing module 10 for extracting frames from the captured video to obtain the first A training set, marking the target images in the first training set to form a second training set, and then using a Gaussian filter to generate a density map of the target image to form a third training set
  • the network training module 20 passes the
  • the real-time vehicle flow detection system based on ghost convolution feature fusion neural network in this embodiment includes: a data preprocessing module 10, which is used to extract frames from a captured video to obtain a first training set, and perform training on the first training set.
  • the concentrated target images are labeled to form the second training set, and then the Gaussian filter is used to generate the density map of the target image to form the third training set, which is conducive to training the network;
  • the network training module 20 uses the third training set to The network is trained to obtain a network model, where the network includes multiple ghost convolutional layers, so that the network can reduce the amount of parameters while ensuring the original performance, reduce the consumption of hardware resources, and speed up the operation; target information prediction module 30 Extract frames from the captured video and use it as a test image. Input the test image into the network model for prediction, and obtain the target information of the test image. Not only is the method simple, but the error after detection is small, and good results are obtained. Performance.
  • frames are extracted from the video shot by the surveillance camera, and the training set is obtained by extracting the frames
  • X i is the i-th training set images
  • the size of m ⁇ n is the number of training sets.
  • Training set available after labeling Wherein X i is the i-th training set images, the size of m ⁇ n; P i is the i-th information center coordinates of the target images, the size c i ⁇ 2, wherein the first center point as the target in the image The abscissa, the second column is the ordinate of the target center point in the image, c i is the number of targets in the image; N is the number of training sets.
  • the Gaussian filter After the Gaussian filter is used to generate the density map of the target image, it also includes the step of normalizing all the images, converting the pixel value of each channel of the input image from 0-255 to 0-1.
  • the unified processing can speed up the solution speed of gradient descent and improve the convergence speed of the model.
  • the processed training set can be obtained Used to train the network.
  • the network includes a first layer, a second layer, a third layer, a fourth layer, and a fifth layer, wherein the first layer is a three-column convolutional layer And the convolution kernel sizes of the three-column convolution layers are different, and the multi-column structure with different convolution kernel sizes enables the network to capture feature information of different scales, which is beneficial to improve the accuracy of target counting;
  • the second layer It is a convolutional layer with five layers of convolution kernels all having the same size;
  • the fifth layer is a feature fusion layer with two layers of convolution kernels having the same size, which helps to effectively improve the quality of the density map generated by the network.
  • the first layer of the network is a three-column convolutional layer, the channel numbers of the three columns are 10, 14 and 16, respectively, the convolution kernel sizes are 9, 7 and 5 respectively, and the multi-column structure with different convolution kernel sizes It enables the network to capture feature information of different scales and improves the accuracy of target counting.
  • the feature maps obtained by the three-column convolution of the first layer are concatenated and processed by a maximum pooling layer.
  • the network model also includes an activation function. All activation functions used in the network of the present invention are ReLU activation functions, and all convolutional layers are followed by batch normalization layers. Since if the activation function is not introduced into the network, even if the number of network layers is increased, it will only form a linear transformation, and the complexity of the linear function is limited, so the ability to learn complex function mapping from the data is small.
  • the present invention introduces an activation function, which performs a nonlinear conversion on the feature map, and all the feature maps generated by the convolutional layer are processed by the batch normalization layer and then processed by the ReLU activation function, so it is beneficial to improve the learning of complex function mapping. ability.
  • the network model includes a loss function.
  • the definition of the loss function is: Wherein, N is the number of training samples, X i is the i-th training sample, ⁇ is the parameter learning, F (X i; ⁇ ) is the i-th network density map prediction samples, M i is the i th The true density map of the sample.
  • frames are extracted from the video shot by the surveillance camera as a test image.
  • test image P input it into the trained network model for prediction, and the predicted density map M p of image P can be obtained.
  • the data set contains a total of 1244 annotated images, and a total of 46,796 vehicle center coordinate information has been annotated in the data set.
  • 823 images are used for training, and the remaining 421 images are used for testing.
  • the images are all taken from road monitoring and the image size is 480*640.
  • the training set with annotated information can be directly obtained
  • X i is the i TRANCOS images
  • size is 480 * 640
  • P i is the i-th information center coordinates of the target images, the size of c ⁇ 2, wherein the first center point as the target in the image The abscissa, the second column is the ordinate of the target center point in the image
  • c is the number of targets in the i-th image
  • N is the number of training sets in TRANCOS.
  • Training set D′ can get processed training set after step (2) Used to train the network.
  • the processed training set is used Trained on the feature fusion convolutional neural network shown in Figure 2.
  • the first layer of this network is a three-column convolutional layer, the channel numbers of the three columns are 10, 14 and 16, respectively, and the convolution kernel size is 9, 7 and 5 respectively.
  • the multi-column structure with different convolution kernel sizes enables the network to capture feature information of different scales and improve the accuracy of target counting.
  • the feature maps obtained by the three-column convolution of the first layer are concatenated and processed by a maximum pooling layer.
  • Next is a convolutional layer with five convolution kernel sizes of 3, and the feature maps obtained by convolution of the third and fourth layers are processed by the maximum pooling layer.
  • the loss function L( ⁇ ) of the network is defined as follows: Where N is the number of training samples, X i is the i-th training sample, ⁇ is the parameter learning, F (X i; ⁇ ) is the density of the network of FIG prediction samples i, M i is the i th sample The true density map.
  • the present invention proposes a feature fusion neural network structure based on ghost convolution based on the C-CNN network, which can not only capture feature information with inconsistent scales in images, but also fuse features of different scales and further extract and generate high-quality prediction density maps. Can further improve the target counting performance. In addition, it also greatly reduces the amount of network parameters, making the network more suitable for running on low-performance devices and mobile devices.
  • This network model and the C-CNN network model compare target counts on the same data set. It can be seen from the results in Figure 3 that the network model proposed by the present invention achieves better performance than the C-CNN model in terms of mean absolute error (MAE) and mean square error (MSE). At the same time, the parameters in Figure 4 Compared with C-CNN, the amount has also dropped significantly.
  • MAE mean absolute error
  • MSE mean square error
  • this embodiment provides a real-time traffic flow detection method based on ghost convolutional feature fusion neural network.
  • the principle of solving the problem is the same as that of the real-time traffic flow detection system based on ghost convolution feature fusion neural network. Repeat it again.
  • this embodiment provides a method for real-time traffic flow detection based on ghost convolution feature fusion neural network, which includes the following steps:
  • Step S1 Extract frames from the captured video to obtain a first training set, mark the target image in the first training set to form a second training set, and then use a Gaussian filter to generate a density map of the target image to form the first training set.
  • Step S2 Train the network through the third training set to obtain a network model, where the network includes multiple ghost convolutional layers;
  • Step S3 extracting frames from the captured video as a test image, inputting the test image into the network model for prediction, and obtaining target information of the test image.
  • this application can be provided as a method, a system, or a computer program product. Therefore, this application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
  • a computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device.
  • the device implements the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.
  • These computer program instructions can also be loaded on a computer or other programmable data processing equipment, so that a series of operation steps are executed on the computer or other programmable equipment to produce computer-implemented processing, so as to execute on the computer or other programmable equipment.
  • the instructions provide steps for implementing functions specified in a flow or multiple flows in the flowchart and/or a block or multiple blocks in the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (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)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

本发明涉及一种基于Ghost卷积特征融合神经网络实时车流量检测系统及方法,包括：数据预处理模块，用于从拍摄的视频中进行抽帧，得到第一训练集，对所述第一训练集中的目标图像进行标记形成第二训练集，再利用高斯滤波器生成目标图像的密度图，形成第三训练集；网络训练模块，通过所述第三训练集对网络进行训练得到网络模型，其中所述网络包括多个Ghost卷积层；目标信息预测模块，从拍摄的视频中进行抽帧，作为测试图像，将所述测试图像输入至所述网络模型中进行预测，得到测试图像的目标信息。本发明误差小，有利于获得良好的性能。

Description

基于Ghost卷积特征融合神经网络实时车流量检测系统及方法 技术领域
本发明涉及车流量检测的技术领域，尤其是指一种基于Ghost卷积特征融合神经网络实时车流量检测系统及方法。
背景技术
近年来，随着交通车辆数量高速增长，交通监管面临巨大挑战。随着智能城市智能交通系统(Intelligent Transport System,ITS)的推广，其应用的核心技术得到了迅速的发展。车流量检测，作为构建交通状况视频监控的一项关键技术，一直是其重要的组成部分。
现有车流量检测的方法大致可以归于三类：第一类是在地下埋感应线圈，当车辆经过感应线圈区域时，由感应线圈装置受到压力发出脉冲信号，从而判断是否有车辆。常用的方法有环形线圈检测法、地磁检测法、电磁检测法等。这类方法较在安装和维修时需要破坏地面，安装成本较高，同时也给车辆的出行造成了不便。第二类是悬挂式检测方法，当车辆驶入检测范围时，检测器发出的微波、超声波或红外线，以此来判断是否有车辆。这类检测方法易于安装，对路面不会造成破坏，但当两辆车并排进入检测范围时，只能计数一辆车，检测精度受到了影响，检测范围较小。第三类为通过摄像机拍摄交通视频，利用计算机视觉技术以及图像处理技术对视频进行处理。相对于前两类传统方法，视频检测方法具有明显的优势：安装和维护装置方便，能对整个路面的路况进行拍摄，检测到的范围比较广，获取到的交通信息比较全面。
传统的基于视觉的车辆检测方法通常是手动提取特征，费时费力，泛化能力差，易受环境变化影响。随着深度学习理论和实践的迅速发展，基于深 度学习的目标检测与分类进入了一个新的阶段。不同于传统的特征提取算法，卷积神经网络有很强的泛化性，可以克服改变车辆外观的困难，能够自适应的训练数据驱动下构建的特征描述，具有更大的灵活性和综合能力。目前多数的基于视频的车流量检测方法都是采用了目标检测的方案，采用目标检测的方案优点在于算法不但能够统计车辆的数量，更能给出车辆在图像中的位置。不过这种方案的缺点也很明显，那就是目标检测算法需要很高性能的硬件支持，才能满足其运行时的算力需求。虽然有SSD(Single ShotMultibox Detector)、YOLO(You Only Look Once)等其他快速目标检测网络致力于在保持高速度的同时实现高精度检测，但在低算力移动设备上，这类算法依旧很难满足实时检测车流量的需求。
基于目标检测的方法速度之慢在于这些方法将大部分的算力消耗在车辆的定位和种类识别上，实际上对于车流量的检测这种定位和分类是非必要的，并且基于目标检测的计数方法在车流量大、车辆重叠遮挡严重的场景下识别率会大打折扣。Shi等人在论文“A Real-Time Deep Network For Crowd Counting”中提出了一种实时检测人群密度的轻量化卷积神经网络(Compact Convolutional Neural Network，C-CNN)。该网络对硬件性能要求较低且同样可用于车流量计数，同属于目标计数的领域的技术，但是C-CNN网络结构较为简单且作为轻量化的网络本身没有使用模型压缩技术，因此检测后的误差大，不能获得良好的性能。
发明内容
为此，本发明所要解决的技术问题在于克服现有技术中检测后的误差大，不能获得良好性能的问题，从而提供一种检测后的误差小，且获得良好的性能的基于Ghost卷积特征融合神经网络实时车流量检测系统及方法。
为解决上述技术问题，本发明的一种基于Ghost卷积特征融合神经网络实时车流量检测系统，包括：数据预处理模块，用于从拍摄的视频中进行抽帧，得到第一训练集，对所述第一训练集中的目标图像进行标记形成第二训练集，再利用高斯滤波器生成目标图像的密度图，形成第三训练集；网络训 练模块，通过所述第三训练集对网络进行训练得到网络模型，其中所述网络包括多个Ghost卷积层；目标信息预测模块，从拍摄的视频中进行抽帧，作为测试图像，将所述测试图像输入至所述网络模型中进行预测，得到测试图像的目标信息。
在本发明的一个实施例中，所述网络包括第一层、第二层、第三层、第四层以及第五层，其中所述第一层是三列卷积层，且所述三列卷积层的卷积核大小不同，所述第二层是五层卷积核大小均相同的卷积层，所述第五层是两层卷积核大小均相同的特征融合层。
在本发明的一个实施例中，所述第一层的三列卷积得到的特征图连结后经过最大池化层处理。
在本发明的一个实施例中，所述第三层和第四层卷积得到的特征图均经过最大池化层处理。
在本发明的一个实施例中，所述网络模型包括损失函数。
在本发明的一个实施例中，利用高斯滤波器生成目标图像的密度图后，还包括对所有图像进行归一化处理的步骤。
在本发明的一个实施例中，所述网络模型还包括激活函数。
本发明还提供了一种基于Ghost卷积特征融合神经网络实时车流量检测方法，包括如下步骤：步骤S1:从拍摄的视频中进行抽帧，得到第一训练集，对所述第一训练集中的目标图像进行标记形成第二训练集，再利用高斯滤波器生成目标图像的密度图，形成第三训练集；步骤S2:通过所述第三训练集对网络进行训练得到网络模型，其中所述网络包括多个Ghost卷积层；步骤S3:从拍摄的视频中进行抽帧，作为测试图像，将所述测试图像输入至所述网络模型中进行预测，得到测试图像的目标信息。
在本发明的一个实施例中，对所述第一训练集中的目标图像进行标记的方法为:利用标注工具对所述第一训练集中的目标图像进行标记。
在本发明的一个实施例中，将所述测试图像输入至所述网络模型中进行 预测时，得到测试图像的预测密度图，对预测密度图进行求和运算就可以得到测试图像的目标信息。
本发明的上述技术方案相比现有技术具有以下优点：
本发明所述的基于Ghost卷积特征融合神经网络实时车流量检测系统及方法，在数据预处理模块中，使用图像中已经标注好的目标位置信息，利用高斯滤波器生成目标图像的密度图，并将所有图像归一化；在网络训练模块，使用处理好的图像及生成的密度图对网络进行训练；在目标信息预测模块中，给定一张图像，使用训练好的网络进行预测图像中的目标信息。本发明所述网络包括多个Ghost卷积层，使得所述网络在保证原有性能的同时能够降低参数量，降低对硬件资源的消耗并加快运行速度；不但方法简单，而且检测后的误差小，能获得良好的性能。
附图说明
为了使本发明的内容更容易被清楚的理解，下面根据本发明的具体实施例并结合附图，对本发明作进一步详细的说明，其中
图1是本发明基于Ghost卷积特征融合神经网络实时车流量检测系统流程图；
图2是本发明网络模型的示意图；
图3是本发明目标计数结果在TRANCOS数据集上的对比；
图4是本发明目标计数结果在TRANCOS数据集上参数量对比；
图5是基于Ghost卷积特征融合神经网络实时车流量检测方法的流程图。
说明书附图标记说明：10-数据预处理模块，20-网络训练模块，30-目标信息预测模块。
具体实施方式
实施例一
如图1和图2所示，本实施例提供一种基于Ghost卷积特征融合神经网络实时车流量检测系统，包括：数据预处理模块10，用于从拍摄的视频中进行抽帧，得到第一训练集，对所述第一训练集中的目标图像进行标记形成第二训练集，再利用高斯滤波器生成目标图像的密度图，形成第三训练集；网络训练模块20，通过所述第三训练集对网络进行训练得到网络模型，其中所述网络包括多个Ghost卷积层；目标信息预测模块30从拍摄的视频中进行抽帧，作为测试图像，将所述测试图像输入至所述网络模型中进行预测，得到测试图像的目标信息。
本实施例所述基于Ghost卷积特征融合神经网络实时车流量检测系统，包括：数据预处理模块10，用于从拍摄的视频中进行抽帧，得到第一训练集，对所述第一训练集中的目标图像进行标记形成第二训练集，再利用高斯滤波器生成目标图像的密度图，形成第三训练集，有利于对网络进行训练；网络训练模块20，通过所述第三训练集对网络进行训练得到网络模型，其中所述网络包括多个Ghost卷积层，使得所述网络在保证原有性能的同时能够降低参数量，降低对硬件资源的消耗并加快运行速度；目标信息预测模块30从拍摄的视频中进行抽帧，作为测试图像，将所述测试图像输入至所述网络模型中进行预测，得到测试图像的目标信息，不但方法简单，而且检测后的误差小，能获得良好的性能。
所述数据预处理模块10中，从监控摄像头拍摄的视频中进行抽帧，抽帧得到训练集
Figure PCTCN2020120742-appb-000001
其中X i为训练集中第i张图像，大小为m×n，N为训练集的数量。
使用标注工具标注训练集
Figure PCTCN2020120742-appb-000002
中图像X i中所有目标的中心位置。标注完成可得训练集
Figure PCTCN2020120742-appb-000003
其中X i为训练集中第i张图像，大小为m×n；P i是第i张图像中目标中心坐标的信息，大小为c i×2，其中第一列为目标中心点在图像中的横坐标，第二列为目标中心点在图像中的纵坐标，c i为该图像中的目标数；N为训练集的数量。
采用一个大小为15×15、方差为σ的高斯滤波器，根据P i中目标中心的 坐标信息，生成X i图像的密度图M i。
利用高斯滤波器生成目标图像的密度图后，还包括对所有图像进行归一化处理的步骤，将输入图像各通道的像素值由0-255转换到0-1的区间内，通过所述归一化处理能够加快梯度下降的求解速度，提升模型的收敛速度。
训练集D经过上述处理后就可以得到处理好的训练集
Figure PCTCN2020120742-appb-000004
用于训练网络。
如图2所示，所述网络训练模块20中，所述网络包括第一层、第二层、第三层、第四层以及第五层，其中所述第一层是三列卷积层，且所述三列卷积层的卷积核大小不同，卷积核大小不同的多列结构使得网络能够捕捉到不同尺度的特征信息，有利于提高目标计数的准确率；所述第二层是五层卷积核大小均相同的卷积层；所述第五层是两层卷积核大小均相同的特征融合层，从而有利于能够有效提升网络生成密度图的质量。
具体地，所有卷积层全部使用“same”模式填充，卷积层参数含义Ghost Conv-(卷积核数量)-(卷积核大小)。所述网络第一层是一个三列卷积层，三列的通道数分别为10、14和16，卷积核大小分别为9、7和5，所述卷积核大小不同的多列结构使得网络能够捕捉到不同尺度的特征信息，提高目标计数的准确率。第一层的三列卷积得到的特征图连结后经过一层最大池化层处理。接着是五层卷积核大小均为3的卷积层，其中所述第三层和第四层卷积得到的特征图均经过最大池化层处理。最后是两层卷积核大小为3的特征融合层，该层的作用是将网络提取到的不同尺度的特征进行融合并进一步提取生成最终预测密度图。由于网络全部采用Ghost卷积层，保证性能的同时网络参数量大幅下降，更适合部署在移动设备和其他低性能设备上运行。
所述网络模型还包括激活函数。本发明网络中使用的所有激活函数均为ReLU激活函数，所有卷积层后均接批标准化层。由于如果网络中没有引入激活函数，那么即使网络层数再多形成的也只是线性变换，而线性函数复杂性有限，因此从数据中学习复杂函数映射的能力小。本发明通过引入激活函数，所述激活函数对特征图进行非线性转换，所有的卷积层生成的特征图经 批标准化层处理后再经ReLU激活函数处理，因此有利于提高学习复杂函数映射的能力。
使用处理好的训练集
Figure PCTCN2020120742-appb-000005
训练图2所示网络，所述网络模型包括损失函数。所述损失函数的定义为：
Figure PCTCN2020120742-appb-000006
其中，N为训练样本的个数，X i为第i个训练样本，Θ为网络学习的参数，F(X i；Θ)为网络预测第i个样本的密度图，M i为第i个样本的真实密度图。
所述目标信息预测模块30中，从监控摄像头拍摄的视频中进行抽帧，作为测试图像。对任意测试图像P，将其输入到训练好的网络模型中进行预测，可以得到图像P的预测密度图M p，对M p进行求和运算得到的值即为图像P中预测的目标数：p_count＝sum(M p)。
下面结合目标计数数据集TRaffic ANd COngestionS(简称TRANCOS)上的测试具体说明：
所述数据集中一共包含1244张已标注的图像，数据集中共有46796个车辆中心坐标信息已标注。该数据集中823张用于训练，其余421张用于测试，其中的图像均是路面监控拍摄而得，图像尺寸均为480*640。
对于所述数据预处理模块10中，由于TRANCOS数据集图像均已标注,可直接得到有标注信息的训练集
Figure PCTCN2020120742-appb-000007
其中X i为TRANCOS中的第i张图像，大小为480*640；P i是第i张图像中目标中心坐标的信息，大小为c×2，其中第一列为目标中心点在图像中的横坐标，第二列为目标中心点在图像中的纵坐标，c为第i张图像中的目标数；N为TRANCOS中训练集的数量。
(2)采用一个大小为15*15、方差为σ的高斯滤波器，根据P i中目标中心的坐标信息，生成X i图像的密度图M i。
(3)训练集D′经过步骤(2)后就可以得到处理好的训练集
Figure PCTCN2020120742-appb-000008
用于训练网络。
所述网络训练模块20中，使用处理好的训练集
Figure PCTCN2020120742-appb-000009
在如 图2所示的特征融合卷积神经网络上进行训练。本网络第一层是一个三列卷积层，三列的通道数分别为10、14和16，卷积核大小分别为9、7和5。卷积核大小不同的多列结构使得网络能够捕捉到不同尺度的特征信息，提高目标计数的准确率。第一层的三列卷积得到的特征图连结后经过一层最大池化层处理。接下来是五层卷积核大小均为3的卷积层，其中第三层和第四层卷积得到的特征图均经过最大池化层处理。最后是两层卷积核大小为3的特征融合层，该层的作用是将网络提取到的不同尺度的特征进行融合并进一步提取生成最终预测密度图。由于网络全部采用Ghost卷积层，保证性能的同时网络参数量大幅下降，更适合部署在移动设备和其他低性能设备上运行。本发明网络中使用的所有激活函数均为ReLU激活函数，所有卷积层后均接批标准化层。
所述网络的损失函数L(Θ)定义如下：
Figure PCTCN2020120742-appb-000010
其中N为训练样本的个数，X i为第i个训练样本，Θ为网络学习的参数，F(X i；Θ)为网络预测第i个样本的密度图，M i为第i个样本的真实密度图。
所述目标信息预测模块30中，给定一张测试图像P，将图像P输入到训练好的网络模型中进行预测，可以得到图像P的预测密度图M p，对M p进行求和运算得到的值即为图像P中预测的目标：p_count＝sum(M p)。
本发明的效果可以通过如下实验验证：
本发明基于C-CNN网络提出了基于Ghost卷积的特征融合神经网络结构，它既可捕捉图像中尺度不一致的特征信息，也能够将不同尺度的特征融合并进一步提取生成高质量预测密度图。能够进一步提高目标计数性能。除此之外它还大幅降低网络参数量，使网络更适合运行在低性能设备和移动设备上。本网络模型与C-CNN网络模型在相同的数据集上做目标计数比较。从图3的结果中可以看出，本发明提出的网络模型在平均绝对误差(MAE)和均方误差(MSE)上都比C-CNN模型获得了更好的性能，同时在图4中参数量相较于C-CNN也有大幅下降。
实施例二
基于同一发明构思，本实施例提供了基于Ghost卷积特征融合神经网络实时车流量检测方法，其解决问题的原理与所述基于Ghost卷积特征融合神经网络实时车流量检测系统，重复之处不再赘述。
如图5所示，本实施例提供一种基于Ghost卷积特征融合神经网络实时车流量检测方法，包括如下步骤：
步骤S1：从拍摄的视频中进行抽帧，得到第一训练集，对所述第一训练集中的目标图像进行标记形成第二训练集，再利用高斯滤波器生成目标图像的密度图，形成第三训练集；
步骤S2：通过所述第三训练集对网络进行训练得到网络模型，其中所述网络包括多个Ghost卷积层；
步骤S3：从拍摄的视频中进行抽帧，作为测试图像，将所述测试图像输入至所述网络模型中进行预测，得到测试图像的目标信息。
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
显然，上述实施例仅仅是为清楚地说明所作的举例，并非对实施方式的限定。对于所属领域的普通技术人员来说，在上述说明的基础上还可以做出其它不同形式变化或变动。这里无需也无法对所有的实施方式予以穷举。而由此所引伸出的显而易见的变化或变动仍处于本发明创造的保护范围之中。

Claims (10)

  1. 一种基于Ghost卷积特征融合神经网络实时车流量检测系统，其特征在于，包括：
    数据预处理模块，用于从拍摄的视频中进行抽帧，得到第一训练集，对所述第一训练集中的目标图像进行标记形成第二训练集，再利用高斯滤波器生成目标图像的密度图，形成第三训练集；
    网络训练模块，通过所述第三训练集对网络进行训练得到网络模型，其中所述网络包括多个Ghost卷积层；
    目标信息预测模块，从拍摄的视频中进行抽帧，作为测试图像，将所述测试图像输入至所述网络模型中进行预测，得到测试图像的目标信息。
  2. 根据权利要求1所述的基于Ghost卷积特征融合神经网络实时车流量检测系统，其特征在于：所述网络包括第一层、第二层、第三层、第四层以及第五层，其中所述第一层是三列卷积层，且所述三列卷积层的卷积核大小不同，所述第二层是五层卷积核大小均相同的卷积层，所述第五层是两层卷积核大小均相同的特征融合层。
  3. 根据权利要求2所述的基于Ghost卷积特征融合神经网络实时车流量检测系统，其特征在于：所述第一层的三列卷积得到的特征图连结后经过最大池化层处理。
  4. 根据权利要求2所述的基于Ghost卷积特征融合神经网络实时车流量检测系统，其特征在于：所述第三层和第四层卷积得到的特征图均经过最大池化层处理。
  5. 根据权利要求1所述的基于Ghost卷积特征融合神经网络实时车流量检测系统，其特征在于：对所述第一训练集中的目标图像进行标记的方法为:利用标注工具对所述第一训练集中的目标图像进行标记。
  6. 根据权利要求1所述的基于Ghost卷积特征融合神经网络实时车流量检测系统，其特征在于：将所述测试图像输入至所述网络模型中进行预测时，得到测试图像的预测密度图，对预测密度图进行求和运算就可以得到测试图像的目标信息。
  7. 根据权利要求1所述的基于Ghost卷积特征融合神经网络实时车流量检测系统，其特征在于：所述网络模型包括损失函数。
  8. 根据权利要求1所述的基于Ghost卷积特征融合神经网络实时车流量检测系统，其特征在于：利用高斯滤波器生成目标图像的密度图后，还包括对所有图像进行归一化处理的步骤。
  9. 根据权利要求1所述的基于Ghost卷积特征融合神经网络实时车流量检测系统，其特征在于：所述网络模型还包括激活函数。
  10. 一种基于Ghost卷积特征融合神经网络实时车流量检测方法，其特征在于，包括如下步骤：
    步骤S1:从拍摄的视频中进行抽帧，得到第一训练集，对所述第一训练集中的目标图像进行标记形成第二训练集，再利用高斯滤波器生成目标图像的密度图，形成第三训练集；
    步骤S2:通过所述第三训练集对网络进行训练得到网络模型，其中所述网络包括多个Ghost卷积层；
    步骤S3:从拍摄的视频中进行抽帧，作为测试图像，将所述测试图像输入至所述网络模型中进行预测，得到测试图像的目标信息。
PCT/CN2020/120742 2020-05-29 2020-10-14 基于Ghost卷积特征融合神经网络实时车流量检测系统及方法 WO2021238019A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010475998.6 2020-05-29
CN202010475998.6A CN111640101B (zh) 2020-05-29 2020-05-29 基于Ghost卷积特征融合神经网络实时车流量检测系统及方法

Publications (1)

Publication Number Publication Date
WO2021238019A1 true WO2021238019A1 (zh) 2021-12-02

Family

ID=72331616

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/120742 WO2021238019A1 (zh) 2020-05-29 2020-10-14 基于Ghost卷积特征融合神经网络实时车流量检测系统及方法

Country Status (2)

Country Link
CN (1) CN111640101B (zh)
WO (1) WO2021238019A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114419330A (zh) * 2022-01-24 2022-04-29 西北大学 一种二维深度假图泛化检测方法、系统、设备及存储介质
CN114782863A (zh) * 2022-04-07 2022-07-22 中国科学院宁波材料技术与工程研究所 基于i-p帧特征融合的视频目标检测方法
CN115018788A (zh) * 2022-06-02 2022-09-06 常州晋陵电力实业有限公司 基于智能机器人的架空线异常检测方法和系统
CN115134168A (zh) * 2022-08-29 2022-09-30 成都盛思睿信息技术有限公司 基于卷积神经网络的云平台隐蔽通道检测方法及系统
CN116187398A (zh) * 2022-05-16 2023-05-30 山东巍然智能科技有限公司 一种无人机海洋图像检测用轻量神经网络搭建方法及设备
CN116524474A (zh) * 2023-07-04 2023-08-01 武汉大学 一种基于人工智能的车辆目标检测方法及系统
TWI843285B (zh) 2022-11-11 2024-05-21 國立澎湖科技大學 基於深度學習之密度估計網路架構的蝦苗計數系統與方法

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111640101B (zh) * 2020-05-29 2022-04-29 苏州大学 基于Ghost卷积特征融合神经网络实时车流量检测系统及方法
CN112750150B (zh) * 2021-01-18 2023-04-07 西安电子科技大学 基于车辆检测和多目标跟踪的车流量统计方法
CN112819771A (zh) * 2021-01-27 2021-05-18 东北林业大学 一种基于改进YOLOv3模型的木材缺陷检测方法
CN113052254B (zh) * 2021-04-06 2022-10-04 安徽理工大学 多重注意力幽灵残差融合分类模型及其分类方法
CN114387470B (zh) * 2022-01-11 2024-05-31 中船（浙江）海洋科技有限公司 一种基于轻量化卷积神经网络的船舶分类方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106650913A (zh) * 2016-12-31 2017-05-10 中国科学技术大学 一种基于深度卷积神经网络的车流密度估计方法
CN109858461A (zh) * 2019-02-21 2019-06-07 苏州大学 一种密集人群计数的方法、装置、设备以及存储介质
US20200053857A1 (en) * 2013-03-18 2020-02-13 Signify Holding B.V. Methods and apparatus for information management and control of outdoor lighting networks
CN111640101A (zh) * 2020-05-29 2020-09-08 苏州大学 基于Ghost卷积特征融合神经网络实时车流量检测系统及方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110651310B (zh) * 2017-04-05 2023-09-08 卡内基梅隆大学 估计对象密度和/或流量的深度学习方法及相关方法和软件
CN108876774A (zh) * 2018-06-07 2018-11-23 浙江大学 一种基于卷积神经网络的人群计数方法
CN109241858A (zh) * 2018-08-13 2019-01-18 湖南信达通信息技术有限公司 一种基于轨道交通列车的客流密度检测方法及装置
CN110659718B (zh) * 2019-09-12 2021-06-18 中南大学 基于深度卷积神经网络的小卷积核细胞计数方法及系统
CN110879982B (zh) * 2019-11-15 2023-06-20 苏州大学 一种人群计数系统及方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200053857A1 (en) * 2013-03-18 2020-02-13 Signify Holding B.V. Methods and apparatus for information management and control of outdoor lighting networks
CN106650913A (zh) * 2016-12-31 2017-05-10 中国科学技术大学 一种基于深度卷积神经网络的车流密度估计方法
CN109858461A (zh) * 2019-02-21 2019-06-07 苏州大学 一种密集人群计数的方法、装置、设备以及存储介质
CN111640101A (zh) * 2020-05-29 2020-09-08 苏州大学 基于Ghost卷积特征融合神经网络实时车流量检测系统及方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
HAN KAI; WANG YUNHE; TIAN QI; GUO JIANYUAN; XU CHUNJING; XU CHANG: "GhostNet: More Features from Cheap Operations.", 2020 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), 13 June 2020 (2020-06-13), pages 1577 - 1586, XP033805516, DOI: 10.1109/CVPR42600.2020.00165 *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114419330A (zh) * 2022-01-24 2022-04-29 西北大学 一种二维深度假图泛化检测方法、系统、设备及存储介质
CN114419330B (zh) * 2022-01-24 2024-02-09 西北大学 一种二维深度假图泛化检测方法、系统、设备及存储介质
CN114782863A (zh) * 2022-04-07 2022-07-22 中国科学院宁波材料技术与工程研究所 基于i-p帧特征融合的视频目标检测方法
CN114782863B (zh) * 2022-04-07 2023-12-19 中国科学院宁波材料技术与工程研究所 基于i-p帧特征融合的视频目标检测方法
CN116187398A (zh) * 2022-05-16 2023-05-30 山东巍然智能科技有限公司 一种无人机海洋图像检测用轻量神经网络搭建方法及设备
CN116187398B (zh) * 2022-05-16 2023-08-25 山东巍然智能科技有限公司 一种无人机海洋图像检测用轻量神经网络搭建方法及设备
CN115018788A (zh) * 2022-06-02 2022-09-06 常州晋陵电力实业有限公司 基于智能机器人的架空线异常检测方法和系统
CN115018788B (zh) * 2022-06-02 2023-11-14 常州晋陵电力实业有限公司 基于智能机器人的架空线异常检测方法和系统
CN115134168A (zh) * 2022-08-29 2022-09-30 成都盛思睿信息技术有限公司 基于卷积神经网络的云平台隐蔽通道检测方法及系统
TWI843285B (zh) 2022-11-11 2024-05-21 國立澎湖科技大學 基於深度學習之密度估計網路架構的蝦苗計數系統與方法
CN116524474A (zh) * 2023-07-04 2023-08-01 武汉大学 一种基于人工智能的车辆目标检测方法及系统
CN116524474B (zh) * 2023-07-04 2023-09-15 武汉大学 一种基于人工智能的车辆目标检测方法及系统

Also Published As

Publication number Publication date
CN111640101A (zh) 2020-09-08
CN111640101B (zh) 2022-04-29

Similar Documents

Publication Publication Date Title
WO2021238019A1 (zh) 基于Ghost卷积特征融合神经网络实时车流量检测系统及方法
Zhao et al. Detection method based on automatic visual shape clustering for pin-missing defect in transmission lines
CN107545263B (zh) 一种物体检测方法及装置
Chen et al. Moving vehicle detection based on optical flow estimation of edge
CN109886159B (zh) 一种非限定条件下的人脸检测方法
CN107301376B (zh) 一种基于深度学习多层刺激的行人检测方法
Liu et al. Pedestrian detection algorithm based on improved SSD
CN110991444A (zh) 面向复杂场景的车牌识别方法及装置
Kuang et al. Real-time pedestrian detection using convolutional neural networks
CN106650647A (zh) 基于传统算法和深度学习算法级联的车辆检测方法及系统
CN111915583A (zh) 复杂场景中基于车载红外热像仪的车辆和行人检测方法
Liu et al. D-CenterNet: An anchor-free detector with knowledge distillation for industrial defect detection
Zheng et al. Feature enhancement for multi-scale object detection
Yin Object Detection Based on Deep Learning: A Brief Review
CN112766353A (zh) 一种加强局部注意的双分支车辆再识别方法
Xiang et al. A real-time vehicle traffic light detection algorithm based on modified YOLOv3
Wu et al. Research on asphalt pavement disease detection based on improved YOLOv5s
CN116486166A (zh) 一种基于边缘计算的输电线异物识别检测方法
CN112070048B (zh) 基于RDSNet的车辆属性识别方法
Yang et al. Improved Object Detection Algorithm Based on Multi-scale and Variability Convolutional Neural Networks
Zhang et al. PLOT: a 3D point cloud object detection network for autonomous driving
Zhang et al. Channel compression optimization oriented bus passenger object detection
Wang et al. Research on Road Object Detection Model Based on YOLOv4 of Autonomous Vehicle
CN108830166B (zh) 一种公交车客流量实时统计方法
Bhuptani et al. Automating traffic signals based on traffic density estimation in bangalore using YOLO

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: 20937938

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: 20937938

Country of ref document: EP

Kind code of ref document: A1