WO2016155371A1 - 识别交通标志的方法和装置 - Google Patents

识别交通标志的方法和装置 Download PDF

Info

Publication number
WO2016155371A1
WO2016155371A1 PCT/CN2015/098903 CN2015098903W WO2016155371A1 WO 2016155371 A1 WO2016155371 A1 WO 2016155371A1 CN 2015098903 W CN2015098903 W CN 2015098903W WO 2016155371 A1 WO2016155371 A1 WO 2016155371A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
traffic sign
window image
neural network
trained
Prior art date
Application number
PCT/CN2015/098903
Other languages
English (en)
French (fr)
Inventor
郝志会
王宁
葛雷鸣
Original Assignee
百度在线网络技术（北京）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 百度在线网络技术（北京）有限公司 filed Critical 百度在线网络技术（北京）有限公司
Priority to JP2016559945A priority Critical patent/JP6400117B2/ja
Priority to KR1020167027213A priority patent/KR101856584B1/ko
Publication of WO2016155371A1 publication Critical patent/WO2016155371A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
    • G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
    • G06V20/582—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of traffic signs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
    • G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based 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
    • G06T2207/20—Special algorithmic details
    • G06T2207/20021—Dividing image into blocks, subimages or windows
    • 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]

Definitions

  • the present application relates to the field of computer technologies, and in particular, to the field of computer object recognition technologies, and in particular, to a method and apparatus for identifying a traffic sign.
  • the traditional method of identifying traffic signs relies mainly on manual processing.
  • the specific program software and basemap data are deployed to the collection vehicle in advance, and the trained field personnel work with the vehicle: after observing the traffic signage, the operator will The type and visual distance are manually entered into the software; after the field collection is completed, the internal staff compares the base map before and after the operation to update the valid information to the road network database.
  • This process requires the support of a variety of software tools, but the accuracy of the data results mainly depends on the individual ability and concentration of the operator, while the complex acquisition process reduces the update efficiency of the road network data.
  • the purpose of the present application is to propose a method and apparatus for identifying a traffic sign to solve the technical problems mentioned in the background section above.
  • the present application provides a method for identifying a traffic sign, the method comprising: acquiring a feature value obtained by a predetermined feature channel by a predetermined feature algorithm by a scan window image obtained by dividing a panoramic spherical image; according to the scan window image The eigenvalue and the pre-trained detection classifier model, detecting the scanned window image, and obtaining the traffic sign window to be confirmed An image, wherein the pre-trained detection classifier model is trained according to samples of the scan window image and its eigenvalues; according to the pre-trained convolutional neural network model, the traffic sign window image to be confirmed is identified, and the traffic sign category is obtained, wherein The pre-trained convolutional neural network model is trained based on the samples of the confirmed traffic sign window image and its traffic sign categories.
  • the present application provides an apparatus for identifying a traffic sign, the apparatus comprising: a feature value acquiring module, configured to acquire a feature of a scan window image obtained by dividing a panoramic spherical image by a predetermined feature algorithm in a predetermined integration channel; a traffic sign detecting module, configured to detect a scan window image according to a feature value of the scan window image and a pre-trained test classifier model, to obtain a to-be-confirmed traffic sign window image, wherein the pre-trained detection classifier The model is trained according to the sample of the scan window image and its characteristic value; the traffic sign recognition module is configured to identify the image of the traffic sign window to be confirmed according to the pre-trained convolutional neural network model, and obtain a traffic sign category, wherein the pre-training The convolutional neural network model is trained based on the samples of the confirmed traffic sign window image and its traffic sign categories.
  • the method and apparatus for identifying a traffic sign obtained by the present application obtains a feature value obtained by a predetermined feature algorithm by a predetermined feature algorithm by acquiring a scan window image obtained by dividing a panoramic spherical image, and then according to the feature value of the scan window image and pre-training
  • the detection classifier model detects the scanning window image, obtains the image of the traffic sign window to be confirmed, and then identifies the image of the traffic sign window to be confirmed according to the pre-trained convolutional neural network model, obtains the traffic sign category, and realizes the detection through the pre-training.
  • the classifier model detects the image of the traffic sign window to be confirmed, identifies the traffic sign category through the pre-trained convolutional neural network model, improves the accuracy of detecting and identifying the traffic sign in the panoramic image, and improves the update efficiency of the road network data. .
  • FIG. 1 illustrates an exemplary flow chart of a method of identifying a traffic sign in accordance with an embodiment of the present application
  • FIG. 2 illustrates an exemplary flow chart of a method of training a pre-trained detection model in accordance with an embodiment of the present application
  • FIG. 3 illustrates an exemplary flowchart of a method of acquiring an image of a traffic sign window to be confirmed according to an embodiment of the present application
  • FIG. 4 illustrates an exemplary flow chart of a method of training a pre-trained convolutional neural network model in accordance with an embodiment of the present application
  • FIG. 5 illustrates an exemplary structural diagram of a preset convolutional neural network model in accordance with an embodiment of the present application
  • FIG. 6 is a diagram showing an example of the structure of an apparatus for identifying a traffic sign according to an embodiment of the present application
  • FIG. 7 shows an exemplary structural diagram of a detection classifier model training device according to an embodiment of the present application.
  • FIG. 8 shows an exemplary structural diagram of an apparatus for acquiring an image of a traffic sign window to be confirmed
  • FIG. 9 illustrates an exemplary structural diagram of an apparatus for training a pre-trained convolutional neural network model in accordance with an embodiment of the present application
  • FIG. 10 is a schematic structural diagram of a computer system according to an embodiment of the present application.
  • FIG. 1 illustrates an exemplary flow chart of a method of identifying a traffic sign in accordance with an embodiment of the present application.
  • a method 100 of identifying a traffic sign can include:
  • Step 110 Acquire a feature value obtained by the predetermined feature channel by a predetermined feature algorithm of the scan window image obtained by dividing the panoramic spherical image.
  • the panoramic spherical image that needs to be identified by the traffic sign can be first obtained, and then the panoramic spherical image is scanned by the window of the preset size to obtain the scanned window image, and then the scanned window image is acquired through the predetermined feature algorithm in the predetermined integration channel. The resulting feature value.
  • the panoramic spherical image refers to an image that can be displayed by combining a plurality of fisheye images.
  • a window of a preset size is usually a window that is preset according to the size of the target object.
  • the predetermined integration channel may be any one or more integration channels in the prior art in the field of image processing. Any one or more of the integration channels that may be developed in the future technology; the predetermined feature algorithm may be any one or more of the feature algorithms in the prior art in the field of image processing, or may be any possible development in the future technology. One or more feature algorithms.
  • the application does not limit the specific manner of obtaining the feature value, and the user can select according to the actual use requirement.
  • a haar eigenvalue can be obtained by a Hal algorithm in a gray image channel
  • a histogram eigenvalue can be obtained by a histogram eigenvalue algorithm at a gradient image channel of different angle parameters, while being in red, green, and blue monochrome.
  • the image channel obtains a random pair of feature values and the like by randomly eigenvalue algorithm.
  • the feature value of the scan window image may be acquired by the integral map: optionally, in the above embodiment, step 110 may include: step 111, acquiring the image by the panoramic spherical image at a predetermined time The integral map of the integration channel divides the integrated map of the scanned window image; and in step 112, the feature value of the scanned window image is obtained according to the integral map of the scanned window image.
  • the feature value of the scan window image can be accelerated, and the calculation efficiency of the feature value of the scan window image can be improved.
  • the sample of the traffic sign window image may be determined according to the traffic sign category.
  • the feature value and a predetermined integration channel and a predetermined feature algorithm for acquiring the feature values of the scanned window image For example, for the ban class Marks such as speed limit can enhance the description on the red channel.
  • the predetermined integration channel may include one or more of the following: a grayscale image channel, a red, green, and blue monochrome image channel, a gradient image channel of different angle parameters, and an edge detection image channel; and a predetermined feature algorithm may It includes one or more of the following: Haar feature algorithm, local binary pattern feature algorithm, histogram feature algorithm and random pair feature algorithm.
  • a predetermined feature algorithm may It includes one or more of the following: Haar feature algorithm, local binary pattern feature algorithm, histogram feature algorithm and random pair feature algorithm.
  • the predetermined integral channel and the predetermined feature algorithm for acquiring the feature value of the sample of the traffic sign window image and the feature value of the scan window image determined according to the traffic sign category can describe the characteristics of the target from a plurality of different angles, thereby overcoming the target because Changes from different angles and lighting.
  • Step 120 Detecting a scan window image according to the feature value of the scan window image and the pre-trained detection classifier model, and obtaining a traffic sign window image to be confirmed, wherein the pre-trained detection classifier model is based on the sample of the scan window image and its feature Value training is obtained.
  • the pre-trained detection classifier model first receives the manual annotation, and determines a scan window image containing the traffic sign and a scan window image not including the traffic sign as the sample of the traffic sign window image in the scan window image, and then acquires the sample.
  • the eigenvalue is then used to train the parameters of the classifier model set according to the actual demand using the sample and its eigenvalues, thereby obtaining a pre-trained detection classifier model.
  • the scan window image may be detected according to the feature value of the scan window image and the pre-trained detection classifier model to obtain a window in which the traffic sign is suspected.
  • the image serves as a window image of the traffic sign to be confirmed.
  • Step S130 identifying a traffic sign window image to be confirmed according to the pre-trained convolutional neural network model, and obtaining a traffic sign category, wherein the pre-trained convolutional neural network model is based on the confirmed traffic sign window image sample and its traffic sign Category training is available.
  • the pre-trained convolutional neural network model first receives the manual annotation, and divides the sample of the to-be-confirmed traffic sign window image detected by the detection classifier model. Class, marked as a specific traffic sign category or non-traffic category, and then use the sample of the traffic sign window image to be confirmed and its traffic sign category to train the convolutional neural network model set according to actual needs, thus obtaining pre-training Convolutional neural network model.
  • the traffic sign window image to be confirmed may be identified according to the to-be-confirmed traffic sign window image and the pre-trained convolutional neural network model to obtain the traffic to be confirmed.
  • the traffic sign category of the maximum probability corresponding to the flag window image is used as the identified traffic sign category.
  • the method for identifying a traffic sign in the above embodiment of the present application improves the accuracy of detecting and identifying a traffic sign in a panoramic image, and improves the update efficiency of the road network data.
  • FIG. 2 illustrates an exemplary flow chart of a method of training a pre-trained detection model in accordance with an embodiment of the present application.
  • the method 200 of training a pre-trained detection classifier model can include:
  • a positive sample and a negative sample in a sample of the scan window image are acquired, wherein the positive sample includes a window image of the traffic sign or a window image including the traffic sign and the surrounding extended preset pixels, and the negative sample includes the positive sample removed. Scan the window image.
  • the positive sample and the negative sample in the sample of the scan window image are obtained by first acquiring the sample in the scan window image, and then including the window image of the traffic sign in the sample of the scan window image or including the traffic according to the received manual annotation.
  • the window image of the marker and its surrounding preset pixels is used as a positive sample, and the scan window image in the sample from which the positive sample is removed is taken as a negative sample.
  • the above positive sample can improve the efficiency of calculating the feature value of the positive sample when only the window image of the traffic sign is included. Considering that the pixel around the target object can also effectively describe the target object itself, the positive sample can improve the accuracy of describing the feature value of the positive sample when the window image of the preset pixel is expanded including the traffic sign and the surrounding.
  • step 202 the feature values obtained by the predetermined feature and the negative sample in the predetermined integration channel by the predetermined feature algorithm are obtained.
  • the predetermined integration channel when acquiring the feature values obtained by the predetermined feature algorithm in the predetermined integration channel by the positive sample and the negative sample, the predetermined integration channel may be in the prior art in the image processing field. Any one or more of the integration channels may also be any one or more integration channels that may be developed in the future technology; the predetermined feature algorithm may be any one or more of the prior art algorithms in the image processing field. It can also be any one or more feature algorithms that may be developed in future technologies.
  • the present application does not limit the specific pre-integration channel and the predetermined feature algorithm for acquiring feature values, which can be selected by the user according to actual use requirements.
  • the haar eigenvalues obtained by the Haar feature algorithm in the gray image channel the histogram eigenvalues obtained by the gradient image channel histogram feature algorithm at different angle parameters, in the red, green, and blue monochrome images
  • the channel passes the random pair of feature values obtained by the feature algorithm randomly.
  • the traffic sign window may be determined according to the traffic sign category.
  • a predetermined integration channel and a predetermined feature algorithm for the samples of the image For example, for ban-like signs such as speed limit, the description on the red channel can be enhanced.
  • the predetermined integration channel may include one or more of the following: a grayscale image channel, a red, green, and blue monochrome image channel, a gradient image channel of different angle parameters, and an edge detection image channel; and the predetermined feature algorithm may include One or more of the following: a Hal feature algorithm, a local binary pattern feature algorithm, a histogram feature algorithm, and a random pair feature algorithm. It should be understood by those skilled in the art that the predetermined integration channel and the predetermined feature algorithm in the above embodiments may be combined as needed to obtain the feature values obtained by the predetermined feature and the negative sample in the predetermined integration channel by the predetermined feature algorithm.
  • step 203 the classifier model is trained by the boosting algorithm according to the positive sample and the negative sample and the feature values obtained by the predetermined feature algorithm in the predetermined integration channel, and the pre-trained detection classifier model is obtained.
  • the classifier model can be trained by the boosting algorithm to obtain a high-accuracy result classifier as the pre-training. Detection classifier model.
  • FIG. 3 illustrates an exemplary flow chart of a method of acquiring a to-be-confirmed traffic sign window image in accordance with an embodiment of the present application.
  • the method 300 for obtaining an image of a traffic sign window to be confirmed includes:
  • step 301 the panoramic spherical image is continuously downsampled to obtain an image pyramid.
  • the panoramic spherical image may be subjected to pyramid transformation, such as Laplacian pyramid transformation or the like, to transform the panoramic spherical image from the original size to a different scale space, thereby obtaining an image pyramid including the multi-level image.
  • pyramid transformation such as Laplacian pyramid transformation or the like
  • Step 302 Acquire an integral graph of the scan window image of each level image in the image pyramid obtained by dividing the image pyramid in the integral map of the predetermined integration channel.
  • each level image of the image pyramid may be first integrated in a predetermined integration channel to obtain an integral map of the image pyramid, and then the integral map of the image pyramid is divided by a window of a preset size to obtain An integral map of the scanned window image for each level of image.
  • Step 303 Acquire feature values of the scan window image of each level image according to the integral map of the scan window image of each level image.
  • the integral map of the scan window image of each level image may be used to perform the operation according to the integral map of the scan window image of each level image, and the image of each level is quickly obtained. Scans the feature values of the window image.
  • Step 304 Detect a scan window image of each level image according to a feature value of the scan window image of each level image and a pre-trained detection classifier model, and obtain a traffic sign window image to be confirmed.
  • the scan window image of each level image may be detected by the pre-trained detection classifier model according to the feature value of the scan window image of each level image, and the obtained The suspected traffic sign window image is used as the to-be-confirmed traffic sign window image, thereby obtaining all the to-be-confirmed traffic sign window images in the image pyramid.
  • the method for obtaining a to-be-confirmed traffic sign window image in the above embodiment of the present application acquires a scan window image in a pyramid image acquired by a pyramidal image by a pyramid transformation, and then detects a scan window image of each level image to obtain a to-be-confirmed traffic sign window image.
  • the possibility of missing the traffic sign window to be confirmed in the panoramic spherical image is reduced, thereby improving the accuracy of obtaining the traffic sign window to be confirmed.
  • the integral value is used to acquire the feature value of the scan window image of each level image, the acquisition speed of the feature value is improved.
  • FIG. 4 illustrates an exemplary flow chart of a method of training a pre-trained convolutional neural network model in accordance with an embodiment of the present application.
  • a method 400 of training a pre-trained convolutional neural network model includes:
  • Step 401 Initialize weights of the convolutional layer and the fully connected layer of the preset convolutional neural network model according to the Gaussian distribution, where the preset convolutional neural network model includes convolution layers, extraction layers, and Fully connected layer and normalized layer.
  • the successively connected convolution layer, extraction layer, fully connected layer, and normalization layer may include: one convolution layer, one extraction layer, one full connection layer, and one normalization layer; or may include multiple volumes The layer and the extraction layer corresponding to the convolution layer, one or more fully connected layers and one normalized layer.
  • the convolution layer set therein can enhance the original signal feature and reduce the noise by convolution operation; the set extraction layer can use the principle of image local correlation to subsample the image while reducing the amount of data processing. Keep useful information.
  • FIG. 5 shows an exemplary structural diagram of a preset convolutional neural network model in accordance with an embodiment of the present application.
  • the preset convolutional neural network model 500 includes: a convolutional layer conv1, an extraction layer pool1, a convolutional layer conv2, an extraction layer pool2, a convolution layer conv3, an extraction layer pool3, and a full connection layer.
  • Fc1 fully connected layer fc2.
  • conv1 has 16 convolution kernels with size 5*5*3
  • conv2 has 32 convolution kernels with size 5*5*16
  • conv3 has 64 convolution kernels with size 5*5*32.
  • the two fully connected layers each have 512 and 120 neurons.
  • the preset convolutional neural network model may further include a loss function layer.
  • the loss function can be calculated using a multi-class logarithmic loss logarithmic loss function, ie
  • L is the loss function
  • N is the number of training samples
  • M is the number of categories
  • p i,j is the probability that the i-th sample of the network output is j-th class
  • y i,j is the true value of the sample, if the ith sample It belongs to the jth class, it is 1, otherwise it is 0.
  • step 402 according to the confirmed sample of the traffic sign window image and its traffic sign category, the weight of both the convolutional layer and the fully connected layer is iterated by the error backpropagation BP algorithm.
  • the sample of the confirmed traffic sign window image is a sample of the window image including the traffic sign confirmed according to the received manual mark in the sample extracted from the image of the traffic sign window to be confirmed.
  • the sample of the confirmed traffic sign window image may include: a sample of the original confirmed traffic sign window image, and normalizing the sample of the original confirmed traffic sign window image to the preset window after one or more of the following processing Dimensional images: rotate, pan, and zoom.
  • iterating through the error back propagation BP algorithm for the weights of both the convolution layer and the fully connected layer may include: through the loss function and the BP algorithm, the convolution layer and the full connection layer
  • the weights of the individuals are iterated separately. For example, after initializing the weight of the Gaussian distribution, the network model is iterated multiple times by using a stochastic gradient descent algorithm. Each time, the loss function L of the network is calculated by the loss function formula (a), and then reversed. The layer calculates the gradient of L relative to each layer weight W i , and then updates the weight W i according to the gradient of each layer weight W i , ie
  • ⁇ is the learning rate of the preset gradient descent, It is the gradient of the loss function L with respect to the weight W i .
  • the convolutional neural network model of the optimal weight may be determined by step S403 or step S404.
  • Step S403 If the difference between the weight after the current iteration and the weight after the last iteration is less than a preset value, the weight after the current iteration is determined as the optimal weight.
  • Step S404 If the weight after the iteration with the lowest error rate occurs, the weight after the iteration with the lowest error rate is determined as the optimal weight.
  • Step S405 The convolutional neural network model including the optimal weight is set as a pre-trained convolutional neural network model.
  • step 103 in FIG. 1 may include: inputting the image of the traffic sign window to be confirmed into the convolutional neural network model to obtain The traffic sign category with the largest weight output from the normalized layer; the traffic sign category with the largest weight is set as the identified traffic sign category.
  • FIG. 6 shows an example of the structure of an apparatus for identifying a traffic sign according to an embodiment of the present application.
  • the apparatus 600 for identifying a traffic sign may include: a feature value acquisition module 610, a traffic sign detection module 620, and a traffic sign recognition module 630.
  • the feature value obtaining module 610 is configured to acquire the feature value obtained by the predetermined feature channel by the predetermined feature channel by the scan window image obtained by the panoramic spherical image segmentation.
  • the feature value obtaining module 610 may include: an integral map acquiring first sub-module 611, configured to acquire an integral map of the scan window image obtained by dividing the panoramic spherical image in the integral map of the predetermined integration channel; and acquiring the feature value A sub-module 612 is configured to acquire a feature value of the scan window image according to the integral map of the scan window image.
  • the traffic sign detecting module 620 is configured to detect a scan window image according to the feature value of the scan window image and the pre-trained test classifier model, and obtain a to-be-confirmed traffic sign window image, wherein the pre-trained detection classifier model is based on the scan window image
  • the samples and their eigenvalues are trained.
  • the traffic sign recognition module 630 is configured to identify a traffic sign window image to be confirmed according to the pre-trained convolutional neural network model, and obtain a traffic sign category, wherein the pre-trained convolutional neural network model is based on the confirmed traffic sign window image.
  • the sample and its traffic sign categories were trained.
  • the device for identifying a traffic sign in the above embodiment of the present application improves the accuracy of detecting and identifying a traffic sign in a panoramic image, and improves the update efficiency of the road network data.
  • the pre-trained detection classifier model described above may be obtained by acquiring a positive sample and a negative sample in a sample of the scan window image, wherein the positive sample includes a window image of the traffic sign or includes a traffic sign and its surrounding expansion a window image of a preset pixel, the negative sample includes a scan window image from which the positive sample is removed; the feature value obtained by the predetermined feature algorithm in the predetermined integration channel by the positive sample and the negative sample; and the positive sample and the negative sample and the The feature value obtained by the predetermined feature algorithm is trained by the boosting algorithm to detect the classifier model, and a pre-trained detection classifier model is obtained.
  • the above method for obtaining a pre-trained detection classifier model can be illustrated by FIG.
  • the detection classifier model training device is implemented.
  • FIG. 7 shows an exemplary structural diagram of a detection classifier model training device according to an embodiment of the present application.
  • the detection classifier model training apparatus 700 may include a sample acquisition module 701, a sample feature value acquisition module 702, and a detection model training module 703.
  • the sample obtaining module 701 is configured to obtain a positive sample and a negative sample in the sample of the scan window image, where the positive sample includes a window image of the traffic sign or a window image including the traffic sign and the surrounding extended preset pixels, and the negative sample includes the positive sample The scanned window image of the sample.
  • the sample feature value obtaining module 702 is configured to obtain the feature values obtained by the predetermined feature channel by the predetermined feature algorithm in the predetermined integration channel.
  • the predetermined integration channel includes one or more of the following: a grayscale image channel, a red, green, and blue monochrome image channel, a gradient image channel with different angle parameters, and an edge detection image channel; and the predetermined feature algorithm includes the following one Or multiple: Hal feature algorithm, local binary pattern feature algorithm, histogram feature algorithm and random pair feature algorithm.
  • the detection model training module 703 is configured to train the detection classifier model by using a boosting algorithm according to the positive sample and the negative sample and the feature values obtained by the predetermined feature algorithm in the predetermined integration channel, to obtain a pre-trained detection classifier model.
  • FIG. 8 shows an exemplary structural diagram of an apparatus for acquiring an image of a traffic sign window to be confirmed.
  • the apparatus 800 for acquiring a to-be-confirmed traffic sign window image may include: a downsampling sub-module 801, an integral map acquisition second sub-module 802, a feature value second acquisition sub-module 803, and a multi-level detection sub-module. 804.
  • the downsampling sub-module 801 is configured to continuously downsample the panoramic spherical image to obtain an image pyramid.
  • the integral map acquires a second sub-module 802 for acquiring an integral map of the scan window image of each level image in the image pyramid obtained by dividing the image pyramid by the integral map of the predetermined integration channel.
  • Feature value second acquisition sub-module 803 for scanning window image according to each level of image The integral map of the image of the scanned window image of each level of image is obtained.
  • the multi-level detection sub-module 804 is configured to detect a scan window image of each level image according to the feature value of the scan window image of each level image and the pre-trained detection classifier model, and obtain a to-be-confirmed traffic sign window image.
  • the device for obtaining the image of the traffic sign window to be confirmed by the foregoing embodiment of the present application can further improve the accuracy of obtaining the image of the traffic sign window to be confirmed and the speed of detecting the image of the traffic sign window to be confirmed.
  • the pre-trained convolutional neural network model in FIG. 6 above may be obtained by: initializing the weights of both the convolutional layer and the fully connected layer of the preset convolutional neural network model according to a Gaussian distribution, wherein The preset convolutional neural network model includes a convolution layer, an extraction layer, a full connection layer, and a normalization layer connected in sequence; according to the confirmed sample of the traffic sign window image and its traffic sign category, the error is reversed
  • the propagation BP algorithm iterates the weights of the convolution layer and the fully connected layer.
  • the current iteration is The weight is determined as the optimal weight, and the convolutional neural network model including the optimal weight is set as the pre-trained convolutional neural network model; or if the weight after the iteration with the lowest error rate occurs, the error rate is the lowest.
  • the iterative weight is determined as the optimal weight, and the convolutional neural network model including the optimal weight is set as the pre-trained convolutional neural network model.
  • the sample of the confirmed traffic sign window image is a sample of the window image including the traffic sign confirmed according to the received manual mark in the sample extracted from the image of the traffic sign window to be confirmed.
  • sample of the confirmed traffic sign window image may include: a sample of the original confirmed traffic sign window image, and normalizing the sample of the original confirmed traffic sign window image to the preset window after one or more of the following processing Dimensional images: rotate, pan, and zoom.
  • the convolution layer, the decimation layer, the full connection layer and the normalization layer sequentially connected include: a plurality of convolution layers and an extraction layer corresponding to the convolution layer, one or more fully connected layers and one normalized layer .
  • the preset convolutional neural network model may further include a loss function layer.
  • the convolutional neural network model corresponding to the preset also includes a loss function layer, which is reversed by error. Iterating to the weighting of the convolutional layer and the fully connected layer by the propagation BP algorithm may include iterating the weights of both the convolutional layer and the fully connected layer by a loss function and a BP algorithm, respectively.
  • the above method of obtaining a pre-trained convolutional neural network model can be implemented by the apparatus for training a pre-trained convolutional neural network model shown in FIG.
  • FIG. 9 illustrates an exemplary structural diagram of an apparatus for training a pre-trained convolutional neural network model in accordance with an embodiment of the present application.
  • an apparatus 900 for training a pre-trained convolutional neural network model includes:
  • the initialization module 901 is configured to initialize weights of the convolutional layer and the fully connected layer of the preset convolutional neural network model according to the Gaussian distribution, where the preset convolutional neural network model includes convolution layers connected in sequence, Extract layer, fully connected layer, and normalized layer.
  • the convolution layer, the decimation layer, the full connection layer and the normalization layer sequentially connected include: a plurality of convolution layers and an extraction layer corresponding to the convolution layer, one or more fully connected layers and one normalized layer .
  • the preset convolutional neural network model further includes a loss function layer.
  • the weight iteration module 902 is configured to iterate the weights of the convolutional layer and the fully connected layer by using an error backpropagation BP algorithm according to the sample of the confirmed traffic sign window image and its traffic sign category;
  • the first optimal weight determining module 903 is configured to determine the weight after the current iteration as the optimal weight if the difference between the weight after the current iteration and the weight after the last iteration is less than a preset value.
  • the second optimal weight determination module 904 is configured to determine, after the iterative weight with the lowest error rate, the weight of the iteration with the lowest error rate as the optimal weight.
  • the model setting module 905 is configured to set the convolutional neural network model including the optimal weight as a pre-trained convolutional neural network model.
  • the convolutional neural network model corresponding to the preset further includes a loss function layer, and the weight determination module 902 is further configured to iterate the weights of both the convolutional layer and the fully connected layer by the loss function and the BP algorithm, respectively.
  • the device 900 for training a pre-trained convolutional neural network model is trained in advance After the trained convolutional neural network model, the traffic sign recognition module 630 in FIG. 6 may include: a maximum weight identification module (not shown) for inputting the image of the traffic sign window to be confirmed into the convolutional neural network model. A traffic sign category having the largest weight of the output of the layer; and a flag category setting module (not shown) for setting the traffic sign category having the largest weight as the identified traffic sign class.
  • the units recited in apparatus 600 correspond to the various steps in the method described with reference to FIG.
  • the units described in apparatus 700 correspond to the various steps in the method described with reference to FIG. 2.
  • the units described in apparatus 800 correspond to the various steps in the method described with reference to FIG.
  • the units described in apparatus 900 correspond to the various steps in the method described with reference to FIG.
  • the operations and features described above for the method of identifying traffic signs are equally applicable to the apparatus 600 and the units contained therein, and the operations and features described above for the method of training the pre-trained detection model are equally applicable to the apparatus 700 and therein.
  • the included elements, the operations and features described above for the method of obtaining the image of the traffic sign window to be confirmed are equally applicable to the apparatus 800 and the units contained therein, the operations described above for the method of training the pre-trained convolutional neural network model and The features are equally applicable to the device 900 and the units contained therein, and are not described herein again.
  • Corresponding units in devices 600, 700, 800, and 900 can cooperate with units in the terminal device and/or server to implement the solution of the embodiments of the present application.
  • the modules involved in the embodiments of the present application may be implemented by software or by hardware.
  • the described modules may also be provided in the processor, for example, as described in one of the following: a processor includes a feature value acquisition module, a detection module, and an identification module.
  • the name of these modules does not constitute a limitation on the module itself in some cases.
  • the feature value acquisition module may also be described as "for acquiring a scan window image obtained by dividing the panoramic spherical image in a predetermined integration channel.
  • a module of feature values obtained by a predetermined feature algorithm may be implemented by software or by hardware.
  • the described modules may also be provided in the processor, for example, as described in one of the following: a processor includes a feature value acquisition module, a detection module, and an identification module.
  • the name of these modules does not constitute a limitation on the module itself in some cases.
  • the feature value acquisition module may also be described as "for acquiring a scan window image obtained by dividing the panoramic spherical image in a pre
  • FIG. 10 is a schematic structural diagram of a computer system according to an embodiment of the present application.
  • FIG. 10 there is shown a block diagram of a computer system 700 suitable for use in implementing the apparatus of the embodiments of the present application.
  • computer system 1000 includes a central processing unit (CPU) 1001 that can be based on a program stored in read only memory (ROM) 1002 or from a storage portion. 708 loads the program into random access memory (RAM) 1003 to perform various appropriate actions and processes.
  • CPU central processing unit
  • RAM random access memory
  • the CPU 1001 executes the method described in the present application.
  • various programs and data required for the operation of the system 1000 are also stored.
  • the CPU 1001, the ROM 1002, and the RAM 1003 are connected to each other through a bus 1004.
  • An input/output (I/O) interface 1005 is also coupled to bus 1004.
  • the following components are connected to the I/O interface 1005: an input portion 1006 including a keyboard, a mouse, etc.; an output portion 1007 including a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a speaker; a storage portion 1008 including a hard disk or the like And a communication portion 1009 including a network interface card such as a LAN card, a modem, or the like.
  • the communication section 1009 performs communication processing via a network such as the Internet.
  • Driver 1010 is also coupled to I/O interface 1005 as needed.
  • a removable medium 1011 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory or the like is mounted on the drive 1010 as needed so that a computer program read therefrom is installed into the storage portion 1008 as needed.
  • embodiments of the present disclosure include a computer program product comprising a computer program tangibly embodied on a machine readable medium, the computer program comprising program code for executing the method illustrated in the flowchart.
  • the computer program can be downloaded and installed from the network via the communication portion 1009, and/or installed from the removable medium 1011.
  • each block in the flowchart or block diagram can represent a module, a program segment, or a portion of code, and a module, a program segment, or a portion of code includes one or more Executable instructions.
  • the functions noted in the blocks may also occur in a different order than that illustrated in the drawings. For example, two successively represented blocks may in fact be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams and/or flowcharts, and combinations of blocks in the block diagrams and/or flowcharts can be implemented in a dedicated hardware-based system that performs the specified function or operation. Or you can use dedicated hardware Implemented in combination with computer instructions.
  • the present application further provides a computer readable storage medium, which may be a computer readable storage medium included in the apparatus described in the foregoing embodiment, or may exist separately, not A computer readable storage medium that is assembled into a terminal.
  • the computer readable storage medium stores one or more programs that are used by one or more processors to perform the method of identifying a traffic sign as described herein.

Abstract

一种识别交通标志的方法和装置。所述方法的一具体实施方式包括：获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值(110)；根据所述扫描窗口图像的特征值和预先训练的检测分类器模型，检测扫描窗口图像，得到待确认交通标志窗口图像，其中，所述预先训练的检测分类器模型根据扫描窗口图像的样本及其特征值训练得到(120)；根据预先训练的卷积神经网络模型，识别待确认交通标志窗口图像，得到交通标志类别，其中，所述预先训练的卷积神经网络模型根据已确认的交通标志窗口图像的样本及其交通标志类别训练得到(130)。该实施方式提高了在全景图像中检测和识别交通标志的准确性，同时提高了路网数据的更新效率。

Description

识别交通标志的方法和装置
相关申请的交叉引用
本申请要求于2015年03月31日提交的中国专利申请号为“201510150525.8”的优先权，其全部内容作为整体并入本申请中。
技术领域
本申请涉及计算机技术领域，具体涉及计算机目标识别技术领域，尤其涉及识别交通标志的方法和装置。
背景技术
为了提供准确而全面的导航数据，需要对交通路网中的交通标志进行识别。
传统的识别交通标志的方法，主要依靠人工处理，预先将特定的程序软件和底图数据部署到采集车上，由经过培训的外业人员随车作业：观察到交通标志标牌以后，作业员将类型和目测的距离手动录入到软件内；外业采集结束后，再由内业人员比较作业前后的底图，将有效的信息更新到路网数据库。这个过程需要多种软件工具的支持，但数据成果的准确性主要取决于作业人员的个人能力和专注度，同时，复杂的采集流程降低了路网数据的更新效率。
发明内容
本申请的目的在于提出一种识别交通标志的方法和装置，来解决以上背景技术部分提到的技术问题。
一方面，本申请提供了一种识别交通标志的方法，所述方法包括：获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值；根据所述扫描窗口图像的特征值和预先训练的检测分类器模型，检测扫描窗口图像，得到待确认交通标志窗口 图像，其中，所述预先训练的检测分类器模型根据扫描窗口图像的样本及其特征值训练得到；根据预先训练的卷积神经网络模型，识别待确认交通标志窗口图像，得到交通标志类别，其中，所述预先训练的卷积神经网络模型根据已确认的交通标志窗口图像的样本及其交通标志类别训练得到。
第二方面，本申请提供了一种识别交通标志的装置，所述装置包括：特征值获取模块，用于获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值；交通标志检测模块，用于根据所述扫描窗口图像的特征值和预先训练的检测分类器模型，检测扫描窗口图像，得到待确认交通标志窗口图像，其中，所述预先训练的检测分类器模型根据扫描窗口图像的样本及其特征值训练得到；交通标志识别模块，用于根据预先训练的卷积神经网络模型，识别待确认交通标志窗口图像，得到交通标志类别，其中，所述预先训练的卷积神经网络模型根据已确认的交通标志窗口图像的样本及其交通标志类别训练得到。
本申请提供的识别交通标志的方法和装置，通过获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值，随后根据所述扫描窗口图像的特征值和预先训练的检测分类器模型，检测扫描窗口图像，得到待确认交通标志窗口图像，而后根据预先训练的卷积神经网络模型，识别待确认交通标志窗口图像，得到交通标志类别，实现了通过预先训练的检测分类器模型检测待确认交通标志窗口图像，通过预先训练的卷积神经网络模型识别出交通标志类别，提高了在全景图像中检测和识别交通标志的准确性，同时提高了路网数据的更新效率。
附图说明
通过阅读参照以下附图所作的对非限制性实施例所作的详细描述，本申请的其它特征、目的和优点将会变得更明显：
图1示出了根据本申请实施例的识别交通标志的方法的示例性流程图；
图2示出了根据本申请实施例的训练预先训练的检测模型的方法的示例性流程图；
图3示出了根据本申请实施例的获取待确认交通标志窗口图像的方法的一种示例性流程图；
图4示出了根据本申请实施例的训练预先训练的卷积神经网络模型的方法的示例性流程图；
图5示出了根据本申请实施例的预设的卷积神经网络模型的一种示例性结构图；
图6示出了根据本申请实施例的识别交通标志的装置的结构示例图；
图7示出了根据本申请实施例的检测分类器模型训练装置的示例性结构图；
图8示出了用于获取待确认交通标志窗口图像的装置的示例性结构图；
图9示出了根据本申请实施例的用于训练预先训练的卷积神经网络模型的装置的示例性结构图；
图10示出了根据本申请实施例提供的一种计算机系统的结构示意图。
具体实施方式
下面结合附图和实施例对本申请作进一步的详细说明。可以理解的是，此处所描述的具体实施例仅仅用于解释相关发明，而非对该发明的限定。另外还需要说明的是，为了便于描述，附图中仅示出了与有关发明相关的部分。
需要说明的是，在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互组合。下面将参考附图并结合实施例来详细说明本申请。
图1示出了根据本申请实施例的识别交通标志的方法的示例性流程图。
如图1所示，识别交通标志的方法100可以包括：
步骤110，获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值。
在这里，可以首先获取需要进行交通标志识别的全景球面图像，然后通过预设尺寸的窗口对全景球面图像进行扫描，从而得到扫描窗口图像，之后再获取扫描窗口图像在预定积分通道通过预定特征算法的得到的特征值。
其中，全景球面图像是指由多张鱼眼图像拼接而成的可以显示全景的图像。预设尺寸的窗口通常为根据目标物体的尺寸预设的窗口。
需要说明的是，在获取扫描窗口图像在预定的积分通道通过预定的特征算法得到的特征值时，预定的积分通道可以为图像处理领域现有技术中的任一种或多种积分通道，也可以为或未来技术中可能发展的任一种或多种积分通道；预定的特征算法可以为图像处理领域现有技术中任一种或多种特征算法，也可以为未来技术中可能发展的任一种或多种特征算法。本申请对具体的获取特征值的方式并不做限制，其可以由用户根据实际使用需求进行选择。例如，可以在灰度图像通道通过哈尔算法获取哈尔(haar)特征值，在不同角度参数的梯度图像通道通过直方图特征值算法获取直方图特征值，同时在红、绿、蓝单色图像通道通过随机对特征值算法获取随机对特征值等。
为了进一步提高获取扫描窗口图像的特征值的速度，可以通过积分图获取扫描窗口图像的特征值：可选地，在上述实施例中，步骤110可以包括：步骤111，获取由全景球面图像在预定积分通道的积分图划分得到的扫描窗口图像的积分图；以及步骤112，根据扫描窗口图像的积分图，获取扫描窗口图像的特征值。
通过上述使用全景球面图像的积分图计算扫描窗口图像的特征值，可以加速计算扫描窗口图像的特征值，提高扫描窗口图像的特征值的计算效率。
可选地，为了加速扫描窗口图像的特征值与预先训练的监测分类器模型中的交通标志窗口图像的样本的特征值的比对，可以根据交通标志类别，来确定获取交通标志窗口图像的样本的特征值和获取扫描窗口图像的特征值的预定积分通道和预定特征算法。例如针对禁令类 标志如限速等，可以加强在红色通道上的描述。优选地，预定积分通道可以包括以下一项或多项：灰度图像通道，红、绿、蓝单色图像通道，不同角度参数的梯度图像通道，和边缘检测图像通道等；以及预定特征算法可以包括以下一项或多项：哈尔特征算法、局部二值模式特征算法、直方图特征算法和随机对特征算法等。本领域技术人员应当理解，上述实施例中的预定积分通道和预定特征算法可以根据需要进行组合，以得到进行对扫描窗口图像进行检测所需要的特征值。
上述根据交通标志类别确定的获取交通标志窗口图像的样本的特征值和获取扫描窗口图像的特征值的预定积分通道和预定特征算法，能够从多个不同的角度描述目标的特性，从而克服目标因为不同角度和光照带来的变化。
步骤120，根据扫描窗口图像的特征值和预先训练的检测分类器模型，检测扫描窗口图像，得到待确认交通标志窗口图像，其中，预先训练的检测分类器模型根据扫描窗口图像的样本及其特征值训练得到。
在这里，预先训练的检测分类器模型，是首先接收人工标注，在扫描窗口图像中确定包含交通标志的扫描窗口图像和未包含交通标志的扫描窗口图像作为交通标志窗口图像的样本，之后获取样本的特征值，然后使用样本及其特征值，对根据实际需求设定的分类器模型的参数进行训练，从而得到的预先训练的检测分类器模型。
当在上述步骤101中获取到扫描窗口图像的特征值之后，就可以根据扫描窗口图像的特征值和上述预先训练的检测分类器模型，对扫描窗口图像进行检测，以得到其中疑似交通标志的窗口图像作为待确认交通标志窗口图像。
步骤S130，根据预先训练的卷积神经网络模型，识别待确认交通标志窗口图像，得到交通标志类别，其中，预先训练的卷积神经网络模型根据已确认的交通标志窗口图像的样本及其交通标志类别训练得到。
在这里，预先训练的卷积神经网络模型，是首先接收人工标注，将检测分类器模型检测出来的待确认交通标志窗口图像的样本进行分 类，标注为具体的交通标志类别或非交通类别，之后使用待确认交通标志窗口图像的样本及其交通标志类别，对根据实际需要设定的卷积神经网络模型进行训练，从而得到的预先训练的卷积神经网络模型。
当在上述步骤102中获取到待确认交通标志窗口图像之后，就可以根据待确认交通标志窗口图像和上述预先训练的卷积神经网络模型，对待确认交通标志窗口图像进行识别，以得到待确认交通标志窗口图像对应的最大概率的交通标志类别作为识别出的交通标志类别。
本申请上述实施例的识别交通标志的方法，提高了在全景图像中检测和识别交通标志的准确性，同时提高了路网数据的更新效率。
下面结合图2描述训练预先训练的检测模型的方法。
图2示出了根据本申请实施例的训练预先训练的检测模型的方法的示例性流程图。
如图2所示，训练预先训练的检测分类器模型的方法200可以包括：
在步骤201中，获取扫描窗口图像的样本中的正样本和负样本，其中正样本包括交通标志的窗口图像或包括交通标志及其周围扩充预设像素的窗口图像，负样本包括去除正样本的扫描窗口图像。
在这里，扫描窗口图像的样本中的正样本和负样本，是首先在扫描窗口图像中获取样本，之后按照接收的人工标注，将上述扫描窗口图像的样本中包括交通标志的窗口图像或包括交通标志及其周围扩充预设像素的窗口图像作为正样本，将样本中去除正样本的扫描窗口图像作为负样本。
上述正样本在仅包括交通标志的窗口图像时，可以提高计算正样本的特征值的效率。考虑到目标物体周围像素也能有效的描述目标物体本身，因此正样本在包括交通标志及其周围扩充预设像素的窗口图像时，可以提高描述正样本的特征值的准确性。
在步骤202中，获取正样本和负样本在预定积分通道通过预定特征算法得到的特征值。
在这里，当获取正样本和负样本在预定积分通道通过预定特征算法得到的特征值时，预定的积分通道可以为图像处理领域现有技术中 的任一种或多种积分通道，也可以为或未来技术中可能发展的任一种或多种积分通道；预定的特征算法可以为图像处理领域现有技术中任一种或多种特征算法，也可以为未来技术中可能发展的任一种或多种特征算法。本申请对具体的获取特征值的预定积分通道和预定特征算法不做限制，其可以由用户根据实际使用需求进行选择。例如，在灰度图像通道通过哈尔特征算法获取的哈尔(haar)特征值，在不同角度参数的梯度图像通道直方图特征算法获取的直方图特征值，在红、绿、蓝单色图像通道通过随机对特征算法获取的随机对特征值等。
可选地，在进行样本的特征值的具体计算时，为了快速并有效的得到正样本和负样本在预定积分通道通过预定特征算法得到的特征值，可以根据交通标志类别，来确定交通标志窗口图像的样本的预定积分通道和预定特征算法。例如针对禁令类标志如限速等，可以加强在红色通道上的描述。优选地，预定积分通道可以包括以下一项或多项：灰度图像通道，红、绿、蓝单色图像通道，不同角度参数的梯度图像通道，和边缘检测图像通道；以及预定特征算法可以包括以下一项或多项：哈尔特征算法、局部二值模式特征算法、直方图特征算法和随机对特征算法。本领域技术人员应当理解，上述实施例中的预定积分通道和预定特征算法可以根据需要进行组合，以得到正样本和负样本在预定积分通道通过预定特征算法得到的特征值。
在步骤203中，根据正样本和负样本及其在预定积分通道通过预定特征算法得到的特征值，通过boosting算法训练检测分类器模型，得到预先训练的检测分类器模型。
在这里，在上述步骤202获取正样本和负样本在预定积分通道通过预定特征算法得到的特征值之后，可以通过boosting算法训练检测分类器模型，从而得到一个高准确度的结果分类器作为预先训练的检测分类器模型。
以下结合图3，描述在上述实施例的基础上获取待确认交通标志窗口图像的方法。
图3示出了根据本申请实施例的获取待确认交通标志窗口图像的方法的一种示例性流程图。
如图3所示，获取待确认交通标志窗口图像的方法300包括：
步骤301，对全景球面图像连续降采样，获得图像金字塔。
在这里，可以对全景球面图像进行金字塔变换，例如拉普拉斯金字塔变换等，将全景球面图像从原尺寸变换到不同的尺度空间，从而得到包括多级图像的图像金字塔。
步骤302，获取由图像金字塔在预定积分通道的积分图划分得到的图像金字塔中每级图像的扫描窗口图像的积分图。
在上述步骤301获得图像金字塔之后，可以首先对图像金字塔的每一级图像在预定积分通道进行积分，得到图像金字塔的积分图，之后通过预设尺寸的窗口对图像金字塔的积分图进行划分，得到每级图像的扫描窗口图像的积分图。
步骤303，根据每级图像的扫描窗口图像的积分图，获取每级图像的扫描窗口图像的特征值。
在上述步骤302得到每级图像的扫描窗口图像的积分图之后，可以根据每级图像的扫描窗口图像的积分图，利用每级图像的扫描窗口图像的积分图进行运算，快速获取每级图像的扫描窗口图像的特征值。
步骤304，根据每级图像的扫描窗口图像的特征值和预先训练的检测分类器模型，检测每级图像的扫描窗口图像，得到待确认交通标志窗口图像。
在上述步骤303获取每级图像的扫描窗口图像的特征值之后，可以根据每级图像的扫描窗口图像的特征值，通过预先训练的检测分类器模型检测每级图像的扫描窗口图像，得到其中的疑似交通标志窗口图像作为待确认交通标志窗口图像，从而得到图像金字塔中所有待确认交通标志窗口图像。
本申请上述实施例的获取待确认交通标志窗口图像的方法，在全景球面图像通过金字塔变换获取的金字塔图像中获取扫描窗口图像，之后检测每级图像的扫描窗口图像以得到待确认交通标志窗口图像，减少了在全景球面图像中遗漏待确认交通标志窗口的可能性，因此提高了获取待确认交通标志窗口的准确度。又由于采用了积分图获取每级图像的扫描窗口图像的特征值，提高了特征值的获取速度。
以下结合图4，描述训练预先训练的卷积神经网络模型的方法。
图4示出了根据本申请实施例的训练预先训练的卷积神经网络模型的方法的示例性流程图。
如图4所示，训练预先训练的卷积神经网络模型的方法400包括：
步骤401，按照高斯分布，初始化预设的卷积神经网络模型的卷积层和全连接层两者的权重，其中，预设的卷积神经网络模型包括依次连接的卷积层、抽取层、全连接层和归一化层。
在这里，依次连接的卷积层、抽取层、全连接层和归一化层可以包括：一个卷积层、一个抽取层、一个全连接层及一个归一化层；也可以包括多个卷积层以及与卷积层一一对应的抽取层、一个以上全连接层和一个归一化层。
其中设置的卷积层，可以通过卷积运算，使原信号特征增强，并且降低噪音；设置的抽取层，可以利用图像局部相关性的原理，对图像进行子抽样，在减少数据处理量的同时保留有用信息。
下面以图5为例，描述预设的卷积神经网络模型。
图5示出了根据本申请实施例的预设的卷积神经网络模型的一种示例性结构图。
如图5所示，预设的卷积神经网络模型500包括：依次连接的卷积层conv1、抽取层pool1、卷积层conv2、抽取层pool2、卷积层conv3、抽取层pool3、全接连层fc1、全连接层fc2。
其中，conv1有16个大小为5*5*3的卷积核，conv2有32个大小为5*5*16的卷积核，conv3有64个大小为5*5*32的卷积核。两个全连接层各有512和120个神经元。
可选地，为了增大相似目标与非相似目标的输出差别，预设的卷积神经网络模型还可以包括损失函数层。例如，损失函数可以用多类对数损失logarithmic loss函数计算，即
Figure PCTCN2015098903-appb-000001
其中L为损失函数，N为训练样本数量，M为类别数量，pi，j为网络输出的第i个样本是第j类的概率，yi，j为样本真值，如果第i个样本是属于第j类，则为1，否则为0。
返回图4，步骤402，根据已确认的交通标志窗口图像的样本及其交通标志类别，通过误差反向传播BP算法对卷积层和全连接层两者的权重进行迭代。
其中，已确认的交通标志窗口图像的样本，是在待确认交通标志窗口图像中提取的样本中，根据接收的人工标注确认后的包括交通标志的窗口图像的样本。
进一步地，已确认的交通标志窗口图像的样本可以包括：原始已确认交通标志窗口图像的样本，以及将原始已确认交通标志窗口图像的样本进行以下一项或多项处理后归一至预设窗口尺寸的图像：旋转、平移和缩放。通过扩充样本，可以提高样本识别的准确率。
与步骤401中设置损失函数层相对应，通过误差反向传播BP算法对卷积层和全连接层两者的权重进行迭代可以包括：通过损失函数和BP算法对卷积层和全连接层两者的权重分别进行迭代。例如，在上述用高斯分布初始化所有权重之后，采用随机梯度下降算法对网络模型进行多次迭代，每迭代一次先由上述损失函数公式(a)正向计算网络的损失函数L，然后反向逐层计算L相对于每层权重Wi的梯度，之后根据每层权重Wi的梯度更新权重Wi，即
Figure PCTCN2015098903-appb-000002
其中，α为预设的梯度下降的学习率，
Figure PCTCN2015098903-appb-000003
为损失函数L相对于权重Wi的梯度。
之后，可以通过步骤S403或步骤S404确定最优权重的卷积神经网络模型。
步骤S403：若当前迭代后的权重与上一次迭代后的权重的差值小于预设值，则将当前迭代后的权重确定为最优权重。
步骤S404：若出现错误率最低的迭代后的权重，则将错误率最低的迭代后的权重确定为最优权重。
步骤S405：将包括最优权重的卷积神经网络模型设定为预先训练的卷积神经网络模型。
在训练出预先训练的卷积神经网络模型之后，则图1中的步骤103可以包括：将待确认交通标志窗口图像输入卷积神经网络模型，得到 归一化层输出的权重最大的交通标志类别；将权重最大的交通标志类别设定为识别得到的交通标志类别。
图6示出了根据本申请实施例的识别交通标志的装置的结构示例图。
如图6所示，所述识别交通标志的装置600可以包括：特征值获取模块610，交通标志检测模块620和交通标志识别模块630。
特征值获取模块610，用于获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值。
可选地，特征值获取模块610可以包括：积分图获取第一子模块611，用于获取由全景球面图像在预定积分通道的积分图划分得到的扫描窗口图像的积分图；以及特征值获取第一子模块612，用于根据扫描窗口图像的积分图，获取扫描窗口图像的特征值。
交通标志检测模块620，用于根据扫描窗口图像的特征值和预先训练的检测分类器模型，检测扫描窗口图像，得到待确认交通标志窗口图像，其中，预先训练的检测分类器模型根据扫描窗口图像的样本及其特征值训练得到。
交通标志识别模块630，用于根据预先训练的卷积神经网络模型，识别待确认交通标志窗口图像，得到交通标志类别，其中，预先训练的卷积神经网络模型根据已确认的交通标志窗口图像的样本及其交通标志类别训练得到。
本申请上述实施例的识别交通标志的装置，提高了在全景图像中检测和识别交通标志的准确性，同时提高了路网数据的更新效率。
可选地，上述的预先训练的检测分类器模型可以通过如下方法获得：获取扫描窗口图像的样本中的正样本和负样本，其中正样本包括交通标志的窗口图像或包括交通标志及其周围扩充预设像素的窗口图像，负样本包括去除正样本的扫描窗口图像；获取正样本和负样本在预定积分通道通过预定特征算法得到的特征值；根据正样本和负样本及其在预定积分通道通过预定特征算法得到的特征值，通过boosting算法训练检测分类器模型，得到预先训练的检测分类器模型。
上述获得预先训练的检测分类器模型的方法，可以通过图7示出 的检测分类器模型训练装置来实现。
图7示出了根据本申请实施例的检测分类器模型训练装置的示例性结构图。
如图7所示，检测分类器模型训练装置700可以包括：样本获取模块701，样本特征值获取模块702和检测模型训练模块703。
样本获取模块701，用于获取扫描窗口图像的样本中的正样本和负样本，其中正样本包括交通标志的窗口图像或包括交通标志及其周围扩充预设像素的窗口图像，负样本包括去除正样本的扫描窗口图像。
样本特征值获取模块702，用于获取正样本和负样本在预定积分通道通过预定特征算法得到的特征值。
其中，预定积分通道包括以下一项或多项：灰度图像通道，红、绿、蓝单色图像通道，不同角度参数的梯度图像通道，和边缘检测图像通道；以及预定特征算法包括以下一项或多项：哈尔特征算法、局部二值模式特征算法、直方图特征算法和随机对特征算法。
检测模型训练模块703，用于根据正样本和负样本及其在预定积分通道通过预定特征算法得到的特征值，通过boosting算法训练检测分类器模型，得到预先训练的检测分类器模型。
以下结合图8，在上述实施例的基础上，描述获取待确认交通标志窗口图像的装置。
图8示出了用于获取待确认交通标志窗口图像的装置的示例性结构图。
如图8所示，用于获取待确认交通标志窗口图像的装置800可以包括：降采样子模块801、积分图获取第二子模块802、特征值第二获取子模块803和多级检测子模块804。
降采样子模块801，用于对全景球面图像连续降采样，获得图像金字塔。
积分图获取第二子模块802，用于获取由图像金字塔在预定积分通道的积分图划分得到的图像金字塔中每级图像的扫描窗口图像的积分图。
特征值第二获取子模块803，用于根据每级图像的扫描窗口图像 的积分图，获取每级图像的扫描窗口图像的特征值。
多级检测子模块804，用于根据每级图像的扫描窗口图像的特征值和预先训练的检测分类器模型，检测每级图像的扫描窗口图像，得到待确认交通标志窗口图像。
通过本申请实施例上述获取待确认交通标志窗口图像的装置，可以进一步提高获取待确认交通标志窗口图像的正确率以及提高检测待确认交通标志窗口图像的速度。
可选地，上述图6中的预先训练的卷积神经网络模型可以通过如下方法获得：按照高斯分布，初始化预设的卷积神经网络模型的卷积层和全连接层两者的权重，其中，所述预设的卷积神经网络模型包括依次连接的卷积层、抽取层、全连接层和归一化层；根据已确认的交通标志窗口图像的样本及其交通标志类别，通过误差反向传播BP算法对所述卷积层和所述全连接层两者的权重进行迭代，若当前迭代后的权重与上一次迭代后的权重的差值小于预设值，则将当前迭代后的权重确定为最优权重，将包括最优权重的卷积神经网络模型设定为所述预先训练的卷积神经网络模型；或若出现错误率最低的迭代后的权重，则将错误率最低的迭代后的权重确定为最优权重，将包括最优权重的卷积神经网络模型设定为所述预先训练的卷积神经网络模型。
其中，已确认的交通标志窗口图像的样本，是在待确认交通标志窗口图像中提取的样本中，根据接收的人工标注确认后的包括交通标志的窗口图像的样本。
进一步地，已确认的交通标志窗口图像的样本可以包括：原始已确认交通标志窗口图像的样本，以及将原始已确认交通标志窗口图像的样本进行以下一项或多项处理后归一至预设窗口尺寸的图像：旋转、平移和缩放。
其中，依次连接的卷积层、抽取层、全连接层和归一化层包括：多个卷积层以及与卷积层一一对应的抽取层、一个以上全连接层和一个归一化层。
进一步地，预设的卷积神经网络模型还可以包括损失函数层。
对应于预设的卷积神经网络模型还包括损失函数层，通过误差反 向传播BP算法对所述卷积层和所述全连接层两者的权重进行迭代可以包括：通过损失函数和BP算法对卷积层和全连接层两者的权重分别进行迭代。
上述获得预先训练的卷积神经网络模型的方法可以通过图9示出的训练预先训练的卷积神经网络模型的装置来实现。
图9示出了根据本申请实施例的用于训练预先训练的卷积神经网络模型的装置的示例性结构图。
如图9所示，用于训练预先训练的卷积神经网络模型的装置900包括：
初始化模块901，用于按照高斯分布，初始化预设的卷积神经网络模型的卷积层和全连接层两者的权重，其中，预设的卷积神经网络模型包括依次连接的卷积层、抽取层、全连接层和归一化层。
其中，依次连接的卷积层、抽取层、全连接层和归一化层包括：多个卷积层以及与卷积层一一对应的抽取层、一个以上全连接层和一个归一化层。
可选地，预设的卷积神经网络模型还包括损失函数层。
权重迭代模块902，用于根据已确认的交通标志窗口图像的样本及其交通标志类别，通过误差反向传播BP算法对卷积层和全连接层两者的权重进行迭代；
第一最优权重确定模块903，用于若当前迭代后的权重与上一次迭代后的权重的差值小于预设值，则将当前迭代后的权重确定为最优权重。
第二最优权重确定模块904，用于若出现错误率最低的迭代后的权重，则将错误率最低的迭代后的权重确定为最优权重。
模型设定模块905，用于将包括最优权重的卷积神经网络模型设定为预先训练的卷积神经网络模型。
对应于预设的卷积神经网络模型还包括损失函数层，权重确定模块902进一步用于通过损失函数和BP算法对卷积层和全连接层两者的权重分别进行迭代。
在用于训练预先训练的卷积神经网络模型的装置900训练出预先 训练的卷积神经网络模型之后，则图6中的交通标志识别模块630可以包括：最大权重识别模块(未示出)，用于将待确认交通标志窗口图像输入卷积神经网络模型，得到归一化层输出的权重最大的交通标志类别；以及标志类别设定模块(未示出)，用于将权重最大的交通标志类别设定为识别得到的交通标志类。
应当理解，装置600中记载的诸单元与参考图1描述的方法中的各个步骤相对应。装置700中记载的诸单元与参考图2描述的方法中的各个步骤相对应。装置800中记载的诸单元与参考图3描述的方法中的各个步骤相对应。装置900中记载的诸单元与参考图4描述的方法中的各个步骤相对应。由此，上文针对识别交通标志的方法描述的操作和特征同样适用于装置600及其中包含的单元，上文针对训练预先训练的检测模型的方法描述的操作和特征同样适用于装置700及其中包含的单元，上文针对获取待确认交通标志窗口图像的方法描述的操作和特征同样适用于装置800及其中包含的单元，上文针对训练预先训练的卷积神经网络模型的方法描述的操作和特征同样适用于装置900及其中包含的单元，在此不再赘述。装置600、700、800和900中的相应单元可以与终端设备和/或服务器中的单元相互配合以实现本申请实施例的方案。
描述于本申请实施例中所涉及到的模块可以通过软件的方式实现，也可以通过硬件的方式来实现。所描述的模块也可以设置在处理器中，例如，可以描述为：一种处理器包括特征值获取模块，检测模块和识别模块。其中，这些模块的名称在某种情况下并不构成对该模块本身的限定，例如，特征值获取模块还可以被描述为“用于获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值的模块”。
图10是本申请的实施例提供的一种计算机系统的结构示意图。
下面参考图10，其示出了适于用来实现本申请实施例的设备的计算机系统700的结构示意图。
如图10所示，计算机系统1000包括中央处理单元(CPU)1001，其可以根据存储在只读存储器(ROM)1002中的程序或者从存储部分 708加载到随机访问存储器(RAM)1003中的程序而执行各种适当的动作和处理。当存储在ROM 1002中的程序或者从存储部分1008加载到RAM 1003中的程序被CPU 1001执行时，CPU 1001执行本申请描述的方法。在RAM 1003中，还存储有系统1000操作所需的各种程序和数据。CPU 1001、ROM 1002以及RAM 1003通过总线1004彼此相连。输入/输出(I/O)接口1005也连接至总线1004。
以下部件连接至I/O接口1005：包括键盘、鼠标等的输入部分1006；包括诸如阴极射线管(CRT)、液晶显示器(LCD)等以及扬声器等的输出部分1007；包括硬盘等的存储部分1008；以及包括诸如LAN卡、调制解调器等的网络接口卡的通信部分1009。通信部分1009经由诸如因特网的网络执行通信处理。驱动器1010也根据需要连接至I/O接口1005。可拆卸介质1011，诸如磁盘、光盘、磁光盘、半导体存储器等等，根据需要安装在驱动器1010上，以便于从其上读出的计算机程序根据需要被安装入存储部分1008。
特别地，根据本公开的实施例，上文参考流程图描述的过程可以被实现为计算机软件程序。例如，本公开的实施例包括一种计算机程序产品，其包括有形地包含在机器可读介质上的计算机程序，计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中，该计算机程序可以通过通信部分1009从网络上被下载和安装，和/或从可拆卸介质1011被安装。
附图中的流程图和框图，图示了按照本发明各种实施例的系统、方法的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件 与计算机指令的组合来实现。
作为另一方面，本申请还提供了一种计算机可读存储介质，该计算机可读存储介质可以是上述实施例中所述装置中所包含的计算机可读存储介质；也可以是单独存在，未装配入终端中的计算机可读存储介质。所述计算机可读存储介质存储有一个或者一个以上程序，所述程序被一个或者一个以上的处理器用来执行描述于本申请的识别交通标志的方法。
以上描述仅为本申请的较佳实施例以及对所运用技术原理的说明。本领域技术人员应当理解，本申请中所涉及的发明范围，并不限于上述技术特征的特定组合而成的技术方案，同时也应涵盖在不脱离所述发明构思的情况下，由上述技术特征或其等同特征进行任意组合而形成的其它技术方案。例如上述特征与本申请中公开的(但不限于)具有类似功能的技术特征进行互相替换而形成的技术方案。

Claims (16)

  1. 一种识别交通标志的方法，其特征在于，包括：
    获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值；
    根据所述扫描窗口图像的特征值和预先训练的检测分类器模型，检测扫描窗口图像，得到待确认交通标志窗口图像，其中，所述预先训练的检测分类器模型根据扫描窗口图像的样本及其特征值训练得到；
    根据预先训练的卷积神经网络模型，识别待确认交通标志窗口图像，得到交通标志类别，其中，所述预先训练的卷积神经网络模型根据已确认的交通标志窗口图像的样本及其交通标志类别训练得到。
  2. 根据权利要求1所述的方法，其特征在于，通过如下方法获得所述预先训练的检测分类器模型：
    获取扫描窗口图像的样本中的正样本和负样本，其中所述正样本包括交通标志的窗口图像或包括交通标志及其周围扩充预设像素的窗口图像，所述负样本包括去除正样本的扫描窗口图像；
    获取所述正样本和负样本在预定积分通道通过预定特征算法得到的特征值；
    根据所述正样本和负样本及其在预定积分通道通过预定特征算法得到的特征值，通过boosting算法训练检测分类器模型，得到所述预先训练的检测分类器模型。
  3. 根据权利要求1所述的方法，其特征在于，所述获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值包括：
    获取由全景球面图像在预定积分通道的积分图划分得到的扫描窗口图像的积分图；
    根据所述扫描窗口图像的积分图，获取所述扫描窗口图像的特征 值。
  4. 根据权利要求3所述的方法，其特征在于，所述获取由全景球面图像在预定积分通道的积分图划分得到的扫描窗口图像的积分图包括：对所述全景球面图像连续降采样，获得图像金字塔；获取由图像金字塔在预定积分通道的积分图划分得到的图像金字塔中每级图像的扫描窗口图像的积分图；
    所述根据所述扫描窗口图像的积分图，获取所述扫描窗口图像的特征值包括：根据所述每级图像的扫描窗口图像的积分图，获取每级图像的扫描窗口图像的特征值；以及
    所述根据所述扫描窗口图像的特征值和预先训练的检测分类器模型，检测扫描窗口图像，得到待确认交通标志窗口图像包括：根据所述每级图像的扫描窗口图像的特征值和预先训练的检测分类器模型，检测所述每级图像的扫描窗口图像，得到待确认交通标志窗口图像。
  5. 根据权利要求1-4之一所述的方法，其特征在于，通过如下方法获得所述预先训练的卷积神经网络模型：
    按照高斯分布，初始化预设的卷积神经网络模型的卷积层和全连接层两者的权重，其中，所述预设的卷积神经网络模型包括依次连接的卷积层、抽取层、全连接层和归一化层；
    根据已确认的交通标志窗口图像的样本及其交通标志类别，通过误差反向传播BP算法对所述卷积层和所述全连接层两者的权重进行迭代，若当前迭代后的权重与上一次迭代后的权重的差值小于预设值，则将当前迭代后的权重确定为最优权重，将包括最优权重的卷积神经网络模型设定为所述预先训练的卷积神经网络模型；或若出现错误率最低的迭代后的权重，则将错误率最低的迭代后的权重确定为最优权重，将包括最优权重的卷积神经网络模型设定为所述预先训练的卷积神经网络模型。
  6. 根据权利要求5所述的方法，其特征在于，其特征在于，所述 预设的卷积神经网络模型还包括损失函数层；以及
    所述通过误差反向传播BP算法对所述卷积层和所述全连接层两者的权重进行迭代包括：通过损失函数和BP算法对所述卷积层和所述全连接层两者的权重分别进行迭代。
  7. 根据权利要求5所述的方法，其特征在于，所述根据预先训练的卷积神经网络模型，识别待确认交通标志窗口图像，得到交通标志类别包括：
    将所述待确认交通标志窗口图像输入所述卷积神经网络模型，得到归一化层输出的权重最大的交通标志类别；
    将所述权重最大的交通标志类别设定为识别得到的交通标志类别。
  8. 一种识别交通标志的装置，其特征在于，包括：
    特征值获取模块，用于获取由全景球面图像划分得到的扫描窗口图像在预定积分通道通过预定特征算法得到的特征值；
    交通标志检测模块，用于根据所述扫描窗口图像的特征值和预先训练的检测分类器模型，检测扫描窗口图像，得到待确认交通标志窗口图像，其中，所述预先训练的检测分类器模型根据扫描窗口图像的样本及其特征值训练得到；
    交通标志识别模块，用于根据预先训练的卷积神经网络模型，识别待确认交通标志窗口图像，得到交通标志类别，其中，所述预先训练的卷积神经网络模型根据已确认的交通标志窗口图像的样本及其交通标志类别训练得到。
  9. 根据权利要求8所述的装置，其特征在于，所述预先训练的检测分类器模型通过如下方法获得：
    获取扫描窗口图像的样本中的正样本和负样本，其中所述正样本包括交通标志的窗口图像或包括交通标志及其周围扩充预设像素的窗口图像，所述负样本包括去除正样本的扫描窗口图像；
    获取所述正样本和负样本在预定积分通道通过预定特征算法得到的特征值；
    根据所述正样本和负样本及其在预定积分通道通过预定特征算法得到的特征值，通过boosting算法训练检测分类器模型，得到所述预先训练的检测分类器模型。
  10. 根据权利要求8所述的装置，其特征在于，所述特征值获取模块包括：
    积分图获取第一子模块，用于获取由全景球面图像在预定积分通道的积分图划分得到的扫描窗口图像的积分图；
    特征值获取第一子模块，用于根据所述扫描窗口图像的积分图，获取所述扫描窗口图像的特征值。
  11. 根据权利要求10所述的装置，其特征在于，所述积分图获取第一子模块包括：降采样子模块，用于对所述全景球面图像连续降采样，获得图像金字塔；积分图获取第二子模块，用于获取由图像金字塔在预定积分通道的积分图划分得到的图像金字塔中每级图像的扫描窗口图像的积分图；
    所述特征值获取第一子模块包括：特征值第二获取子模块，用于根据所述每级图像的扫描窗口图像的积分图，获取每级图像的扫描窗口图像的特征值；以及
    所述检测模块包括：多级检测子模块，用于根据所述每级图像的扫描窗口图像的特征值和预先训练的检测分类器模型，检测所述每级图像的扫描窗口图像，得到待确认交通标志窗口图像。
  12. 根据权利要求8-11之一所述的装置，其特征在于，所述预先训练的卷积神经网络模型通过如下方法获得：
    按照高斯分布，初始化预设的卷积神经网络模型的卷积层和全连接层两者的权重，其中，所述预设的卷积神经网络模型包括依次连接的卷积层、抽取层、全连接层和归一化层；
    根据已确认的交通标志窗口图像的样本及其交通标志类别，通过误差反向传播BP算法对所述卷积层和所述全连接层两者的权重进行迭代，若当前迭代后的权重与上一次迭代后的权重的差值小于预设值，则将当前迭代后的权重确定为最优权重，将包括最优权重的卷积神经网络模型设定为所述预先训练的卷积神经网络模型；或若出现错误率最低的迭代后的权重，则将错误率最低的迭代后的权重确定为最优权重，将包括最优权重的卷积神经网络模型设定为所述预先训练的卷积神经网络模型。
  13. 根据权利要求12所述的装置，其特征在于，其特征在于，所述预设的卷积神经网络模型还包括损失函数层；以及
    所述权重确定模块进一步用于通过损失函数和BP算法对所述卷积层和所述全连接层两者的权重分别进行迭代。
  14. 根据权利要求12所述的装置，其特征在于，所述交通标志识别模块包括：
    最大权重识别模块，用于将所述待确认交通标志窗口图像输入所述卷积神经网络模型，得到归一化层输出的权重最大的交通标志类别；
    标志类别设定模块，用于将所述权重最大的交通标志类别设定为识别得到的交通标志类。
  15. 一种设备，包括：
    处理器；和
    存储器，
    所述存储器中存储有能够被所述处理器执行的计算机可读指令，在所述计算机可读指令被执行时，所述处理器执行权利要求1至7中任一项所述的方法。
  16. 一种非易失性计算机存储介质，所述计算机存储介质存储有能够被处理器执行的计算机可读指令，当所述计算机可读指令被处理 器执行时，所述处理器执行权利要求1至7中任一项所述的方法。
PCT/CN2015/098903 2015-03-31 2015-12-25 识别交通标志的方法和装置 WO2016155371A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
JP2016559945A JP6400117B2 (ja) 2015-03-31 2015-12-25 交通標識を認識する方法及び装置
KR1020167027213A KR101856584B1 (ko) 2015-03-31 2015-12-25 교통 표지의 인식 방법 및 장치

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201510150525.8A CN104700099B (zh) 2015-03-31 2015-03-31 识别交通标志的方法和装置
CN201510150525.8 2015-03-31

Publications (1)

Publication Number Publication Date
WO2016155371A1 true WO2016155371A1 (zh) 2016-10-06

Family

ID=53347198

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/098903 WO2016155371A1 (zh) 2015-03-31 2015-12-25 识别交通标志的方法和装置

Country Status (4)

Country Link
JP (1) JP6400117B2 (zh)
KR (1) KR101856584B1 (zh)
CN (1) CN104700099B (zh)
WO (1) WO2016155371A1 (zh)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2018173946A (ja) * 2017-03-06 2018-11-08 バイドゥ ユーエスエイ エルエルシーＢａｉｄｕ ＵＳＡ ＬＬＣ 自動運転車両に用いる畳み込みニューラルネットワークモデルのコンボリューション／デコンボリューション層とバッチ正規化層のオフライン組み合わせ
CN109255279A (zh) * 2017-07-13 2019-01-22 深圳市凯立德科技股份有限公司 一种交通标志检测识别的方法及系统
CN109919017A (zh) * 2019-01-28 2019-06-21 平安科技（深圳）有限公司 人脸识别优化方法、装置、计算机设备和存储介质
CN110175500A (zh) * 2019-04-03 2019-08-27 平安科技（深圳）有限公司 指静脉比对方法、装置、计算机设备及存储介质
US20190286917A1 (en) * 2018-03-19 2019-09-19 Kabushiki Kaisha Toshiba Recognition device, vehicle system and storage medium
CN110348396A (zh) * 2019-07-15 2019-10-18 南京信息工程大学 一种基于深度学习的道路上方文字交通标志识别方法及装置
CN110889378A (zh) * 2019-11-28 2020-03-17 湖南率为控制科技有限公司 一种多视图融合交通标志检测识别方法及其系统
CN111062885A (zh) * 2019-12-09 2020-04-24 中国科学院自动化研究所 基于多阶段迁移学习的标志检测模型训练及标志检测方法
CN111199217A (zh) * 2020-01-09 2020-05-26 上海应用技术大学 一种基于卷积神经网络的交通标志识别方法和系统
CN111310844A (zh) * 2020-02-26 2020-06-19 广州华工邦元信息技术有限公司 车辆识别模型构建方法及装置、识别方法及装置
CN111597875A (zh) * 2020-04-01 2020-08-28 吉利汽车研究院（宁波）有限公司 一种交通标志识别方法、装置、设备及存储介质
CN111881944A (zh) * 2020-07-08 2020-11-03 贵州无忧天空科技有限公司 图像鉴别的方法、电子设备和计算机可读介质
CN113409393A (zh) * 2019-05-17 2021-09-17 百度在线网络技术（北京）有限公司 用于识别交通标志牌的方法及装置

Families Citing this family (39)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104700099B (zh) * 2015-03-31 2017-08-11 百度在线网络技术（北京）有限公司 识别交通标志的方法和装置
CN105069472B (zh) * 2015-08-03 2018-07-27 电子科技大学 一种基于卷积神经网络自适应的车辆检测方法
CN105335754A (zh) * 2015-10-29 2016-02-17 小米科技有限责任公司 文字识别方法及装置
CN105678322A (zh) * 2015-12-31 2016-06-15 百度在线网络技术（北京）有限公司 样本标注方法和装置
CN105869016A (zh) * 2016-03-28 2016-08-17 天津中科智能识别产业技术研究院有限公司 一种基于卷积神经网络的点击通过率预估方法
CN105956524A (zh) * 2016-04-22 2016-09-21 北京智芯原动科技有限公司 一种交通标识识别方法及装置
CN106022380A (zh) * 2016-05-25 2016-10-12 中国科学院自动化研究所 基于深度学习的个体身份识别方法
CN106778835B (zh) * 2016-11-29 2020-03-24 武汉大学 融合场景信息和深度特征的遥感图像机场目标识别方法
US10282999B2 (en) * 2017-03-17 2019-05-07 GM Global Technology Operations LLC Road construction detection systems and methods
CN106886801B (zh) 2017-04-14 2021-12-17 北京图森智途科技有限公司 一种图像语义分割方法及装置
CN107067043B (zh) * 2017-05-25 2020-07-24 哈尔滨工业大学 一种农作物病虫害检测方法
KR101834778B1 (ko) 2017-05-30 2018-03-07 만도헬라일렉트로닉스(주) 교통 표지판 인식장치 및 방법
WO2019009490A1 (ko) 2017-07-06 2019-01-10 삼성전자 주식회사 영상을 부호화/복호화 하는 방법 및 그 장치
US10986356B2 (en) 2017-07-06 2021-04-20 Samsung Electronics Co., Ltd. Method for encoding/decoding image and device therefor
WO2019009447A1 (ko) * 2017-07-06 2019-01-10 삼성전자 주식회사 영상을 부호화/복호화 하는 방법 및 그 장치
US9947103B1 (en) * 2017-10-03 2018-04-17 StradVision, Inc. Learning method and learning device for improving image segmentation and testing method and testing device using the same
CN107679577A (zh) * 2017-10-12 2018-02-09 理光图像技术（上海）有限公司 基于深度学习的图像检测方法
CN109726728B (zh) * 2017-10-31 2020-12-15 阿里巴巴（中国）有限公司 一种训练数据生成方法及装置
CN107944437B (zh) * 2017-12-31 2018-12-14 广州二元科技有限公司 一种基于神经网络和积分图像的人脸定位方法
CN110096937A (zh) * 2018-01-31 2019-08-06 北京四维图新科技股份有限公司 一种用于辅助车辆自动驾驶的图像识别的方法及装置
CN108549927A (zh) * 2018-03-16 2018-09-18 深圳市商汤科技有限公司 物体检测网络的训练方法和特征监督图生成方法
CN108520212A (zh) * 2018-03-27 2018-09-11 东华大学 基于改进的卷积神经网络的交通标志检测方法
KR102102161B1 (ko) * 2018-05-18 2020-04-20 오드컨셉 주식회사 이미지 내 객체의 대표 특성을 추출하는 방법, 장치 및 컴퓨터 프로그램
CN108960308A (zh) * 2018-06-25 2018-12-07 中国科学院自动化研究所 交通标志识别方法、装置、车载终端及车辆
KR102106029B1 (ko) * 2018-07-19 2020-05-04 네이버랩스 주식회사 간판 검출 성능 향상을 위한 방법 및 시스템
KR102096926B1 (ko) * 2018-07-19 2020-04-06 네이버랩스 주식회사 관심 영역 변화를 검출하는 방법 및 시스템
CN109102017B (zh) * 2018-08-09 2021-08-03 百度在线网络技术（北京）有限公司 神经网络模型处理方法、装置、设备及可读存储介质
CN110879950A (zh) * 2018-09-06 2020-03-13 北京市商汤科技开发有限公司 多级目标分类及交通标志检测方法和装置、设备、介质
CN109492628B (zh) * 2018-10-08 2021-11-16 杭州电子科技大学 一种应用于课堂违纪抓取的球形定位系统的实现方法
WO2020080765A1 (en) 2018-10-19 2020-04-23 Samsung Electronics Co., Ltd. Apparatuses and methods for performing artificial intelligence encoding and artificial intelligence decoding on image
WO2020080665A1 (en) 2018-10-19 2020-04-23 Samsung Electronics Co., Ltd. Methods and apparatuses for performing artificial intelligence encoding and artificial intelligence decoding on image
US10402977B1 (en) * 2019-01-25 2019-09-03 StradVision, Inc. Learning method and learning device for improving segmentation performance in road obstacle detection required to satisfy level 4 and level 5 of autonomous vehicles using laplacian pyramid network and testing method and testing device using the same
KR102436512B1 (ko) 2019-10-29 2022-08-25 삼성전자주식회사 부호화 방법 및 그 장치, 복호화 방법 및 그 장치
CN110796109A (zh) * 2019-11-05 2020-02-14 哈尔滨理工大学 一种基于模型融合的驾驶员分心行为识别方法
CN111178153A (zh) * 2019-12-09 2020-05-19 武汉光庭信息技术股份有限公司 一种交通标志检测方法及系统
CN111178371B (zh) * 2019-12-17 2023-12-01 深圳市优必选科技股份有限公司 目标检测方法、装置和计算机存储介质
KR102287942B1 (ko) 2020-02-24 2021-08-09 삼성전자주식회사 전처리를 이용한 영상의 ａｉ 부호화 및 ａｉ 복호화 방법, 및 장치
KR102189926B1 (ko) * 2020-03-30 2020-12-11 네이버랩스 주식회사 관심 영역 변화를 검출하는 방법 및 시스템
CN112085056B (zh) * 2020-08-05 2023-12-29 深圳市优必选科技股份有限公司 目标检测模型生成方法、装置、设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130272575A1 (en) * 2011-11-01 2013-10-17 Intel Corporation Object detection using extended surf features
CN103366177A (zh) * 2012-03-28 2013-10-23 佳能株式会社 对象检测分类器生成方法和设备、图像对象检测方法和设备
CN104036323A (zh) * 2014-06-26 2014-09-10 叶茂 一种基于卷积神经网络的车辆检测方法
CN104700099A (zh) * 2015-03-31 2015-06-10 百度在线网络技术（北京）有限公司 识别交通标志的方法和装置

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH08221378A (ja) * 1995-02-10 1996-08-30 Ricoh Co Ltd 学習機械
JP2004191339A (ja) * 2002-12-13 2004-07-08 Sharp Corp 位置情報検索方法、位置情報検索装置、位置情報検索端末、及び、位置情報検索システム
JP2005311691A (ja) * 2004-04-21 2005-11-04 Toyota Central Res & Dev Lab Inc 物体検出装置及び方法
JP5364578B2 (ja) * 2006-07-12 2013-12-11 コファックス， インコーポレイテッド トランスダクティブデータ分類のための方法およびシステム、ならびに機械学習手法を用いたデータ分類方法
JP2013196454A (ja) * 2012-03-21 2013-09-30 Honda Elesys Co Ltd 画像処理装置、画像処理方法および画像処理プログラム
JP2014010633A (ja) * 2012-06-29 2014-01-20 Honda Elesys Co Ltd 画像認識装置、画像認識方法、及び画像認識プログラム
JP2014071832A (ja) * 2012-10-01 2014-04-21 Toshiba Corp 物体検出装置及びその検出方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130272575A1 (en) * 2011-11-01 2013-10-17 Intel Corporation Object detection using extended surf features
CN103366177A (zh) * 2012-03-28 2013-10-23 佳能株式会社 对象检测分类器生成方法和设备、图像对象检测方法和设备
CN104036323A (zh) * 2014-06-26 2014-09-10 叶茂 一种基于卷积神经网络的车辆检测方法
CN104700099A (zh) * 2015-03-31 2015-06-10 百度在线网络技术（北京）有限公司 识别交通标志的方法和装置

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2018173946A (ja) * 2017-03-06 2018-11-08 バイドゥ ユーエスエイ エルエルシーＢａｉｄｕ ＵＳＡ ＬＬＣ 自動運転車両に用いる畳み込みニューラルネットワークモデルのコンボリューション／デコンボリューション層とバッチ正規化層のオフライン組み合わせ
CN109255279A (zh) * 2017-07-13 2019-01-22 深圳市凯立德科技股份有限公司 一种交通标志检测识别的方法及系统
US10984256B2 (en) * 2018-03-19 2021-04-20 Kabushiki Kaisha Toshiba Recognition device, vehicle system and storage medium
US20190286917A1 (en) * 2018-03-19 2019-09-19 Kabushiki Kaisha Toshiba Recognition device, vehicle system and storage medium
CN109919017A (zh) * 2019-01-28 2019-06-21 平安科技（深圳）有限公司 人脸识别优化方法、装置、计算机设备和存储介质
CN109919017B (zh) * 2019-01-28 2023-09-01 平安科技（深圳）有限公司 人脸识别优化方法、装置、计算机设备和存储介质
CN110175500A (zh) * 2019-04-03 2019-08-27 平安科技（深圳）有限公司 指静脉比对方法、装置、计算机设备及存储介质
CN110175500B (zh) * 2019-04-03 2024-01-19 平安科技（深圳）有限公司 指静脉比对方法、装置、计算机设备及存储介质
CN113409393B (zh) * 2019-05-17 2023-10-03 百度在线网络技术（北京）有限公司 用于识别交通标志牌的方法及装置
CN113409393A (zh) * 2019-05-17 2021-09-17 百度在线网络技术（北京）有限公司 用于识别交通标志牌的方法及装置
CN110348396A (zh) * 2019-07-15 2019-10-18 南京信息工程大学 一种基于深度学习的道路上方文字交通标志识别方法及装置
CN110889378A (zh) * 2019-11-28 2020-03-17 湖南率为控制科技有限公司 一种多视图融合交通标志检测识别方法及其系统
CN111062885A (zh) * 2019-12-09 2020-04-24 中国科学院自动化研究所 基于多阶段迁移学习的标志检测模型训练及标志检测方法
CN111062885B (zh) * 2019-12-09 2023-09-12 中国科学院自动化研究所 基于多阶段迁移学习的标志检测模型训练及标志检测方法
CN111199217B (zh) * 2020-01-09 2023-03-28 上海应用技术大学 一种基于卷积神经网络的交通标志识别方法和系统
CN111199217A (zh) * 2020-01-09 2020-05-26 上海应用技术大学 一种基于卷积神经网络的交通标志识别方法和系统
CN111310844A (zh) * 2020-02-26 2020-06-19 广州华工邦元信息技术有限公司 车辆识别模型构建方法及装置、识别方法及装置
CN111597875A (zh) * 2020-04-01 2020-08-28 吉利汽车研究院（宁波）有限公司 一种交通标志识别方法、装置、设备及存储介质
CN111881944A (zh) * 2020-07-08 2020-11-03 贵州无忧天空科技有限公司 图像鉴别的方法、电子设备和计算机可读介质

Also Published As

Publication number Publication date
KR101856584B1 (ko) 2018-05-10
JP2017516197A (ja) 2017-06-15
CN104700099A (zh) 2015-06-10
KR20160132331A (ko) 2016-11-17
CN104700099B (zh) 2017-08-11
JP6400117B2 (ja) 2018-10-03

Similar Documents

Publication Publication Date Title
WO2016155371A1 (zh) 识别交通标志的方法和装置
CN108197326B (zh) 一种车辆检索方法及装置、电子设备、存储介质
CN111898523A (zh) 一种基于迁移学习的遥感图像特种车辆目标检测方法
WO2016150240A1 (zh) 身份认证方法和装置
CN111461213B (zh) 一种目标检测模型的训练方法、目标快速检测方法
CN110033040B (zh) 一种火焰识别方法、系统、介质和设备
CN113177456B (zh) 基于单阶段全卷积网络和多特征融合的遥感目标检测方法
CN110705565A (zh) 淋巴结肿瘤区域识别方法与装置
JP2017033197A (ja) 変化領域検出装置、方法、及びプログラム
CN113052170B (zh) 一种无约束场景下的小目标车牌识别方法
CN111860106B (zh) 一种无监督的桥梁裂缝识别方法
CN109858438B (zh) 一种基于模型拟合的车道线检测方法
CN111881944A (zh) 图像鉴别的方法、电子设备和计算机可读介质
CN109584206B (zh) 零件表面瑕疵检测中神经网络的训练样本的合成方法
CN111539456B (zh) 一种目标识别方法及设备
CN114495010A (zh) 一种基于多特征学习的跨模态行人重识别方法及系统
CN110458019B (zh) 稀缺认知样本条件下的排除倒影干扰的水面目标检测方法
CN111444816A (zh) 一种基于Faster RCNN的多尺度密集行人检测方法
CN111832463A (zh) 一种基于深度学习的交通标志检测方法
CN111046861B (zh) 识别红外影像的方法、构建识别模型的方法及应用
CN114445788A (zh) 车辆停放检测方法、装置、终端设备和可读存储介质
CN107679528A (zh) 一种基于AdaBoost‑SVM集成学习算法的行人检测方法
CN114627424A (zh) 一种基于视角转化的步态识别方法和系统
JP2021009493A (ja) 画像処理装置、画像処理装置の制御方法及びプログラム
JP2007139421A (ja) 形態分類装置及び方法、並びに変化領域抽出装置及び方法

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2016559945

Country of ref document: JP

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 20167027213

Country of ref document: KR

Kind code of ref document: A

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

Ref document number: 15887321

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: 15887321

Country of ref document: EP

Kind code of ref document: A1