CN110705474A - Pedestrian attribute identification method and device - Google Patents
Pedestrian attribute identification method and device Download PDFInfo
- Publication number
- CN110705474A CN110705474A CN201910943815.6A CN201910943815A CN110705474A CN 110705474 A CN110705474 A CN 110705474A CN 201910943815 A CN201910943815 A CN 201910943815A CN 110705474 A CN110705474 A CN 110705474A
- Authority
- CN
- China
- Prior art keywords
- layer
- attribute
- composite
- pedestrian
- feature
- Prior art date
- Legal status (The legal status 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 status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The disclosure relates to a pedestrian attribute identification method and device. The method comprises the following steps: performing feature extraction on the sample image according to the convolutional neural network to obtain a plurality of initial feature layers, wherein the sample image is an image in a preset pedestrian attribute sample set, and the image in the preset pedestrian attribute sample set has a plurality of pedestrian attributes; performing layer-by-layer feature fusion on the plurality of initial feature layers from top to bottom to obtain a plurality of composite feature layers; according to the space transformation network and the multiple composite characteristic layers, the attribute positioning identification module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer is determined, so that the accuracy and efficiency of pedestrian attribute identification can be improved.
Description
Technical Field
The present disclosure relates to the field of computer vision technologies, and in particular, to a method and an apparatus for identifying a pedestrian attribute.
Background
Pedestrian Attribute Recognition (Pedestrian Attribute Recognition) refers to the use of a computer to predict and analyze various types of Attribute information related to pedestrians in an image. Common pedestrian attribute identification includes identifying macroscopic attributes of a pedestrian such as gender, skin color, age, posture and the like, and also includes specific character attributes such as a backpack type, a clothes type and color, a trousers type and color, a current action and the like. In recent years, the pedestrian attribute identification technology has attracted much attention, and research work on the pedestrian attribute identification technology has a high value in academic research, industrial application and the like. The existing pedestrian attribute identification algorithm based on deep learning mainly comprises a global type and a local type. The integral algorithm considers the pedestrian attribute identification problem as a multi-label classification problem, usually a Convolutional Neural Network (CNN) is used for extracting required features on the whole input picture, a full connection layer is used for attribute prediction at the top end of the network, and all attributes share the features. The local type algorithm focuses more on some local regions in the input image which are important for classification prediction, a human body posture estimation model is trained in advance to predict human body key points of the input image, and then local regions such as the head, the upper body and the lower body of a human body are roughly positioned according to the key points, so that attribute classification is guided.
Disclosure of Invention
In view of this, the present disclosure provides a method and an apparatus for identifying a pedestrian attribute, so as to effectively improve the accuracy and efficiency of identifying the pedestrian attribute.
According to a first aspect of the present disclosure, there is provided a pedestrian attribute identification method, including: performing feature extraction on a sample image according to a convolutional neural network to obtain a plurality of initial feature layers, wherein the sample image is an image in a preset pedestrian attribute sample set, and the image in the preset pedestrian attribute sample set has a plurality of pedestrian attributes; performing layer-by-layer feature fusion on the plurality of initial feature layers from top to bottom to obtain a plurality of composite feature layers; and determining an attribute positioning and identifying module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer according to the space transformation network and the plurality of composite characteristic layers.
In a possible implementation manner, performing layer-by-layer feature fusion on the plurality of initial feature layers from top to bottom to obtain a plurality of composite feature layers includes: directly determining the highest-level initial feature layer as a corresponding composite feature layer aiming at the highest-level initial feature layer; and performing feature fusion on the non-highest level initial feature layer and a composite feature layer corresponding to a previous level initial feature layer aiming at the non-highest level initial feature layer to obtain the composite feature layer corresponding to the non-highest level initial feature layer.
In one possible implementation, the non-highest level initial feature level is φiSaid initial characteristic layer phiiUpper part ofOne level of initial feature level is phii+1Said initial characteristic layer phii+1Corresponding composite feature layer is Xi+1(ii) a Aiming at the initial feature layer of the non-highest level, performing feature fusion on the initial feature layer of the non-highest level and the composite feature layer corresponding to the initial feature layer of the previous level to obtain the composite feature layer corresponding to the initial feature layer of the non-highest level, wherein the method comprises the following steps: combining the composite feature layer Xi+1Up-sampling and amplifying to the initial characteristic layer phiiAre the same, and the up-sampled composite feature layer X is obtainedi+1(ii) a The up-sampled composite characteristic layer X is processedi+1With respect to the initial feature layer phiiSplicing according to the number of channels to obtain the initial characteristic layer phiiCorresponding composite feature layer XiWherein the composite feature layer XiIs the initial feature layer phii+1And the initial feature layer phiiThe sum of the number of channels.
In a possible implementation manner, determining, according to the spatial transformation network and the plurality of composite feature layers, an attribute location identification module for performing location identification on each pedestrian attribute on each composite feature layer, includes: for any pedestrian attribute, determining a positioning identification result of the pedestrian attribute on each composite characteristic layer according to an attribute positioning identification module which carries out positioning identification on the pedestrian attribute on each composite characteristic layer; aiming at any pedestrian attribute, determining a global recognition result of the pedestrian attribute according to the highest level initial feature layer; according to the positioning recognition result of each pedestrian attribute on each composite feature layer, the global recognition result of each pedestrian attribute and the real attribute mark of each pedestrian attribute in the sample image, training an attribute positioning recognition module for positioning and recognizing each pedestrian attribute on each composite feature layer to obtain an attribute positioning recognition module for positioning and recognizing each pedestrian attribute on each composite feature layer.
In one possible implementation, for any pedestrian attribute, the pedestrian is attributed according to the attribute of each composite feature layerThe attribute positioning and identifying module for performing positioning and identifying determines a positioning and identifying result of the pedestrian attribute on each composite feature layer, and the attribute positioning and identifying module comprises: for any composite characteristic layer, an attribute positioning and identifying module for positioning and identifying the pedestrian attribute on the composite characteristic layer determines a positioning and identifying result of the pedestrian attribute on the composite characteristic layer through the following steps: passing the composite characteristic layer through a first full-connection layer to obtain a transformation parameter sx、sy、txAnd tyWherein the parameter s is transformedxFor scaling the transformation parameters in the horizontal direction, transformation parameters syFor scaling the transformation parameter in the vertical direction, transformation parameter txFor translation transformation parameters in the horizontal direction, transformation parameter tyTranslation transformation parameters in the vertical direction; according to said transformation parameter sx、sy、txAnd tyDetermining local features corresponding to the pedestrian attributes in the composite feature layer; and enabling the local features corresponding to the pedestrian attributes to pass through a second full-connection layer, and obtaining a positioning identification result of the pedestrian attributes on the composite feature layer.
In one possible implementation, the method further includes: performing the following feature calibration process on the composite feature layer before passing the composite feature layer through a first fully connected layer: sequentially passing the composite feature layer through a global average pooling layer, a 1 × 1 convolution layer, a ReLU active layer, a 1 × 1 convolution layer and a Sigmoid active layer to obtain a first calibration vector; multiplying the composite feature layer and the first calibration vector one by one according to channels to obtain a second calibration vector; and adding the composite characteristic layer and the second calibration vector element by element to obtain a calibrated composite characteristic layer.
In a possible implementation, the transformation parameter s is determined according tox、sy、txAnd tyDetermining local features corresponding to the pedestrian attributes in the composite feature layer, including: according to said transformation parameter sx、sy、txAnd tyDetermining a rectangular bounding box in the composite feature layer; in the compoundingAnd extracting the features in the rectangular bounding box from the feature layer, and determining the features as local features corresponding to the pedestrian attributes.
In a possible implementation manner, according to the positioning recognition result of each pedestrian attribute on each composite feature layer, the global recognition result of each pedestrian attribute, and the real attribute label of each pedestrian attribute in the sample image, training an attribute positioning recognition module that performs positioning recognition on each pedestrian attribute on each composite feature layer to obtain an attribute positioning recognition module that performs positioning recognition on each pedestrian attribute on each composite feature layer, includes: and training an attribute positioning and identifying module for positioning and identifying each pedestrian attribute on each composite characteristic layer through the following cross entropy loss functions:
wherein the content of the first and second substances,
the training loss of the ith feature layer is M is the number of the multiple pedestrian attributes, ymLabeling the real attribute of the mth pedestrian attribute,
as a result of identifying the m-th pedestrian attribute on the i-th feature layer, γmAnd the weight of the mth pedestrian attribute is shown, sigma is a preset parameter, and the ith characteristic layer is a composite characteristic layer or an initial characteristic layer.
In one possible implementation, the method further includes: for any pedestrian attribute, determining a positioning identification result of the pedestrian attribute in the test image on each composite feature layer according to an attribute positioning identification module which positions and identifies the pedestrian attribute on each composite feature layer; determining a global identification result of the pedestrian attribute in the test image according to the highest level initial feature layer; and determining the attribute recognition result of the pedestrian attribute in the test image according to the positioning recognition result of the pedestrian attribute in the test image on each composite feature layer and the global recognition result of the pedestrian attribute in the test image.
According to a second aspect of the present disclosure, there is provided a pedestrian attribute identification device including: the characteristic extraction module is used for extracting characteristics of a sample image according to a convolutional neural network to obtain a plurality of initial characteristic layers, wherein the sample image is an image in a preset pedestrian attribute sample set, and the image in the preset pedestrian attribute sample set has a plurality of pedestrian attributes; the characteristic fusion module is used for carrying out layer-by-layer characteristic fusion on the plurality of initial characteristic layers from top to bottom to obtain a plurality of composite characteristic layers; and the attribute positioning and identifying module is used for determining an attribute positioning and identifying module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer according to the space transformation network and the plurality of composite characteristic layers.
Performing feature extraction on the sample image according to the convolutional neural network to obtain a plurality of initial feature layers, wherein the sample image is an image in a preset pedestrian attribute sample set, and the image in the preset pedestrian attribute sample set has a plurality of pedestrian attributes; performing layer-by-layer feature fusion on the plurality of initial feature layers from top to bottom to obtain a plurality of composite feature layers; and determining an attribute positioning and identifying module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer according to the space transformation network and the plurality of composite characteristic layers. The method and the device can locate the corresponding local area of each pedestrian attribute in the image, and then identify the pedestrian attribute based on the local features, thereby improving the accuracy and efficiency of pedestrian attribute identification.
Other features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments, which proceeds with reference to the accompanying drawings.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate exemplary embodiments, features, and aspects of the disclosure and, together with the description, serve to explain the principles of the disclosure.
Fig. 1 shows a flow diagram of a pedestrian attribute identification method according to an embodiment of the disclosure;
FIG. 2 illustrates a schematic diagram of a pedestrian attribute identification system of an embodiment of the present disclosure;
FIG. 3 shows a schematic diagram of an attribute location module of an embodiment of the present disclosure;
FIG. 4 is a schematic diagram illustrating a pedestrian attribute zone location result according to an embodiment of the present disclosure;
fig. 5 is a schematic structural diagram of a pedestrian attribute identification device according to an embodiment of the present disclosure.
Detailed Description
Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. In the drawings, like reference numbers can indicate functionally identical or similar elements. While the various aspects of the embodiments are presented in drawings, the drawings are not necessarily drawn to scale unless specifically indicated.
The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration. Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.
Furthermore, in the following detailed description, numerous specific details are set forth in order to provide a better understanding of the present disclosure. It will be understood by those skilled in the art that the present disclosure may be practiced without some of these specific details. In some instances, methods, means, elements and circuits that are well known to those skilled in the art have not been described in detail so as not to obscure the present disclosure.
Fig. 1 shows a flowchart of a pedestrian attribute identification method according to an embodiment of the present disclosure. As shown in fig. 1, the method may include:
and step S11, extracting the characteristics of the sample image according to the convolutional neural network to obtain a plurality of initial characteristic layers, wherein the sample image is an image in a preset pedestrian attribute sample set, and the preset pedestrian attribute sample set comprises a plurality of pedestrian attributes.
And step S12, performing layer-by-layer feature fusion on the plurality of initial feature layers from top to bottom to obtain a plurality of composite feature layers.
And step S13, determining an attribute positioning and identifying module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer according to the space transformation network and the plurality of composite characteristic layers.
The method and the device can position the corresponding local area of each pedestrian attribute in the image, and then identify the pedestrian attribute based on the local features, thereby not only improving the accuracy and efficiency of pedestrian attribute identification, but also positioning the semantic area of each pedestrian attribute in the image in a display manner, and improving the interpretability of a pedestrian attribute identification algorithm.
Fig. 2 shows a schematic diagram of a pedestrian attribute identification system of an embodiment of the present disclosure. As shown in fig. 2, the preprocessed sample images in the preset sample set of pedestrian attributes are input into the pedestrian attribute identification system, where the sample image contains only one pedestrian and has a resolution of a preset value (e.g., a resolution of 256 × 128). The pedestrian property identification system may output identification results for a plurality of pedestrian properties, wherein the number of the plurality of pedestrian properties may be set by a preset pedestrian property sample set. For example, a common set of pedestrian attribute samples includes 51 pedestrian attributes. For any attribute of the pedestrian, the identification result is whether the attribute exists in the original image. The preset pedestrian attribute sample set comprises real attribute labels of each pedestrian attribute in the sample images, namely, the real identification result of each pedestrian attribute is known in advance aiming at the sample images.
The identification process of the pedestrian attribute identification system shown in fig. 2 will be described in detail below.
And performing feature extraction on the input sample image according to the convolutional neural network to obtain a plurality of initial feature layers. Also taking the above fig. 2 as an example, as shown in fig. 2, feature extraction is performed at three different positions of the basic backbone network to obtain an initial feature layer phi with a resolution of 32 × 161Initial feature layer phi with resolution of 16 x 82And an initial feature layer phi with a resolution of 8 x 43。
In a deep convolutional neural network, features at a high layer generally have stronger semantic information, but at the same time, the resolution is lower, so that the detail information is lacked; the features of the lower layers typically contain rich detail and resolution high enough, but low level of abstraction, semantic information is relatively weak. Therefore, feature fusion from top to bottom is performed on multiple initial feature layers.
In one possible implementation, performing layer-by-layer feature fusion on a plurality of initial feature layers from top to bottom to obtain a plurality of composite feature layers includes: directly determining the highest-level initial feature layer as a corresponding composite feature layer aiming at the highest-level initial feature layer; and performing feature fusion on the non-highest level initial feature layer and the composite feature layer corresponding to the previous level initial feature layer to obtain the composite feature layer corresponding to the non-highest level initial feature layer.
In one possible implementation, the non-highest level initial feature level is φiInitial feature layer phiiThe last level of initial feature layer is phii+1Initial feature layer phii+1Corresponding composite feature layer is Xi+1(ii) a Aiming at the initial feature layer of the non-highest level, performing feature fusion on the initial feature layer of the non-highest level and the composite feature layer corresponding to the initial feature layer of the previous level to obtain the composite feature layer corresponding to the initial feature layer of the non-highest level, wherein the method comprises the following steps: combining the feature layers Xi+1Up-sampling and amplifying to the initial characteristic layer phiiAre the same, and the up-sampled composite feature layer X is obtainedi+1(ii) a The up-sampled composite characteristic layer Xi+1Phi from the initial feature layeriSplicing according to the number of channels to obtain an initial characteristic layer phiiCorresponding composite feature layer XiWherein the characteristic layer X is compoundediThe number of channels is the initial feature layer phii+1And the initial feature layer phiiThe sum of the number of channels.
Also as an example in FIG. 2 above, due to the initial feature layer φ3Is the initial feature layer of the highest level, and the initial feature layer phi is integrated with other feature layers without being fused3Directly determining as corresponding composite feature layer X3(ii) a For the initial feature layer phi2First, compound feature layer X3(initial feature layer phi)3) Up-sampling and amplifying to the initial characteristic layer phi2Resolution ofThe ratio (16 x 8) is the same, and the initial feature layer phi is further adjusted2With the up-sampled composite feature layer X3Splicing according to the number of channels to obtain an initial characteristic layer phi2Corresponding composite feature layer X2(ii) a For the initial feature layer phi1First, compound feature layer X2Up-sampling and amplifying to the initial characteristic layer phi1Is equal to (32 x 16), thereby forming an initial feature layer phi1With the up-sampled composite feature layer X2Splicing according to the number of channels to obtain an initial characteristic layer phi1Corresponding composite feature layer X1. Assuming an initial feature layer phi1Initial characteristic layer phi2And initial feature layer phi3If the number of channels is 256, the composite feature layer X1Has a channel number of 768 and a composite characteristic layer X2Has a channel number of 512, a composite feature layer X3The number of channels of (2) is 256.
In one possible implementation manner, an attribute positioning and identifying module for determining, according to a Spatial Transformation Network (STN) and a plurality of composite feature layers, on each composite feature layer, a positioning and identifying attribute of each pedestrian is determined, and includes: aiming at any pedestrian attribute, determining a positioning identification result of the pedestrian attribute on each composite characteristic layer according to an attribute positioning identification module which carries out positioning identification on the pedestrian attribute on each composite characteristic layer; aiming at any pedestrian attribute, determining a global recognition result of the pedestrian attribute according to the highest level initial feature layer; and training an attribute positioning and identifying module for positioning and identifying the attributes of the pedestrians on each composite characteristic layer according to the positioning and identifying result of the attributes of the pedestrians on each composite characteristic layer, the overall identifying result of the attributes of the pedestrians and the real attribute label of the attributes of the pedestrians in the sample image to obtain the attribute positioning and identifying module for positioning and identifying the attributes of the pedestrians on each composite characteristic layer.
In a possible implementation manner, for any pedestrian attribute, determining a positioning recognition result of the pedestrian attribute on each composite feature layer according to an attribute positioning recognition module that performs positioning recognition on the pedestrian attribute on each composite feature layer includes: for any composite feature layer, the compositeThe attribute positioning and identifying module for positioning and identifying the pedestrian attribute on the composite characteristic layer determines the positioning and identifying result of the pedestrian attribute on the composite characteristic layer through the following steps: passing the composite characteristic layer through a first full-connection layer to obtain a transformation parameter sx、sy、txAnd tyWherein the parameter s is transformedxFor scaling the transformation parameters in the horizontal direction, transformation parameters syFor scaling the transformation parameter in the vertical direction, transformation parameter txFor translation transformation parameters in the horizontal direction, transformation parameter tyTranslation transformation parameters in the vertical direction; according to transformation parameters sx、sy、txAnd tyDetermining local features corresponding to the pedestrian attributes in the composite feature layer; and passing the local features corresponding to the pedestrian attributes through a second full-connection layer to obtain a positioning identification result of the pedestrian attributes on the composite feature layer.
In one possible implementation, the method further includes: before passing the composite feature layer through the first fully connected layer, performing the following feature calibration process on the composite feature layer: sequentially passing the composite characteristic layer through a global average pooling layer, a 1 × 1 convolution layer, a ReLU active layer, a 1 × 1 convolution layer and a Sigmoid active layer to obtain a first calibration vector; multiplying the composite feature layer and the first calibration vector one by one according to the channel to obtain a second calibration vector; and adding the composite characteristic layer and the second calibration vector element by element to obtain a calibrated composite characteristic layer.
In one possible implementation, the transformation parameter s is determined based onx、sy、txAnd tyAnd determining local features corresponding to the attributes of the pedestrians in the composite feature layer, wherein the local features comprise: according to transformation parameters sx、sy、txAnd tyDetermining a rectangular bounding box in the composite feature layer; and extracting the features in the rectangular bounding box from the composite feature layer, and determining the features as local features corresponding to the attributes of the pedestrians.
Still taking the above fig. 2 as an example, as shown in fig. 2, each composite feature layer corresponds to M attribute location identification modules (ALMs), where M is the number of multiple pedestrian attributes. The pedestrian attribute identification system shown in fig. 2 includes a total of 3M attribute locator modules, each of which acts on only one composite feature layer and one pedestrian attribute.
The specific positioning identification process of any attribute positioning module in any composite feature layer is described in detail below.
FIG. 3 shows a schematic diagram of an attribute location module of an embodiment of the present disclosure. As shown in FIG. 3, the input composite feature layer XiFirstly, reducing the resolution to 1 × 1 through a global average pooling layer, then sequentially performing feature processing on the 1 × 1 convolution layer, the ReLU activation layer and the 1 × 1 convolution layer, and then obtaining a first calibration vector through a Sigmoid activation layer (a commonly used Sigmoid activation function); composite feature layer X input therebyiMultiplying the first calibration vector by the first calibration vector one by one according to the channel to obtain a second calibration vector; last entered composite feature layer XiAnd adding the second calibration vector element by element to obtain a calibrated composite feature layer.
Calibrated composite feature layer XiFirstly, four transformation parameters s are obtained through a first full connection layer (FC)x、sy、txAnd tyWherein the parameter s is transformedxFor scaling the transformation parameters in the horizontal direction, transformation parameters syFor scaling the transformation parameter in the vertical direction, transformation parameter txFor translation transformation parameters in the horizontal direction, transformation parameter tyTranslation transformation parameters in the vertical direction; according to four transformation parameters sx、sy、txAnd tyComposite feature layer X that may be after calibrationiDetermining a rectangular bounding box R; finally, the calibrated composite feature layer X is obtained according to the rectangular bounding box RiThe feature in the rectangular bounding box R is extracted and determined as the local feature corresponding to the attribute of the pedestrian, and then the local feature passes through a second full connection layer (FC) to obtain the positioning and identifying result of the attribute positioning and identifying module.
The rectangular bounding box R can be used for positioning the semantic region of the pedestrian attribute in the image in a display mode, and the interpretability of the pedestrian attribute recognition algorithm is improved. The attribute positioning and identifying module can position the corresponding local area of the pedestrian attribute in the image, and then identify the pedestrian attribute based on the local feature, and compared with global identification, the calculation amount can be reduced, so that the accuracy and efficiency of pedestrian attribute identification are improved.
Fig. 4 is a schematic diagram illustrating a positioning result of a pedestrian attribute region according to an embodiment of the disclosure. As shown in fig. 4, each composite signature layer can be displayed with the region-locating results of the pedestrian attribute (plastic bag).
Also taking the above FIG. 2 as an example, as shown in FIG. 2, based on the composite feature layer X1The positioning identification results of M pedestrian attributes can be obtained based on the composite feature layer X2The positioning identification results of M pedestrian attributes can be obtained based on the composite feature layer X3And obtaining the positioning identification results of M pedestrian attributes. Based on the initial feature layer phi1Initial characteristic layer phi2And initial feature layer phi3And obtaining global identification results of M pedestrian attributes.
In the training process, in order to train the attribute positioning modules on different feature layers more fully, the recognition results of all the attribute positioning modules are used to participate in the training together.
In a possible implementation manner, according to the positioning recognition result of each pedestrian attribute on each composite feature layer, the global recognition result of each pedestrian attribute, and the real attribute label of each pedestrian attribute in the sample image, the attribute positioning recognition module for positioning and recognizing each pedestrian attribute on each composite feature layer is trained to obtain the attribute positioning recognition module for positioning and recognizing each pedestrian attribute on each composite feature layer, which includes: and training an attribute positioning and identifying module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer through the following cross entropy loss function:
wherein the content of the first and second substances,
is the ithTraining loss of the feature layer, M is the number of the plurality of pedestrian attributes, ymLabeling the real attribute of the mth pedestrian attribute,
as a result of identifying the m-th pedestrian attribute on the i-th feature layer, γmAnd the weight of the mth pedestrian attribute is shown, sigma is a preset parameter, and the ith characteristic layer is a composite characteristic layer or an initial characteristic layer.
Also as an example in FIG. 2, above, based on the composite feature layer X1Obtaining the positioning identification result of M pedestrian attributes and the real attribute labels of the M pedestrian attributes, and obtaining a composite characteristic layer X based on the cross entropy function1Training loss L1(ii) a Based on composite feature layer X2Obtaining the positioning identification result of M pedestrian attributes and the real attribute labels of the M pedestrian attributes, and obtaining a composite characteristic layer X based on the cross entropy function2Training loss L2(ii) a Based on composite feature layer X3Obtaining the positioning identification result of M pedestrian attributes and the real attribute labels of the M pedestrian attributes, and obtaining a composite characteristic layer X based on the cross entropy function3Training loss L3(ii) a Based on the initial feature layer phi1Initial characteristic layer phi2And initial feature layer phi3Obtaining the global identification result of the M pedestrian attributes and the real attribute labels of the M pedestrian attributes, and obtaining the training loss L based on the cross entropy function4. Finally, the loss L of the attribute positioning recognition system is the sum of training losses of a plurality of feature layers, namely L ═ L1+L2+L3+L4。
In one possible implementation, the method further includes: aiming at any pedestrian attribute, determining a positioning identification result of each composite characteristic layer on the pedestrian attribute in the test image according to an attribute positioning identification module which is used for positioning and identifying the pedestrian attribute on each composite characteristic layer; determining a global identification result of the pedestrian attribute in the test image according to the highest level initial feature layer; and determining the attribute recognition result of the pedestrian attribute in the test image according to the positioning recognition result of the pedestrian attribute in the test image on each composite characteristic layer and the global recognition result of the pedestrian attribute in the test image.
And aiming at the test image, performing feature extraction on the test image according to the convolutional neural network, wherein the extraction position is the same as the feature extraction position of the sample image, and obtaining a plurality of initial feature layers. And performing layer-by-layer feature fusion on the plurality of initial feature layers from top to bottom to obtain a plurality of composite feature layers.
For any composite feature layer, attribute positioning identification can be carried out on the basis of an attribute positioning identification module corresponding to each pedestrian attribute on the composite feature layer, and a positioning identification result of each pedestrian attribute in the test image on the composite feature is obtained. Based on the highest-level initial feature layer, the global identification result of each pedestrian attribute in the test image can be determined, and then the attribute identification result of each pedestrian attribute in the test image is determined by taking the maximum value element by element according to the positioning identification result and the global identification result of each pedestrian attribute in the test image.
For example, for pedestrian property A, composite feature layer X1The obtained positioning identification result is 60 percent (namely the probability of existence of the pedestrian attribute A in the original image is 60 percent), and the composite characteristic layer X2The obtained positioning identification result is 65%, and the characteristic layer X is compounded3The obtained positioning recognition result was 55%, and the initial feature layer phi was obtained3If the obtained global positioning result is 48%, the maximum value of the final attribute identification result of the pedestrian attribute is 65%, and if the attribute identification result is greater than 50%, the original image contains the attribute, and the original image contains the pedestrian attribute a according to the final attribute identification result.
The identification index pair ratio of the pedestrian attribute identification method provided by the disclosure and other four existing pedestrian attribute identification methods (deep mar, GRL, VeSPA, PGDM) is shown in table 1:
TABLE 1
The average accuracy and the F1 value are common indexes for evaluating a pedestrian attribute identification algorithm, and the higher the value is, the higher the precision is; the model size and recognition speed reflect the efficiency of the algorithm, the lower the better. Therefore, compared with the existing pedestrian attribute identification algorithm, the pedestrian attribute identification algorithm can improve the accuracy and efficiency.
Performing feature extraction on the sample image according to the convolutional neural network to obtain a plurality of initial feature layers, wherein the sample image is an image in a preset pedestrian attribute sample set, and the image in the preset pedestrian attribute sample set has a plurality of pedestrian attributes; performing layer-by-layer feature fusion on the plurality of initial feature layers from top to bottom to obtain a plurality of composite feature layers; and determining an attribute positioning and identifying module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer according to the space transformation network and the plurality of composite characteristic layers. The method and the device can locate the corresponding local area of each pedestrian attribute in the image, and then identify the pedestrian attribute based on the local features, thereby improving the accuracy and efficiency of pedestrian attribute identification.
Fig. 5 is a schematic structural diagram of a pedestrian attribute identification device according to an embodiment of the present disclosure. As shown in fig. 5, the apparatus 50 includes:
the feature extraction module 51 is configured to perform feature extraction on the sample image according to the convolutional neural network to obtain a plurality of initial feature layers, where the sample image is an image in a preset pedestrian attribute sample set, and the image included in the preset pedestrian attribute sample set has a plurality of pedestrian attributes;
the feature fusion module 52 is configured to perform feature fusion layer by layer from top to bottom on the plurality of initial feature layers to obtain a plurality of composite feature layers;
and the attribute positioning and identifying module 53 is configured to determine, according to the spatial transform network and the multiple composite feature layers, an attribute positioning and identifying module that performs positioning and identifying on each pedestrian attribute on each composite feature layer.
In one possible implementation, the feature fusion module 52 includes:
the first determining submodule is used for directly determining the highest-level initial characteristic layer as a corresponding composite characteristic layer aiming at the highest-level initial characteristic layer;
and the feature fusion submodule is used for performing feature fusion on the non-highest-level initial feature layer and the composite feature layer corresponding to the previous-level initial feature layer aiming at the non-highest-level initial feature layer to obtain the composite feature layer corresponding to the non-highest-level initial feature layer.
In one possible implementation, the non-highest level initial feature level is φiInitial feature layer phiiThe last level of initial feature layer is phii+1Initial feature layer phii+1Corresponding composite feature layer is Xi+1；
The feature fusion submodule is specifically configured to:
combining the feature layers Xi+1Up-sampling and amplifying to the initial characteristic layer phiiAre the same, and the up-sampled composite feature layer X is obtainedi+1；
The up-sampled composite characteristic layer Xi+1Phi from the initial feature layeriSplicing according to the number of channels to obtain an initial characteristic layer phiiCorresponding composite feature layer XiWherein the characteristic layer X is compoundediThe number of channels is the initial feature layer phii+1And initial feature layer phiiThe sum of the number of channels.
In one possible implementation, the attribute positioning and identifying module 53 includes:
the second determining submodule is used for determining a positioning and identifying result of the pedestrian attribute on each composite characteristic layer according to the attribute positioning and identifying module which positions and identifies the pedestrian attribute on each composite characteristic layer aiming at any pedestrian attribute;
the third determining submodule is used for determining a global recognition result of the pedestrian attribute according to the highest level initial feature layer aiming at any pedestrian attribute;
and the fourth determining submodule is used for training the attribute positioning and identifying module for positioning and identifying the attributes of each pedestrian on each composite characteristic layer according to the positioning and identifying result of each pedestrian attribute on each composite characteristic layer, the global identifying result of each pedestrian attribute and the real attribute label of each pedestrian attribute in the sample image, so as to obtain the attribute positioning and identifying module for positioning and identifying the attributes of each pedestrian on each composite characteristic layer.
In a possible implementation manner, the second determining submodule is specifically configured to:
for any composite characteristic layer, an attribute positioning and identifying module for positioning and identifying the pedestrian attribute on the composite characteristic layer determines the positioning and identifying result of the pedestrian attribute on the composite characteristic layer through the following steps:
passing the composite characteristic layer through a first full-connection layer to obtain a transformation parameter sx、sy、txAnd tyWherein the parameter s is transformedxFor scaling the transformation parameters in the horizontal direction, transformation parameters syFor scaling the transformation parameter in the vertical direction, transformation parameter txFor translation transformation parameters in the horizontal direction, transformation parameter tyTranslation transformation parameters in the vertical direction;
according to transformation parameters sx、sy、txAnd tyDetermining local features corresponding to the pedestrian attributes in the composite feature layer;
and passing the local features corresponding to the pedestrian attributes through a second full-connection layer to obtain a positioning identification result of the pedestrian attributes on the composite feature layer.
In one possible implementation, the apparatus 50 further includes: a feature calibration submodule;
the feature calibration submodule is specifically configured to: before passing the composite feature layer through the first fully connected layer, performing the following feature calibration process on the composite feature layer:
sequentially passing the composite characteristic layer through a global average pooling layer, a 1 × 1 convolution layer, a ReLU active layer, a 1 × 1 convolution layer and a Sigmoid active layer to obtain a first calibration vector;
multiplying the composite feature layer and the first calibration vector one by one according to the channel to obtain a second calibration vector;
and adding the composite characteristic layer and the second calibration vector element by element to obtain a calibrated composite characteristic layer.
In a possible implementation manner, the second determining submodule is specifically configured to:
according to transformation parameters sx、sy、txAnd tyDetermining a rectangular bounding box in the composite feature layer;
and extracting the features in the rectangular bounding box from the composite feature layer, and determining the features as local features corresponding to the attributes of the pedestrians.
In a possible implementation manner, the fourth determining submodule is specifically configured to:
and training an attribute positioning and identifying module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer through the following cross entropy loss function:
wherein the content of the first and second substances,
for the training loss of the ith feature layer, M is the number of multiple pedestrian attributes, ymLabeling the real attribute of the mth pedestrian attribute,
as a result of identifying the m-th pedestrian attribute on the i-th feature layer, γmAnd the weight of the mth pedestrian attribute is shown, sigma is a preset parameter, and the ith characteristic layer is a composite characteristic layer or an initial characteristic layer.
In one possible implementation, the apparatus 50 further includes: the test module is specifically configured to:
aiming at any pedestrian attribute, determining a positioning identification result of each composite characteristic layer on the pedestrian attribute in the test image according to an attribute positioning identification module which is used for positioning and identifying the pedestrian attribute on each composite characteristic layer;
determining a global identification result of the pedestrian attribute in the test image according to the highest level initial feature layer;
and determining the attribute recognition result of the pedestrian attribute in the test image according to the positioning recognition result of the pedestrian attribute in the test image on each composite characteristic layer and the global recognition result of the pedestrian attribute in the test image.
The apparatus 50 provided in the present disclosure can implement each step in the method embodiments shown in fig. 1 to fig. 3, and implement the same technical effect, and for avoiding repetition, details are not described here again.
Having described embodiments of the present disclosure, the foregoing description is intended to be exemplary, not exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terms used herein were chosen in order to best explain the principles of the embodiments, the practical application, or technical improvements to the techniques in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.
Claims (10)
1. A pedestrian attribute identification method is characterized by comprising the following steps:
performing feature extraction on a sample image according to a convolutional neural network to obtain a plurality of initial feature layers, wherein the sample image is an image in a preset pedestrian attribute sample set, and the image in the preset pedestrian attribute sample set has a plurality of pedestrian attributes;
performing layer-by-layer feature fusion on the plurality of initial feature layers from top to bottom to obtain a plurality of composite feature layers;
and determining an attribute positioning and identifying module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer according to the space transformation network and the plurality of composite characteristic layers.
2. The method of claim 1, wherein performing top-down layer-by-layer feature fusion on the plurality of initial feature layers to obtain a plurality of composite feature layers comprises:
directly determining the highest-level initial feature layer as a corresponding composite feature layer aiming at the highest-level initial feature layer;
and performing feature fusion on the non-highest level initial feature layer and a composite feature layer corresponding to a previous level initial feature layer aiming at the non-highest level initial feature layer to obtain the composite feature layer corresponding to the non-highest level initial feature layer.
3. The method of claim 2, wherein the non-highest level initial feature level is φiSaid initial characteristic layer phiiThe last level of initial feature layer is phii+1Said initial characteristic layer phii+1Corresponding composite feature layer is Xi+1；
Aiming at a non-highest level initial feature layer, performing feature fusion on the non-highest level initial feature layer and a composite feature layer corresponding to a previous level initial feature layer to obtain the composite feature layer corresponding to the non-highest level initial feature layer, wherein the method comprises the following steps:
combining the composite feature layer Xi+1Up-sampling and amplifying to the initial characteristic layer phiiAre the same, and the up-sampled composite feature layer X is obtainedi+1；
The up-sampled composite characteristic layer X is processedi+1With respect to the initial feature layer phiiSplicing according to the number of channels to obtain the initial characteristic layer phiiCorresponding composite feature layer XiWherein the composite feature layer XiIs the initial feature layer phii+1And the initial feature layer phiiThe sum of the number of channels.
4. The method of claim 1, wherein determining an attribute location identification module for location identification of each pedestrian attribute on each composite feature layer according to the spatial transformation network and the plurality of composite feature layers comprises:
for any pedestrian attribute, determining a positioning identification result of the pedestrian attribute on each composite characteristic layer according to an attribute positioning identification module which carries out positioning identification on the pedestrian attribute on each composite characteristic layer;
aiming at any pedestrian attribute, determining a global recognition result of the pedestrian attribute according to the highest level initial feature layer;
according to the positioning recognition result of each pedestrian attribute on each composite feature layer, the global recognition result of each pedestrian attribute and the real attribute mark of each pedestrian attribute in the sample image, training an attribute positioning recognition module for positioning and recognizing each pedestrian attribute on each composite feature layer to obtain an attribute positioning recognition module for positioning and recognizing each pedestrian attribute on each composite feature layer.
5. The method according to claim 4, wherein for any pedestrian attribute, determining a positioning identification result of the pedestrian attribute on each composite feature layer according to an attribute positioning identification module for positioning identification of the pedestrian attribute on each composite feature layer comprises:
for any composite characteristic layer, an attribute positioning and identifying module for positioning and identifying the pedestrian attribute on the composite characteristic layer determines a positioning and identifying result of the pedestrian attribute on the composite characteristic layer through the following steps:
passing the composite characteristic layer through a first full-connection layer to obtain a transformation parameter sx、sy、txAnd tyWherein the parameter s is transformedxFor scaling the transformation parameters in the horizontal direction, transformation parameters syFor scaling the transformation parameter in the vertical direction, transformation parameter txFor translation transformation parameters in the horizontal direction, transformation parameter tyTranslation transformation parameters in the vertical direction;
according to said transformation parameter sx、sy、txAnd tyDetermining local features corresponding to the pedestrian attributes in the composite feature layer;
and enabling the local features corresponding to the pedestrian attributes to pass through a second full-connection layer, and obtaining a positioning identification result of the pedestrian attributes on the composite feature layer.
6. The method of claim 5, further comprising:
performing the following feature calibration process on the composite feature layer before passing the composite feature layer through a first fully connected layer:
sequentially passing the composite feature layer through a global average pooling layer, a 1 × 1 convolution layer, a ReLU active layer, a 1 × 1 convolution layer and a Sigmoid active layer to obtain a first calibration vector;
multiplying the composite feature layer and the first calibration vector one by one according to channels to obtain a second calibration vector;
and adding the composite characteristic layer and the second calibration vector element by element to obtain a calibrated composite characteristic layer.
7. Method according to claim 5, characterized in that, according to said transformation parameter sx、sy、txAnd tyDetermining local features corresponding to the pedestrian attributes in the composite feature layer, including:
according to said transformation parameter sx、sy、txAnd tyDetermining a rectangular bounding box in the composite feature layer;
and extracting the features in the rectangular bounding box from the composite feature layer, and determining the features as local features corresponding to the pedestrian attributes.
8. The method according to claim 4, wherein an attribute locating and identifying module for locating and identifying each pedestrian attribute on each composite feature layer is trained according to the locating and identifying result of each pedestrian attribute on each composite feature layer, the global identifying result of each pedestrian attribute, and the real attribute label of each pedestrian attribute in the sample image, so as to obtain an attribute locating and identifying module for locating and identifying each pedestrian attribute on each composite feature layer, and the method comprises the following steps:
and training an attribute positioning and identifying module for positioning and identifying each pedestrian attribute on each composite characteristic layer through the following cross entropy loss functions:
wherein the content of the first and second substances,
the training loss of the ith feature layer is M is the number of the multiple pedestrian attributes, ymLabeling the real attribute of the mth pedestrian attribute,
as a result of identifying the m-th pedestrian attribute on the i-th feature layer, γmAnd the weight of the mth pedestrian attribute is shown, sigma is a preset parameter, and the ith characteristic layer is a composite characteristic layer or an initial characteristic layer.
9. The method of claim 1, further comprising:
for any pedestrian attribute, determining a positioning identification result of the pedestrian attribute in the test image on each composite feature layer according to an attribute positioning identification module which positions and identifies the pedestrian attribute on each composite feature layer;
determining a global identification result of the pedestrian attribute in the test image according to the highest level initial feature layer;
and determining the attribute recognition result of the pedestrian attribute in the test image according to the positioning recognition result of the pedestrian attribute in the test image on each composite feature layer and the global recognition result of the pedestrian attribute in the test image.
10. A pedestrian property identification device characterized by comprising:
the characteristic extraction module is used for extracting characteristics of a sample image according to a convolutional neural network to obtain a plurality of initial characteristic layers, wherein the sample image is an image in a preset pedestrian attribute sample set, and the image in the preset pedestrian attribute sample set has a plurality of pedestrian attributes;
the characteristic fusion module is used for carrying out layer-by-layer characteristic fusion on the plurality of initial characteristic layers from top to bottom to obtain a plurality of composite characteristic layers;
and the attribute positioning and identifying module is used for determining an attribute positioning and identifying module for positioning and identifying the attribute of each pedestrian on each composite characteristic layer according to the space transformation network and the plurality of composite characteristic layers.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910943815.6A CN110705474B (en) | 2019-09-30 | 2019-09-30 | Pedestrian attribute identification method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910943815.6A CN110705474B (en) | 2019-09-30 | 2019-09-30 | Pedestrian attribute identification method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110705474A true CN110705474A (en) | 2020-01-17 |
| CN110705474B CN110705474B (en) | 2022-05-03 |
Family
ID=69197731
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910943815.6A Active CN110705474B (en) | 2019-09-30 | 2019-09-30 | Pedestrian attribute identification method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110705474B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111401343A (en) * | 2020-06-04 | 2020-07-10 | 北京金山云网络技术有限公司 | Method for identifying attributes of people in image and training method and device for identification model |
| WO2023082196A1 (en) * | 2021-11-12 | 2023-05-19 | 京东方科技集团股份有限公司 | Pedestrian attribute recognition system and training method therefor, and pedestrian attribute recognition method |
Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160042252A1 (en) * | 2014-08-05 | 2016-02-11 | Sri International | Multi-Dimensional Realization of Visual Content of an Image Collection |
| US20170293763A1 (en) * | 2013-03-15 | 2017-10-12 | Advanced Elemental Technologies, Inc. | Methods and systems for secure and reliable identity-based computing |
| CN107977656A (en) * | 2017-12-26 | 2018-05-01 | 北京大学 | A kind of pedestrian recognition methods and system again |
| CN108052894A (en) * | 2017-12-11 | 2018-05-18 | 北京飞搜科技有限公司 | More attribute recognition approaches, equipment, medium and the neutral net of a kind of target object |
| CN108921054A (en) * | 2018-06-15 | 2018-11-30 | 华中科技大学 | A kind of more attribute recognition approaches of pedestrian based on semantic segmentation |
| CN108921051A (en) * | 2018-06-15 | 2018-11-30 | 清华大学 | Pedestrian's Attribute Recognition network and technology based on Recognition with Recurrent Neural Network attention model |
| CN109034044A (en) * | 2018-06-14 | 2018-12-18 | 天津师范大学 | A kind of pedestrian's recognition methods again based on fusion convolutional neural networks |
| CN109214366A (en) * | 2018-10-24 | 2019-01-15 | 北京旷视科技有限公司 | Localized target recognition methods, apparatus and system again |
| US20190042862A1 (en) * | 2017-08-01 | 2019-02-07 | Denso Corporation | Vehicle safety determination apparatus, method, and computer-readable storage medium |
| CN109598186A (en) * | 2018-10-12 | 2019-04-09 | 高新兴科技集团股份有限公司 | A kind of pedestrian's attribute recognition approach based on multitask deep learning |
| CN109902548A (en) * | 2018-07-20 | 2019-06-18 | 华为技术有限公司 | A kind of object properties recognition methods, calculates equipment and system at device |
| CN109948709A (en) * | 2019-03-21 | 2019-06-28 | 南京斯玛唯得智能技术有限公司 | A kind of multitask Attribute Recognition system of target object |
| CN110046553A (en) * | 2019-03-21 | 2019-07-23 | 华中科技大学 | A kind of pedestrian weight identification model, method and system merging attributive character |
| CN110046550A (en) * | 2019-03-14 | 2019-07-23 | 中山大学 | Pedestrian's Attribute Recognition system and method based on multilayer feature study |
| CN110188596A (en) * | 2019-01-04 | 2019-08-30 | 北京大学 | Monitor video pedestrian real-time detection, Attribute Recognition and tracking and system based on deep learning |
-
2019
- 2019-09-30 CN CN201910943815.6A patent/CN110705474B/en active Active
Patent Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170293763A1 (en) * | 2013-03-15 | 2017-10-12 | Advanced Elemental Technologies, Inc. | Methods and systems for secure and reliable identity-based computing |
| US20160042252A1 (en) * | 2014-08-05 | 2016-02-11 | Sri International | Multi-Dimensional Realization of Visual Content of an Image Collection |
| US20190042862A1 (en) * | 2017-08-01 | 2019-02-07 | Denso Corporation | Vehicle safety determination apparatus, method, and computer-readable storage medium |
| CN108052894A (en) * | 2017-12-11 | 2018-05-18 | 北京飞搜科技有限公司 | More attribute recognition approaches, equipment, medium and the neutral net of a kind of target object |
| CN107977656A (en) * | 2017-12-26 | 2018-05-01 | 北京大学 | A kind of pedestrian recognition methods and system again |
| CN109034044A (en) * | 2018-06-14 | 2018-12-18 | 天津师范大学 | A kind of pedestrian's recognition methods again based on fusion convolutional neural networks |
| CN108921051A (en) * | 2018-06-15 | 2018-11-30 | 清华大学 | Pedestrian's Attribute Recognition network and technology based on Recognition with Recurrent Neural Network attention model |
| CN108921054A (en) * | 2018-06-15 | 2018-11-30 | 华中科技大学 | A kind of more attribute recognition approaches of pedestrian based on semantic segmentation |
| CN109902548A (en) * | 2018-07-20 | 2019-06-18 | 华为技术有限公司 | A kind of object properties recognition methods, calculates equipment and system at device |
| CN109598186A (en) * | 2018-10-12 | 2019-04-09 | 高新兴科技集团股份有限公司 | A kind of pedestrian's attribute recognition approach based on multitask deep learning |
| CN109214366A (en) * | 2018-10-24 | 2019-01-15 | 北京旷视科技有限公司 | Localized target recognition methods, apparatus and system again |
| CN110188596A (en) * | 2019-01-04 | 2019-08-30 | 北京大学 | Monitor video pedestrian real-time detection, Attribute Recognition and tracking and system based on deep learning |
| CN110046550A (en) * | 2019-03-14 | 2019-07-23 | 中山大学 | Pedestrian's Attribute Recognition system and method based on multilayer feature study |
| CN109948709A (en) * | 2019-03-21 | 2019-06-28 | 南京斯玛唯得智能技术有限公司 | A kind of multitask Attribute Recognition system of target object |
| CN110046553A (en) * | 2019-03-21 | 2019-07-23 | 华中科技大学 | A kind of pedestrian weight identification model, method and system merging attributive character |
Non-Patent Citations (4)
| Title |
|---|
| DANGWEI LI ET AL.: ""Pose Guided Deep Model for Pedestrian Attribute Recognition in Surveillance Scenarios"", 《2018IEEE INTERNATIONAL CONFERENCE ON MULTIMEDIA AND EXPO》 * |
| YIQIANG CHEN ET AL.: ""Pedestrian attribute recognition with part-based CNN and combined feature representations"", 《HAL》 * |
| ZHONG JI ET AL.: ""Image-attribute reciprocally guided attention network for pedestrian attribute recognition"", 《PATTERN RECOGNITION LETTERS》 * |
| 郭志影: ""基于深度学习的室外监控场景下行人属性识别"", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111401343A (en) * | 2020-06-04 | 2020-07-10 | 北京金山云网络技术有限公司 | Method for identifying attributes of people in image and training method and device for identification model |
| CN111401343B (en) * | 2020-06-04 | 2021-04-30 | 北京金山云网络技术有限公司 | Method for identifying attributes of people in image and training method and device for identification model |
| WO2023082196A1 (en) * | 2021-11-12 | 2023-05-19 | 京东方科技集团股份有限公司 | Pedestrian attribute recognition system and training method therefor, and pedestrian attribute recognition method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110705474B (en) | 2022-05-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110348319B (en) | Face anti-counterfeiting method based on face depth information and edge image fusion | |
| CN109471945B (en) | Deep learning-based medical text classification method and device and storage medium | |
| CN110287960A (en) | The detection recognition method of curve text in natural scene image | |
| CN109934293A (en) | Image-recognizing method, device, medium and obscure perception convolutional neural networks | |
| CN109102024B (en) | Hierarchical semantic embedded model for fine object recognition and implementation method thereof | |
| CN110782420A (en) | Small target feature representation enhancement method based on deep learning | |
| CN111563502A (en) | Image text recognition method and device, electronic equipment and computer storage medium | |
| CN102629279B (en) | Method for searching and reordering images or videos | |
| CN109637664A (en) | A kind of BMI evaluating method, device and computer readable storage medium | |
| CN113034495B (en) | Spine image segmentation method, medium and electronic device | |
| CN114332544B (en) | Image block scoring-based fine-grained image classification method and device | |
| CN110705474B (en) | Pedestrian attribute identification method and device | |
| CN113239820B (en) | Pedestrian attribute identification method and system based on attribute positioning and association | |
| CN113936195B (en) | Sensitive image recognition model training method and device and electronic equipment | |
| CN111027481A (en) | Behavior analysis method and device based on human body key point detection | |
| CN113657409A (en) | Vehicle loss detection method, device, electronic device and storage medium | |
| CN111984772A (en) | Medical image question-answering method and system based on deep learning | |
| CN112115879B (en) | Self-supervision pedestrian re-identification method and system with shielding sensitivity | |
| CN111985532B (en) | Scene-level context-aware emotion recognition deep network method | |
| CN110674685A (en) | Human body analytic segmentation model and method based on edge information enhancement | |
| CN111222530A (en) | Fine-grained image classification method, system, device and storage medium | |
| CN107291774A (en) | Error sample recognition methods and device | |
| CN105844605B (en) | Based on the human face portrait synthetic method adaptively indicated | |
| CN105740903B (en) | More attribute recognition approaches and device | |
| CN113762257A (en) | Identification method and device for marks in makeup brand images |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |