Non-motor vehicle rider attribute recognition method based on multi-label image recognition

Abstract

The application discloses a non-motor vehicle rider attribute recognition method based on multi-label image recognition and belongs to the field of computer vision. In view of the non-motor vehicle rider attributes expected to be recognized by an actual application system, there are attributes related to a large range of image regions and attributes only related to small image regions, the recognition method of the application adopts a multi-branch network structure, attributes related to small image regions, attributes related to medium-sized image regions and attributes related to large range image regions are respectively recognized by different branch networks; further, a feature extraction module and a feature pyramid module of the application extract feature maps of multiple scales and different size receptive fields, and the feature maps of appropriate size receptive fields are selected and input into the above various branch networks, so that the various branch networks all have good recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision, and in particular to a non-motor vehicle rider attribute recognition method based on multi-label image recognition. BACKGROUND

[0002] With the continuous deepening of smart city construction, the motor vehicle lane and the non-motor vehicle lane, the industrial park and the residential area in the city are gradually built with a multi-modal, multi-dimensional sensor system mainly composed of video cameras. In the face of massive video data, relying on visual observation cannot meet the needs of event warning, target retrieval and other applications, thus putting forward an urgent demand for video structured analysis technology. At the same time, electric bicycles and bicycles as representatives of non-motor vehicles have always been the main means of transportation for urban personal travel. It is undoubtedly of great significance to predict the attributes of the rider such as gender, whether to wear a helmet, whether to carry a passenger, and the color of the clothes by using artificial intelligence algorithm based on video image as input, to protect the life and property safety of road traffic participants, to assist the traffic management department in maintaining traffic order, and to constrain people to abide by traffic rules.

[0003] The non-motor vehicle rider attributes expected to be recognized by the actual application system include not only attributes related to a larger image area such as non-motor vehicle type and rider gender, but also attributes related to a smaller image area such as whether to wear a helmet and whether to carry a basket. However, how to simultaneously realize attribute detection of different range scales in the image is a technical problem to be solved at present. SUMMARY

[0004] The present application aims to solve the problems existing in the prior art and provides a non-motor vehicle rider attribute recognition method based on multi-label image recognition.

[0005] To achieve the above-mentioned application purpose, the technical solution adopted by the present application is as follows:

[0006] A non-motor vehicle rider attribute recognition method based on multi-label image recognition, characterized in that the method comprises:

[0007] S1, a deep neural network model for realizing multi-label image recognition is constructed, and the deep neural network model is composed of a feature extraction module, a feature pyramid module and a multi-label recognition module;

[0008] The feature extraction module adopts the first three convolutional layer groups of the ConvNeXt network, and a spatial attention module is additionally attached after each convolutional layer group; the image to be recognized input into the deep neural network model is a color image in RGB color space Where W and H are the width and height of the image, respectively. First, x is converted into an image x′ in the HSV color space. Then, the chroma component image and saturation component image of x′ are concatenated with x in the channel direction to form a W×H×5 image, which is used as the input of the first-level convolutional layer group in the feature extraction module. The other two levels of convolutional layer groups are used as the input of the output feature maps of the previous level convolutional layer group and the spatial attention module, respectively.

[0009] The feature pyramid module is used to further process the feature maps output by the convolutional layers and spatial attention module in the feature extraction module, forming feature maps of different sizes. and Feature maps P1, P2, and P3;

[0010] The multi-label recognition module comprises three branch networks. The first branch network takes feature map P1 as input and is used to identify attributes involving image regions within a first scale range. The second branch network takes P2 as input and is used to identify attributes involving image regions within a second scale range. The third branch network takes P3 as input and is used to identify attributes involving image regions within a third scale range. The scales of the first, second, and third scale ranges increase sequentially. For any i-th branch network in the multi-label recognition module, i = 1, 2, 3, the first step is to identify attributes involving image regions with N channels. p Input feature map P i After performing adaptive average pooling, an N is formed. p A vector of size ×1 is generated, and then this vector is fed into a fully connected layer, followed by a sigmoid activation operation to output a K-value. i A vector of ×1 is used as the final output of the i-th branch network, where K i It equals the number of attributes identified by the i-th branch network; the vectors output by the three branch networks are concatenated into a K×1 vector and used as the final predicted label output by the multi-label recognition module. The k-th element in the vector represents the probability that the input image has the k-th attribute, given in probability form, where K = K1 + K2 + K3 is the total number of attribute categories.

[0011] S2. Train the deep neural network model constructed in S1. The parameters of each network layer of the feature extraction module, feature pyramid module and multi-label recognition module are determined by learning.

[0012] S3, For the image to be recognized The input is then fed into the deep neural network model trained in S2, and the multi-label recognition module outputs the corresponding predicted label. If vector The j-th element If the value is greater than the preset threshold τ, then the input image is determined to have the j-th attribute.

[0013] As preferred, in the feature extraction module, a spatial attention module is additionally attached after any i-th convolutional layer group, i = 1, 2, 3, and the internal data processing process is as follows:

[0014] For any i-th convolutional layer group, the output feature map of the last residual block in the convolutional layer group is where w, h and N c are the height, width and channel number of the feature map, respectively; the spatial attention module attached to the i-th convolutional layer group contains four convolutional layers, the first three convolutional layers each contain N c / 16 convolutional kernels, and the fourth convolutional layer contains 1 convolutional kernel; the convolutional kernel size of the first and fourth convolutional layers is 1x1, and the convolutional step is 1; the second and third convolutional layers use a convolution kernel size of 3x3, a dilation rate of 4, a padding size of 4, and a convolution step of 1; the first three convolutional layers are each followed by a batch normalization operation and a ReLU activation operation; the output of the fourth convolutional layer is expanded into a feature map C p of the same size as z' in the channel direction after being calculated by the Sigmoid function, and finally z'x(1+C p ) is taken as the output C i of the i-th convolutional layer group and the spatial attention module.

[0015] As preferred, the feature pyramid module further processes the feature maps output by the convolutional layer groups at each level in the feature extraction module, specifically including:

[0016] The topmost layer of the feature pyramid module takes the feature map C3 output by the spatial attention module after the third convolutional layer group of the ConvNeXt network as input, where the size of C3 is The input feature map C3 first passes through a convolutional layer containing 256 convolutional kernels with a size of 1x1 and a step of 1 to form a feature map Then the feature map passes through a convolutional layer containing 256 convolutional kernels with a size of 3x3 and a step of 1 to form a feature map P3.

[0017] The size of the feature map C2 output by the spatial attention module after the second convolutional layer group of the ConvNeXt network is The feature map C2 first passes through a convolutional layer containing 256 convolutional kernels with a size of 1x1 and a step of 1 to form a feature map which is element-wise summed with the feature map generated by the transposed convolution up-sampling to generate a feature map Then the feature map passes through a convolutional layer containing 256 convolutional kernels with a size of 3x3 and a step of 1 to form a feature map P2.

[0018] The size of the feature map C1 output by the spatial attention module connected after the first convolutional layer group of the ConvNeXt network is The feature map C1 is first passed through a convolutional layer containing 256 convolutional kernels with a size of 1x1 and a step of 1, and the formed feature map is The feature map after the transposed convolution upsampling is element-wise summed to generate the feature map Then the feature map is passed through a convolutional layer containing 256 convolutional kernels with a size of 3x3 and a step of 1, and the formed feature map P1 is.

[0019] As a preferred, the attributes related to the image region of the first scale range include whether to wear a helmet, helmet color, whether to wear a mask, whether to have an article, backpack position, lower garment type, and lower garment color.

[0020] As a preferred, the attributes related to the image region of the second scale range include upper garment type, upper garment color, and vehicle windshield color.

[0021] As a preferred, the attributes related to the image region of the third scale range include cyclist gender, whether to carry a passenger, and non-motor vehicle type.

[0022] As a preferred, in the S2, the training sample set for training the deep neural network model is where N is the number of samples in the set; x l is the lth sample image; y l is the label data annotated for the sample image x l , which is a Kx1 vector, and each element in the vector corresponds to an attribute. If the sample image x l is annotated with the jth attribute, then the element y l,j corresponding to the attribute is 1, otherwise y l,j = 0;

[0023] Taking the sample image x l as the input of the deep neural network model, the vector formed by splicing the outputs of the three branch networks of the multi-label recognition module is The loss function corresponding to the sample image x l is calculated as follows:

[0024]

[0025] In the learning phase, the sample images in the training sample set are input into the model, and the parameters of each network level are updated and adjusted in a backpropagation manner to minimize the sum of the loss functions of all samples, until the model converges.

[0026] The present application has the following beneficial effects relative to the prior art:

[0027] For the non-motor vehicle rider attributes expected to be identified by the actual application system, there are attributes related to a larger range of image regions and attributes only involving smaller image regions. The non-motor vehicle rider attribute identification method based on multi-label image recognition provided by the present application adopts a multi-branch network structure, and attributes involving smaller image regions, attributes involving medium-sized image regions and attributes involving larger range image regions are identified by different branch networks respectively. Further, the feature extraction module and the feature pyramid module of the present application extract feature maps of multiple scales and different size receptive fields, and select feature maps of appropriate size receptive fields to be input into the above-mentioned respective branch networks, thereby ensuring that each branch network has good recognition accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A schematic diagram of the non-motor vehicle rider attribute identification method based on multi-label image recognition of the present application;

[0029] Figure 2 A schematic diagram of the residual block structure in the convolution layer group of the embodiment of the present application;

[0030] Figure 3 A schematic diagram of the spatial attention module in the convolution layer group of the embodiment of the present application;

[0031] Figure 4 A schematic diagram of part of the test set samples;

[0032] Figure 5 The statistical recognition accuracy results. DETAILED DESCRIPTION

[0033] The present application will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the essence of the present application.

[0034] As Figure 1 shown, the present application provides a non-motor vehicle rider attribute identification method based on multi-label image recognition, and the specific technical solutions are as follows:

[0035] (1) A deep neural network model for realizing multi-label image recognition is constructed, which is composed of a feature extraction module, a feature pyramid module and a multi-label identification module.

[0036] The feature extraction module adopts the first three convolution layer groups of the ConvNeXt network, and a spatial attention module is additionally attached after each convolution layer group. Let the image to be identified input into the deep neural network model be a color image in the RGB color space where W and H are the width and height of the image respectively, x is first converted into an image in HSV color space, denoted as x', and the chroma component image and the saturation component image of x' are spliced with x in the channel direction to form a WxHx5 image as the input of the first convolution layer group in the feature extraction module; secondly, the remaining two convolution layer groups take the output feature maps of the previous convolution layer group and the spatial attention module as the input respectively. Specifically, the second convolution layer group takes the output feature maps of the first convolution layer group and the spatial attention module as the input, and the third convolution layer group takes the output feature maps of the second convolution layer group and the spatial attention module as the input.

[0037] Further, a preferred embodiment of the present application takes the aspect ratio of the input image x to be 1:2, i.e. H = 2W.

[0038] The specific structure of the ConvNeXt network belongs to the prior art, and for the convenience of description and understanding, it is briefly described as follows. Referring to Figure 1 , in the first three convolution layer groups of the ConvNeXt network, each convolution layer group contains a down-sampling convolution layer and a plurality of residual blocks, the down-sampling convolution layer in the first convolution layer group contains 128 convolution kernels with a size of 4x4, and the convolution step is 4; the down-sampling convolution layers in the second and third convolution layer groups contain 256 and 512 convolution kernels with a size of 2x2 respectively, and the convolution step is 2; the down-sampling convolution layer in the first convolution layer group is followed by a channel direction layer normalization operation, and the down-sampling layers in the second and third convolution layer groups are each preceded by a channel direction layer normalization operation.

[0039] In a preferred embodiment of the present application, the number of residual blocks in the first, second and third convolution layer groups of the ConvNeXt network used in the above feature extraction module is 3, 3 and 27 respectively. All residual blocks in the same convolution layer group have the same structure, referring to Figure 2 , according to the data flow direction during forward propagation, the input data first passes through a convolution layer using a grouped convolution operation, which contains N c convolution kernels with a size of 7x7, the convolution step is 1, the padding size is 3, and the number of groups is N c , wherein in the first, second and third convolution layer groups, the value of N c is 128, 256 and 512 respectively; the above convolution layer using a grouped convolution operation is followed by a layer normalization operation applied in the channel direction; then two are sequentially connected, each containing 4N c and N cThe convolutional layer of the convolution kernel, all the convolution kernels have a size of 1*1, the convolution step is 1, and a GELU activation operation is connected after the first convolution layer of the two sequentially connected convolution layers; the input feature map of the residual block is z, the output after the operation of each level of the residual block is f(z), and the entire residual block takes z' = z + f(z) as the output.

[0040] The feature extraction module is additionally provided with a spatial attention module after each of the first three convolution layer groups of the ConvNeXt network, so that each level is composed of a convolution layer group and a spatial attention module, Figure 3 The structural diagram of the spatial attention module is shown. Figure 3 The internal data processing process of any i-th convolution layer group and spatial attention module, i = 1, 2, 3, is as follows:

[0041] For any i-th convolution layer group, the output feature map of the last residual block in the convolution layer group is where w, h and N c are the height, width and channel number of the feature map respectively, the spatial attention module additionally provided after the i-th convolution layer group contains four convolution layers, the first three convolution layers each contain N c / 16 convolution kernels, and the fourth convolution layer contains 1 convolution kernel; the convolution kernel size of the first and fourth convolution layers is 1*1, and the convolution step is 1; the second and third convolution layers adopt a dilated convolution with a convolution kernel size of 3*3, a dilation rate of 4, a padding size of 4 and a convolution step of 1; the first three convolution layers are each followed by a batch normalization operation and a ReLU activation operation; the output of the fourth convolution layer is expanded into a feature map with the same size as z' in the channel direction after being calculated by a Sigmoid function, denoted as C p , and finally, z' x (1 + C p ) is taken as the output C i of the i-th convolution layer group and spatial attention module.

[0042] When the feature extraction module extracts the feature map at each level, the feature pyramid module further processes the feature map output by the feature extraction module at each level, forming feature maps with sizes of and , respectively, denoted as P1, P2 and P3, where W and H are the width and height of the image, respectively.

[0043] In a preferred embodiment of the present application, the feature pyramid module further processes the feature map output by the feature extraction module at each level, and the specific manner of forming the feature maps P1, P2 and P3 is as follows:

[0044] The topmost layer of the feature pyramid module takes the feature map C3 output by the spatial attention module after the third convolutional layer group of the ConvNeXt network as input, where the size of C3 is The input feature map C3 is first passed through a convolutional layer containing 256 1x1 convolutional kernels to form a feature map Then the feature map is passed through a convolutional layer containing 256 3x3 convolutional kernels to form a feature map P3; the convolutional step size of both convolutional layers in this step is 1.

[0045] The feature map C2 output by the spatial attention module after the second convolutional layer group of the ConvNeXt network has a size of This feature map C2 is first passed through a convolutional layer containing 256 1x1 convolutional kernels to form a feature map The feature map after transposed convolution up-sampling is subjected to element-wise summation operation to generate a feature map Then the feature map is passed through a convolutional layer containing 256 3x3 convolutional kernels to form a feature map P2; the convolutional step size of both convolutional layers in this step is 1.

[0046] The feature map C1 output by the spatial attention module after the first convolutional layer group of the ConvNeXt network has a size of This feature map C1 is first passed through a convolutional layer containing 256 1x1 convolutional kernels to form a feature map The feature map after transposed convolution up-sampling is subjected to element-wise summation operation to generate a feature map Then the feature map is passed through a convolutional layer containing 256 3x3 convolutional kernels to form a feature map P1; the convolutional step size of both convolutional layers in this step is 1.

[0047] The multi-label recognition module contains three branch networks, where the first branch network takes the feature map P1 as input, is used to recognize attributes such as wearing a helmet, helmet color, wearing a mask, backpack position, lower garment type, and lower garment color, which only involve image regions in the first scale range; the second branch network takes P2 as input, is used to recognize attributes such as upper garment type, upper garment color, vehicle windshield color, which involve image regions in the second scale range; and the third branch network takes P3 as input, is used to recognize attributes such as cyclist gender, whether carrying a passenger, non-motor vehicle type, which involve image regions in the third scale range. Among them, the first scale range is a relatively small image region scale, the second scale range is a relatively medium image region scale, and the third scale range is a relatively large image region scale, so the scales of the first scale range, the second scale range, and the third scale range increase in turn.

[0048] Table 1 specifically shows the specific attribute category and contained attribute name identified by each branch network in the embodiment.

[0049] Table 1 Attributes identified by each branch network

[0050]

[0051] The entire deep neural network model finally generates the identification result of the corresponding input image by each branch network. Specifically, let the input feature map of the i-th branch network be P i , i = 1, 2, 3, the channel number of P i is N p , and an adaptive average pooling operation is first performed on P i to form an N p × 1 vector, then the vector is input to a fully connected layer, and then a K i × 1 vector is output as the final output of the i-th branch network through the subsequent Sigmoid activation operation, where K i is equal to the number of attributes identified by the i-th branch network; the vectors output by the three branch networks are spliced into a K × 1 vector as the final output of the multi-label identification module. The k-th element in the vector represents the possibility of the occurrence of the k-th attribute in the input image in the form of probability, where K is equal to the total number of attribute categories, that is: K = K1+ K2+ K3.

[0052] (2) The deep neural network model constructed in (1) is trained, and the parameters of each network level of the feature extraction module, the feature pyramid module and the multi-label identification module are determined in a learning manner. In the embodiment, the training sample set used for model learning is where N is the number of samples in the set, x l is the l-th sample image, y l is the label data labeled for the sample image x l , which is a K × 1 vector, each element of the vector corresponds to an attribute, and the attributes listed in Table 1 are numbered in order. If the sample image x l is labeled with the j-th attribute, then the element y l,j corresponding to the attribute = 1, otherwise y l,j = 0.

[0053] With the sample image x l as the input of the deep neural network model, the vector formed by the outputs of the three branch networks of the multi-label identification module is The loss function corresponding to the sample image x l is calculated as follows:

[0054]

[0055] In the learning stage, the sample images in the training sample set are input into the model, and the parameters of each network level are updated in a reverse propagation manner to minimize the sum of the loss functions of all samples until the model converges.

[0056] (3) In the inference stage, if a certain to-be-recognized image is input into the deep neural network model trained in (2), the corresponding predicted label is output by the multi-label recognition module. If the jth element of the vector is greater than a certain preset threshold τ, it is determined that the input image has the jth attribute, otherwise it does not have the jth attribute. In a preferred embodiment of the present application, τ = 0.6.

[0057] In order to prove the effect of the recognition method proposed in the present application, 15000 samples are collected for model training and testing, and the training set and the test set are divided in a ratio of 1:4, and the two sets have no intersection. Figure 4 Some samples in the test set are shown, Figure 5 The recognition accuracy results obtained by dividing the different training sets and test sets are shown in the figure, where the horizontal coordinate is the attribute category, and the vertical coordinate is the average recognition accuracy mA value of each category corresponding to different attributes. The results show that the present application has high recognition accuracy for attributes related to different size image regions.

[0058] The above only describes the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any modification or replacement made within the spirit and principles of the present application should be covered within the protection scope of the present application.

Claims

1.A method for identifying attributes of a non-motorized vehicle rider based on multi-label image recognition, characterized by, The method comprises: S1, constructing a deep neural network model for realizing multi-label image recognition, the deep neural network model being composed of a feature extraction module, a feature pyramid module and a multi-label recognition module; The feature extraction module adopts the first three convolution layer groups of the ConvNeXt network, and a spatial attention module is additionally arranged after each convolution layer group; the image to be recognized input to the deep neural network model is a color image in an RGB color space where W and H are the width and height of the image respectively, x is first converted into an image x' in the HSV color space, and the hue component image and the saturation component image of x' are spliced with x in the channel direction to form a WxHx5 image as the input of the first convolution layer group in the feature extraction module, and the output feature maps of the spatial attention module and the first convolution layer group are taken as the inputs of the remaining two convolution layer groups respectively; The feature pyramid module is used for further processing of the feature maps output by the spatial attention module and each level of the convolutional layer group in the feature extraction module to form feature maps P1, P2 and P3 with sizes of and respectively. The multi-label recognition module comprises three branch networks, wherein the first branch network takes feature map P1 as input, is used for recognizing attributes related to image regions of a first scale range; the second branch network takes P2 as input, is used for recognizing attributes related to image regions of a second scale range; the third branch network takes P3 as input, is used for recognizing attributes related to image regions of a third scale range; the first scale range, the second scale range and the third scale range are sequentially increased in scale; wherein for any ith branch network in the multi-label recognition module, i = 1, 2, 3, first adaptive average pooling operation is performed on an input feature map P p with a channel number of N i , an N p ×1 vector is formed, then the vector is input to a fully connected layer, and then a K i ×1 vector is output as the final output of the ith branch network through a subsequent Sigmoid activation operation, wherein K i is equal to the number of attributes recognized by the ith branch network; the vectors output by the three branch networks are spliced into a K×1 vector, which is used as the predicted label of the final output of the multi-label recognition module; the kth element in the vector represents the possibility of the input image appearing the kth attribute in the form of probability, wherein K = K1 + K2 + K3 is the total number of attribute categories. S2, training the deep neural network model constructed in S1, parameters of each network level of the feature extraction module, the feature pyramid module and the multi-label recognition module being determined in a learning manner; S3, For the image to be recognized The input is then fed into the deep neural network model trained in S2, and the multi-label recognition module outputs the corresponding predicted label. If vector The j-th element If the value is greater than the preset threshold τ, then the input image is determined to have the j-th attribute. 2.The non-motor vehicle rider attribute identification method based on multi-label image recognition according to claim 1, characterized in that, In the feature extraction module, a spatial attention module is attached after any i-th convolutional layer group, i = 1, 2, 3, and the internal data processing flow is as follows: For any i-th convolutional layer group, the output feature map of the last residual block in the convolutional layer group is where w, h and N c are the height, width and channel number of the feature map respectively; the additional spatial attention module of the i-th convolutional layer group contains four convolutional layers, the first three convolutional layers each contain N c / 16 convolutional kernels, and the fourth convolutional layer contains 1 convolutional kernel; the convolutional kernel size of the first and fourth convolutional layers is 1×1, and the convolutional step is 1; the second and third convolutional layers adopt the atrous convolution with the convolutional kernel size of 3×3, the atrous rate of 4, the padding size of 4, and the convolutional step of 1; the first three convolutional layers are each followed by a batch normalization operation and a ReLU activation operation; the output of the fourth convolutional layer is expanded into a feature map C p of the same size as z′ in the channel direction in a replicated manner after being calculated by a Sigmoid function, and finally z′×(1+C p ) is taken as the output C i of the i-th convolutional layer group and the spatial attention module. 3.The non-motor vehicle rider attribute identification method based on multi-label image recognition according to claim 1, characterized in that, The feature pyramid module further processes feature maps output by each convolutional layer group in the feature extraction module, and specifically comprises: The topmost layer of the feature pyramid module takes as input the feature map C3 output by the spatial attention module after the third convolutional layer group of the ConvNeXt network, where the size of C3 is The input feature map C3 is first passed through a convolutional layer containing 256 convolutional kernels of size 1x1 and stride 1 to form a feature map Then the feature map is passed through a convolutional layer containing 256 convolutional kernels of size 3x3 and stride 1 to form a feature map P3; The size of the feature map C2 output by the spatial attention module connected after the second convolutional layer group of the ConvNeXt network is The feature map C2 is first passed through a convolutional layer containing 256 convolutional kernels with a size of 1×1 and a step of 1, and the formed feature map is The feature map after transposed convolution upsampling is subjected to element-wise summation operation to generate a feature map Then the feature map is passed through a convolutional layer containing 256 convolutional kernels with a size of 3×3 and a step of 1 to form a feature map P2; The size of the feature map C1 output by the spatial attention module connected after the first convolutional layer group of the ConvNeXt network is The feature map C1 is first passed through a convolutional layer containing 256 convolutional kernels with a size of 1×1 and a step of 1, and the formed feature map is The feature map after transposed convolution upsampling is subjected to element-wise summation operation to generate a feature map Then the feature map After passing through a convolutional layer containing 256 convolutional kernels with a size of 3×3 and a step of 1, a feature map P1 is formed. 4.The non-motor vehicle rider attribute identification method based on multi-label image recognition according to claim 1, characterized in that, The attributes related to the image region of the first scale range include whether a helmet is worn, a helmet color, whether a mask is worn, whether an article is placed, a backpack position, a lower garment type and a lower garment color. 5.The non-motor vehicle rider attribute identification method based on multi-label image recognition according to claim 1, characterized in that, The attributes related to the image region of the second scale range include an upper garment type, an upper garment color and a vehicle windshield color. 6.The non-motor vehicle rider attribute identification method based on multi-label image recognition according to claim 1, characterized in that, The attributes related to the image region of the third scale range include a cyclist gender, whether a person is carried and a non-motor vehicle type. 7.The non-motor vehicle rider attribute identification method based on multi-label image recognition according to claim 1, characterized in that, The training sample set for training the deep neural network model in S2 is Where N is the number of samples in the set; x l is the lth sample image; y l is the label data annotated for the sample image x l , which is a Kx1 vector, each element of the vector corresponds to an attribute, if the sample image x l is annotated with the jth attribute, then the element y l,j corresponding to the attribute = 1, otherwise y l,j = 0; Taking a sample image x l as the input of a deep neural network model, a vector formed by splicing the outputs of the three branch networks of the multi-label recognition module is The loss function corresponding to the sample image x l is calculated as follows: In the learning stage, sample images in a training sample set are input into the model, parameters of each network level are updated and adjusted in a reverse propagation manner with the aim of minimizing the sum of loss functions of all samples, and the model converges until the model converges.

Images