A pedestrian detection method and system based on a single-stage anchor-free framework
By employing a pedestrian detection method based on a single-stage anchorless framework and utilizing residual networks and cross-fusion feature pyramids to train the model, the problem of small differences among pedestrians is solved, thereby improving the accuracy and precision of pedestrian detection.
Patent Information
- Application Number
- CN202111025295.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-02
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2041-09-02
Smart Images

Figure CN113723322B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of pedestrian detection, and in particular to a pedestrian detection method and system based on a single-stage anchor-free framework. Background Art
[0002] Pedestrian detection is a classic problem in computer vision that uses computer vision technology to determine whether there are pedestrians in an image or video sequence and to accurately locate them. Pedestrian detection has important applications and different requirements in video surveillance, vehicle assisted driving, human behavior analysis and other fields. For example, a vehicle assisted driving system must be able to protect the personal safety of pedestrians and the driving safety of drivers, and interact with the driver while automatically detecting pedestrians in real time. Pedestrian detection is one of the subtasks in target detection. The task goal is to locate all pedestrians of different sizes in the key frames of an image or video, which places high demands on the accuracy and speed of the model. In addition, this field also has difficulties such as complex background, occlusion problems, and small differences between classes.
[0003] In the existing technology, solving complex background problems often requires obtaining more contextual information or redefining positive and negative samples; occlusion problems mostly require special processing such as adding random occlusion to the target to increase the diversity of the data set or generating additional auxiliary frames; however, there is no good way to solve the problem of small differences between pedestrians. Summary of the invention
[0004] The purpose of the present invention is to provide a pedestrian detection method and system based on a single-stage anchor-free framework, which can solve the problem of small differences between pedestrians to a certain extent, thereby improving the accuracy of pedestrian detection.
[0005] To achieve the above object, the present invention provides the following solutions:
[0006] A pedestrian detection method based on a single-stage anchor-free framework, comprising:
[0007] Obtain a pedestrian data set; the pedestrian data set is a pedestrian image in which pedestrians are framed;
[0008] Preprocessing the pedestrian data set to obtain an atlas; the atlas includes a mask map, a size map and an offset map; the mask map is a map including a pedestrian center point and a Gaussian distribution mask of a pedestrian image; the size map is a map including pedestrian aspect ratio information; the offset map is a map including pedestrian center point offset information; the pedestrian center point is the center point of the pedestrian data set center box;
[0009] Inputting the pedestrian data set into a network model to obtain a third feature map; the network model includes a residual network, a cross-fusion feature pyramid and a detection head connected in sequence; the normalization method in the residual network includes group normalization and batch normalization;
[0010] Calculating the error between the third feature map and each map in the atlas;
[0011] The network model is trained using a back propagation algorithm according to the error to obtain a pedestrian detection model, and the pedestrian detection model is used for pedestrian detection.
[0012] Optionally, before preprocessing the pedestrian dataset to obtain an atlas, the method further includes:
[0013] Performing data enhancement processing on the pedestrian dataset to obtain an enhanced pedestrian dataset.
[0014] Optionally, performing data enhancement processing on the pedestrian dataset to obtain an enhanced pedestrian dataset specifically includes:
[0015] A mirroring operation and a size adjustment operation are sequentially performed on the pedestrian data set, wherein the size adjustment operation is random expansion or random cropping.
[0016] Optionally, the process of determining the mask image is:
[0017] Processing the pedestrian data set according to the pedestrian center point to generate a center point mask;
[0018] Taking the pedestrian center point as the center, performing Gaussian distribution processing on the pedestrian data set to generate a Gaussian mask;
[0019] The center point mask and the Gaussian mask are combined to obtain a mask image.
[0020] Optionally, inputting the pedestrian dataset into a network model to obtain a third feature map specifically includes:
[0021] Inputting the pedestrian data set into a residual network to obtain a first feature map;
[0022] Using a cross-fusion feature pyramid to process the first feature map to obtain a second feature map;
[0023] The second feature map is subjected to a detection head operation to obtain a third feature map.
[0024] A pedestrian detection system based on a single-stage anchor-free framework, comprising:
[0025] A pedestrian image acquisition module is used to obtain a pedestrian data set; the pedestrian data set is a pedestrian image in which pedestrians are framed;
[0026] An atlas determination module is used to preprocess the pedestrian data set to obtain an atlas; the atlas includes a mask map, a size map and an offset map; the mask map is a map including a pedestrian center point and a Gaussian distribution mask of a pedestrian image; the size map is a map including pedestrian aspect ratio information; the offset map is a map including pedestrian center point offset information; the pedestrian center point is the center point of the pedestrian data set center box;
[0027] A third feature map determination module is used to input the pedestrian data set into a network model to obtain a third feature map; the network model includes a residual network, a cross-fusion feature pyramid and a detection head connected in sequence; the normalization method in the residual network includes group normalization and batch normalization;
[0028] An error calculation module, used for calculating the error between the third feature map and each map in the map collection;
[0029] The model determination module is used to train the network model using a back propagation algorithm according to the error to obtain a pedestrian detection model, and the pedestrian detection model is used to perform pedestrian detection.
[0030] Optionally, the pedestrian detection system based on the single-stage anchor-free framework further includes:
[0031] The enhancement module is used to perform data enhancement processing on the pedestrian dataset to obtain an enhanced pedestrian dataset.
[0032] Optionally, the enhancement module specifically includes:
[0033] The enhancement unit is used to sequentially perform a mirroring operation and a size adjustment operation on the pedestrian data set, wherein the size adjustment operation is a random expansion or a random cropping.
[0034] Optionally, the atlas determination module includes:
[0035] A center point mask generating unit, configured to process the pedestrian data set according to the pedestrian center point to generate a center point mask;
[0036] A Gaussian mask generating unit, used for performing Gaussian distribution processing on the pedestrian data set with the pedestrian center point as the center to generate a Gaussian mask;
[0037] The mask image determining unit is used to combine the center point mask and the Gaussian mask to obtain a mask image.
[0038] Optionally, the third feature map determining module includes:
[0039] A first feature map determining unit, configured to input the pedestrian data set into a residual network to obtain a first feature map;
[0040] A second feature map determining unit, configured to process the first feature map by using a cross-fusion feature pyramid to obtain a second feature map;
[0041] The third feature map determining unit is used to perform a detection head operation on the second feature map to obtain a third feature map.
[0042] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the present invention preprocesses a pedestrian data set to obtain an atlas; the pedestrian data set is input into a network model to obtain a third feature map; the network model includes a residual network, a cross-fusion feature pyramid and a detection head connected in sequence; the error between the third feature map and each map in the atlas is calculated; the network model is trained using a back propagation algorithm based on the error to obtain a pedestrian detection model, and the obtained pedestrian detection model can solve the problem of small differences between pedestrians to a certain extent, thereby improving the accuracy of pedestrian detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0044] Figure 1 A flowchart of a pedestrian detection method based on a single-stage anchor-free framework provided by an embodiment of the present invention;
[0045] Figure 2 A network structure diagram of a pedestrian detection method based on a single-stage anchor-free framework provided in an embodiment of the present invention;
[0046] Figure 3 A schematic diagram of a residual structure in a residual network provided by an embodiment of the present invention;
[0047] Figure 4 A schematic diagram of cross-fusion in a cross-fusion feature pyramid provided by an embodiment of the present invention;
[0048] Figure 5 A block diagram of a pedestrian detection system based on a single-stage anchor-free framework provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0051] At present, most excellent detection algorithms are based on anchor point detection algorithms. However, since most anchor point priors need to be designed manually, anchor point detection algorithms have many disadvantages. With the rapid improvement of detection accuracy of anchor-free detection algorithms, anchor-free detection algorithms have attracted the attention of more and more scholars, and many excellent algorithms have been produced. In addition, more studies have shown that the definition of positive and negative samples in target detection data sets is the fundamental difference between the performance of the two. Therefore, it is feasible to design a reasonable, fast and robust pedestrian detection based on a single-stage anchor-free framework. In summary, this embodiment provides a pedestrian detection method based on a single-stage anchor-free framework, such as Figure 1 As shown, the method includes:
[0052] A pedestrian dataset is obtained; the pedestrian dataset is pedestrian images in which pedestrians are framed.
[0053] The pedestrian data set is preprocessed to obtain an atlas; the atlas includes a mask map, a size map and an offset map; the mask map is a map including a pedestrian center point and a Gaussian distribution mask of a pedestrian image; the size map is a map including pedestrian aspect ratio information; the offset map is a map including pedestrian center point offset information; the pedestrian center point is the center point of the center box of the pedestrian data set; the aspect ratio information is the aspect ratio of the pedestrian, and the pedestrian center point offset information is the offset between the pedestrian center point in the offset map and the pedestrian center point in the pedestrian data set.
[0054] The pedestrian data set is input into a network model to obtain a third feature map; the network model includes a residual network, a cross-fusion feature pyramid and a detection head connected in sequence; the normalization method in the residual network includes group normalization and batch normalization, group normalization and batch normalization are combined according to weights, and the weights can be optimized through training.
[0055] The third feature map is analyzed and regressed (the error between the third feature map and each map in the map collection is calculated; the network model is trained using a back propagation algorithm based on the error to obtain a pedestrian detection model, and the pedestrian detection model is used for pedestrian detection).
[0056] In practical applications, before preprocessing the pedestrian dataset to obtain the atlas, the method further includes:
[0057] The pedestrian data set is subjected to data enhancement processing to obtain an enhanced pedestrian data set, thereby increasing the number of samples. The enhanced pedestrian data set can then be input into a network model to obtain a third feature map.
[0058] In practical applications, after performing data enhancement processing on the pedestrian dataset to obtain an enhanced pedestrian dataset, before preprocessing the pedestrian dataset to obtain an atlas, it may also include zooming in or out on the pedestrian dataset to obtain an operated pedestrian dataset. Subsequently, the operated pedestrian dataset may be preprocessed to obtain an atlas, or the operated pedestrian dataset may be input into a network model to obtain a third feature map.
[0059] In practical applications, the data enhancement processing performed on the pedestrian dataset to obtain an enhanced pedestrian dataset specifically includes:
[0060] The pedestrian dataset is mirrored and resized in sequence, wherein the resizing operation is random expansion or random cropping, specifically: a threshold α (usually α=0.5) is set, a number β in [0,1] is randomly generated for each pedestrian image in the pedestrian dataset, α and β are compared, if α<β, the pedestrian image and label remain unchanged, if α>β, the pedestrian image and label are mirrored (left-right flipped); a number γ in [0.4,1.5] is randomly generated for the pedestrian image (image size is h,w) in the pedestrian dataset obtained by the mirroring operation, and the pedestrian image is operated accordingly according to the random number γ, if γ<1, the pedestrian image is randomly cropped and the label is transformed accordingly; if γ>1, the edges of the pedestrian image are randomly padded with 0 and the label is transformed accordingly; the label is a box selection of the pedestrian in the image.
[0061] In practical applications, a feature map containing Gaussian distribution mask information and pedestrian center points can be generated according to a pedestrian data set. For the pedestrian data set, a center point mask is generated using a label; a Gaussian mask is generated using a Gaussian distribution with the pedestrian center point as the center to retain the pedestrian's feature information; the two masks (center point mask and Gaussian mask) are combined into a feature map with a dimension size of [2, h / n, w / n] (in practical applications, n can be 4), so the determination process of the mask map is specifically as follows:
[0062] The pedestrian data set is processed according to the pedestrian center point to generate a center point mask.
[0063] Taking the pedestrian center point as the center, Gaussian distribution processing is performed on the pedestrian data set to generate a Gaussian mask.
[0064] The center point mask and the Gaussian mask are combined to obtain a mask image.
[0065] In practical applications, the process of determining the size map is as follows: for the pedestrian dataset, a feature map having pedestrian aspect ratio information is generated according to pedestrian target label information.
[0066] In practical applications, the process of determining the offset map is as follows: for the pedestrian data set, a feature map is generated that retains error information caused by the offset of the pedestrian target center point due to operations such as image scaling, mirroring, cropping or expansion.
[0067] In practical applications, inputting the pedestrian dataset into the network model to obtain the third feature map specifically includes:
[0068] The pedestrian data set is input into a residual network to obtain a first feature map.
[0069] The first feature map is processed by using a cross-fusion feature pyramid to obtain a second feature map.
[0070] The second feature map is subjected to a detection head operation to obtain a third feature map.
[0071] In practical applications, performing a detection head operation on the second feature map to obtain a third feature map specifically includes:
[0072] The second feature map is subjected to three convolution operations to obtain a center heat map (gather map and centermap), a scale heat map and an offset heat map, wherein the center heat map can generate a feature map containing a pedestrian target Gaussian mask and center point information through a 1×1×1 convolution kernel, the scale heat map can generate a feature map containing pedestrian target scale information through a 1×1×1 convolution kernel, and the offset heat map can generate a feature map containing pedestrian target center point offset information through a 1×1×2 convolution kernel.
[0073] The center heat map, the scale heat map and the offset heat map are fused to obtain a third feature map, where the dimension of the third feature map is [3, h / n, w / n], where [0, h / n, w / n] contains the Gaussian mask and center point information of the pedestrian target, [1, h / n, w / n] contains the scale information of the pedestrian target, and [2, h / n, w / n] contains the offset information of the center point of the pedestrian target.
[0074] In practical applications, the error between the third feature map and each map in the atlas is calculated as follows:
[0075] Calculate the error L between the mask image and the third feature map [0,h / n,w / n] 51 , that is, the center point error is in Among them, K is the total number of pedestrians in the image, To indicate whether there is a pedestrian center at this position, the parameter α is similar to a coefficient. To reduce the influence of pixels near the pedestrian center, g p is the Gaussian component in the third feature map [0,h / n,w / n], g t is the Gaussian component in the mask image, γ is set to 2, and β is set to 4; β and γ are both pre-set hyperparameters. In order to control the penalty, the clustering error Where K is the total number of pedestrians in the image, c p is the center point component of the pedestrian target in the third feature map [0,h / n,w / n], c t is the center point component of the pedestrian target in the mask image; that is, L 51 =L center +L gather .L center Used to calculate the positioning error of pedestrian targets, L gather Used to help shorten the distance between the predicted pedestrian target and the corresponding pedestrian label.
[0076] Calculate the error L between the size map and the third feature map [1,h / n,w / n] 52 ,Right now Where K is the total number of pedestrian targets in the image, s p is the aspect ratio information component of the pedestrian target in the third feature map [1,h / n,w / n], s t L is the aspect ratio information component of the pedestrian target in the size image. scale Used to calculate the scale error of pedestrian targets.
[0077] Calculate the error L between the offset map and the third feature map [2,h / n,w / n] 53 ,Right now Where K is the total number of pedestrian targets in the image, o p is the offset component of the pedestrian target center point in the third feature map [2,h / n,w / n], o t is the offset component of the pedestrian target center point in the offset map. offset Used to calculate the offset error of the pedestrian target center point.
[0078] In practical applications, the network model is trained using a back propagation algorithm according to the error to obtain a pedestrian detection model, and the pedestrian detection model is used to perform pedestrian detection, specifically including:
[0079] According to L=L 51 +L 52 +L 53 Calculate the total error.
[0080] Back propagation is performed based on the total error to update the network parameters, and the updated network will be used for pedestrian detection.
[0081] In practical applications, such as Figure 2 As shown in the figure, the pedestrian dataset is input into the new Resnet-50 residual network to obtain the first feature map, and then the first feature map is analyzed using the cross-fusion feature pyramid to obtain the second feature map, and finally the third feature map is obtained through the detection head.
[0082] In practical applications, the first feature map is specifically a feature map output by inputting a pedestrian data set (which may be an initially acquired pedestrian data set, a pedestrian data set after data enhancement, or a pedestrian data set obtained after an operation by enlarging or reducing) into the last three layers (the third, fourth, and fifth layers) of the residual network as the first feature map, wherein the residual network may be a new Resnet-50 residual network, and the new Resnet-50 residual network replaces the batch normalization (Batch Normalization, BN) in the Resnet-50 residual network with the batch group normalization (Batch-Group Normalization, BGN), such as Figure 3 As shown. It is also possible to replace batch normalization in other forms of Resnet-50 residual networks with batch group normalization. Figure 3 This is just an example of a structure. BGN replaces the instance normalization (IN) and layer normalization (LN) in switchable normalization (SN) with group normalization (GN), and combines BN and GN with a trainable weight ratio; the combination is as follows: where β, γ and ∈ are small constants, h represents the pixels in the image before normalization (h is only expressed as pixels here), represents the normalized pixels in the image, k∈{bn,gn}, bn represents batch normalization, gn represents group normalization, μ k and are the expectation and variance respectively, w k To obtain the desired weight required for standardization after softmax, λ k represents the control parameter of softmax, w' k is the weight of the variance required for standardization after softmax, λ′ k represents the control parameter of softmax, λ k ,λ′ k They are w k and w' k , the parameters that need to be optimized.
[0083] In practical applications, the cross-fusion feature pyramid starts from the third stage of the Resnet-50 residual network and is divided into three steps. The method for determining the new fifth-layer feature map is as follows: Figure 4 As shown, Figure 2 One of the cross-fusion feature pyramids shown Taking the feature map of the 5th layer of the Resnet-50 residual network as an example, the first feature map is processed by the cross-fusion feature pyramid to obtain the second feature map. The specific scheme is as follows:
[0084] Step 41: Get the feature maps of the 3rd, 4th, and 5th layers generated by the Resnet-50 residual network.
[0085] Step 42: The feature map of the third layer is further extracted through a 3×3 convolution kernel, and its dimension is reduced to the same as the feature map of the fourth layer. Figure 1 The feature map of the 5th layer is upsampled to make it consistent with the dimension of the feature map of the 4th layer.
[0086] Step 43: Copy the feature map of the 4th layer, and merge it with the two feature layers generated in step 42 to obtain a new feature map of the 5th layer.
[0087] In a manner similar to step 41, step 42, and step 43, a new third-layer feature map is obtained (the feature maps of the third, fourth, and fifth layers are all upsampled to obtain a map with the same dimension as the second-layer feature map, and the three feature maps are fused to obtain a new third-layer feature map) and a new fourth-layer feature map (the feature maps of the fourth and fifth layers are upsampled to obtain a map with the same dimension as the third-layer feature map, and the two maps are fused with the feature map of the third layer to obtain a new fourth-layer feature map), and the new third-layer feature layer, fourth-layer feature layer, and fifth-layer feature layer are subjected to steps similar to step 42 and step 43 to obtain a second feature map with a dimension size of [256, h / n, w / n]. The dimension size of the second feature map is consistent with the size of the three maps after preprocessing of the pedestrian dataset.
[0088] The present invention also provides the following solution:
[0089] like Figure 5 As shown, the detection system includes: a pedestrian image acquisition module, a training and optimization module, a detection module and a display module.
[0090] The pedestrian image acquisition module is mainly responsible for the acquisition of pedestrian images and corresponding labels, and transmits the acquired data to the training and optimization module; the training and optimization module includes an atlas determination module, a third feature map determination module, an error calculation module and a model determination module, which can receive pedestrian images and labels from the pedestrian image acquisition module, and use them to train and optimize the network framework, mainly for training network models, and finally obtain a trained model and send it to the detection module; the detection module includes using the trained network model to detect pedestrian images, and outputs the results to the display module; the display module has an image display function, displays the image data sent by the detection module, and obtains the detection result of whether there are pedestrians in the image.
[0091] Specifically, the pedestrian image acquisition module is used to obtain a pedestrian data set; the pedestrian data set is pedestrian images in which pedestrians are framed.
[0092] An atlas determination module is used to preprocess the pedestrian data set to obtain an atlas; the atlas includes a mask map, a size map and an offset map; the mask map is a map including a pedestrian center point and a Gaussian distribution mask of a pedestrian image; the size map is a map including pedestrian aspect ratio information; the offset map is a map including pedestrian center point offset information; the pedestrian center point is the center point of the center box of the pedestrian data set.
[0093] The third feature map determination module is used to input the pedestrian data set into a network model to obtain a third feature map; the network model includes a residual network, a cross-fusion feature pyramid and a detection head connected in sequence; the normalization method in the residual network includes group normalization and batch normalization.
[0094] The error calculation module is used to calculate the error between the third feature map and each map in the map set.
[0095] The model determination module is used to train the network model using a back propagation algorithm according to the error to obtain a pedestrian detection model, and the pedestrian detection model is used to perform pedestrian detection.
[0096] As an optional implementation, the pedestrian detection system based on the single-stage anchor-free framework further includes:
[0097] The enhancement module is used to perform data enhancement processing on the pedestrian dataset to obtain an enhanced pedestrian dataset.
[0098] As an optional implementation, the enhancement module specifically includes:
[0099] The enhancement unit is used to sequentially perform a mirroring operation and a size adjustment operation on the pedestrian data set, wherein the size adjustment operation is a random expansion or a random cropping.
[0100] As an optional implementation, the atlas determination module includes:
[0101] The center point mask generating unit is used to process the pedestrian data set according to the pedestrian center point to generate a center point mask.
[0102] The Gaussian mask generating unit is used to generate a Gaussian mask by performing Gaussian distribution processing on the pedestrian data set with the pedestrian center point as the center.
[0103] The mask image determining unit is used to combine the center point mask and the Gaussian mask to obtain a mask image.
[0104] As an optional implementation manner, the third feature map determination module includes:
[0105] The first feature map determining unit is used to input the pedestrian data set into a residual network to obtain a first feature map.
[0106] The second feature map determining unit is used to process the first feature map by using a cross-fusion feature pyramid to obtain a second feature map.
[0107] The third feature map determining unit is used to perform a detection head operation on the second feature map to obtain a third feature map.
[0108] As an optional implementation manner, the third feature map determining unit includes:
[0109] The convolution operation subunit is used to perform three convolution operations on the second feature map to obtain a center heat map, a scale heat map and an offset heat map.
[0110] The third feature map determination subunit is used to perform image fusion on the center heat map, the scale heat map and the offset heat map to obtain a third feature map.
[0111] The present invention has the following technical effects:
[0112] 1) The model obtained by training the network structure of the present invention using a residual network, a cross-fusion feature pyramid and an error calculation method can shorten the distance between the predicted target and the annotation, thereby solving the problem of small differences between pedestrians to a certain extent and improving the detection accuracy.
[0113] 2) The new BGN normalization method is used to more effectively extract features from the image, filter out useless features, improve the accuracy of pedestrian detection, and will not affect the detection speed during the test, thereby improving the accuracy of pedestrian detection; L gather By shortening the distance between the predicted target and the annotation, the problem of small differences between pedestrians can be solved to a certain extent.
[0114] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0115] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A pedestrian detection method based on a single-stage anchor-free framework, characterized in that: include: Get the pedestrian dataset; The pedestrian dataset is a pedestrian image in which pedestrians are framed; Preprocessing the pedestrian data set to obtain an atlas; the atlas includes a mask map, a size map and an offset map; The mask image is an image including the pedestrian center point and the pedestrian image Gaussian distribution mask; The size map is a map including pedestrian aspect ratio information; the offset map is a map including pedestrian center point offset information; the pedestrian center point is the center point of the pedestrian data center box; Inputting the pedestrian data set into a network model to obtain a third feature map; the network model includes a residual network, a cross-fusion feature pyramid and a detection head connected in sequence; the normalization method in the residual network includes group normalization and batch normalization; Calculating the error between the third feature map and each map in the atlas; The network model is trained using a back propagation algorithm according to the error to obtain a pedestrian detection model, wherein the pedestrian detection model is used for pedestrian detection; Inputting the pedestrian data set into the network model to obtain the third feature map specifically includes: Inputting the pedestrian data set into a residual network to obtain a first feature map; Using a cross-fusion feature pyramid to process the first feature map to obtain a second feature map; The second feature map is subjected to a detection head operation to obtain a third feature map.
2. The pedestrian detection method based on a single-stage anchor-free framework according to claim 1, characterized in that: Before the pedestrian data set is preprocessed to obtain an atlas, the method further includes: Performing data enhancement processing on the pedestrian dataset to obtain an enhanced pedestrian dataset.
3. The pedestrian detection method based on a single-stage anchor-free framework according to claim 2, characterized in that: The performing data enhancement processing on the pedestrian dataset to obtain an enhanced pedestrian dataset specifically includes: A mirroring operation and a size adjustment operation are sequentially performed on the pedestrian data set, wherein the size adjustment operation is random expansion or random cropping.
4. The pedestrian detection method based on a single-stage anchor-free framework according to claim 1, characterized in that: The process of determining the mask image is as follows: Processing the pedestrian data set according to the pedestrian center point to generate a center point mask; Taking the pedestrian center point as the center, performing Gaussian distribution processing on the pedestrian data set to generate a Gaussian mask; The center point mask and the Gaussian mask are combined to obtain a mask image.
5. A pedestrian detection system based on a single-stage anchor-free framework, characterized in that: include: Pedestrian image acquisition module, used to obtain pedestrian data sets; The pedestrian dataset is a pedestrian image in which pedestrians are framed; An atlas determination module, used for preprocessing the pedestrian data set to obtain an atlas; the atlas includes a mask map, a size map and an offset map; the mask map is a map including a pedestrian center point and a Gaussian distribution mask of a pedestrian image; The size map is a map including pedestrian aspect ratio information; the offset map is a map including pedestrian center point offset information; the pedestrian center point is the center point of the pedestrian data center box; A third feature map determination module is used to input the pedestrian data set into a network model to obtain a third feature map; the network model includes a residual network, a cross-fusion feature pyramid and a detection head connected in sequence; the normalization method in the residual network includes group normalization and batch normalization; An error calculation module, used for calculating the error between the third feature map and each map in the map collection; A model determination module, used to train the network model using a back propagation algorithm according to the error to obtain a pedestrian detection model, wherein the pedestrian detection model is used to perform pedestrian detection; The third feature map determination module includes: A first feature map determining unit, configured to input the pedestrian data set into a residual network to obtain a first feature map; A second feature map determining unit, configured to process the first feature map by using a cross-fusion feature pyramid to obtain a second feature map; The third feature map determining unit is used to perform a detection head operation on the second feature map to obtain a third feature map.
6. A pedestrian detection system based on a single-stage anchor-free framework according to claim 5, characterized in that: Also includes: The enhancement module is used to perform data enhancement processing on the pedestrian dataset to obtain an enhanced pedestrian dataset.
7. A pedestrian detection system based on a single-stage anchor-free framework according to claim 6, characterized in that: The enhancement module specifically includes: The enhancement unit is used to sequentially perform a mirroring operation and a size adjustment operation on the pedestrian data set, wherein the size adjustment operation is a random expansion or a random cropping.
8. The pedestrian detection system based on a single-stage anchor-free framework according to claim 5, characterized in that: The atlas determination module comprises: A center point mask generating unit, configured to process the pedestrian data set according to the pedestrian center point to generate a center point mask; A Gaussian mask generating unit, used for performing Gaussian distribution processing on the pedestrian data set with the pedestrian center point as the center to generate a Gaussian mask; The mask image determining unit is used to combine the center point mask and the Gaussian mask to obtain a mask image.
Citation Information
Patent Citations
Pedestrian shading detection method based on deep learning
CN111144203A
Pedestrian detection method based on semantic enhancement multi-scale feature pyramid fusion
CN112446308A
Pedestrian detection method and system based on lightweight YOLOv3
CN112836657A
Multi-level feature fusion pedestrian detection method based on CSP model
CN113033371A
Pedestrian multi-target tracking method based on multivariate difference fusion
CN113221787A
