WO2021217542A1 - 步态识别方法、装置、终端及存储介质 - Google Patents

步态识别方法、装置、终端及存储介质 Download PDF

Info

Publication number
WO2021217542A1
WO2021217542A1 PCT/CN2020/087966 CN2020087966W WO2021217542A1 WO 2021217542 A1 WO2021217542 A1 WO 2021217542A1 CN 2020087966 W CN2020087966 W CN 2020087966W WO 2021217542 A1 WO2021217542 A1 WO 2021217542A1
Authority
WO
WIPO (PCT)
Prior art keywords
gait
image
humanoid
target
human body
Prior art date
Application number
PCT/CN2020/087966
Other languages
English (en)
French (fr)
Inventor
魏丞昊
龙浩
Original Assignee
深圳大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳大学 filed Critical 深圳大学
Priority to PCT/CN2020/087966 priority Critical patent/WO2021217542A1/zh
Publication of WO2021217542A1 publication Critical patent/WO2021217542A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition

Definitions

  • the present invention relates to the field of biological information authentication, in particular to a gait recognition method, device, terminal and storage medium.
  • various external and internal features of people such as face, iris, palm print, gait, etc. can be used to identify identity, for example, face, iris, palm print, and gait can be used to clock in at work and access control systems , Mobile phone unlocking and other aspects.
  • face recognition, iris recognition, and palmprint recognition are all limited by distance. In this way, the recognition operation may not be realized in some cases, while gait recognition does not have this limitation.
  • the traditional gait recognition technology mainly compresses all the corresponding gait contour maps into one image.
  • the recognition operation is realized only through image matching, and the timing information included in the gait is ignored, resulting in the final
  • the accuracy of gait recognition results is poor.
  • the present invention provides a method, device, terminal and storage medium for gait recognition to solve the problem of poor gait recognition accuracy.
  • an embodiment of the present invention provides a method for gait recognition, including:
  • the acquiring multiple image sequences corresponding to the target individual, each of the image sequences including multiple images, and extracting the key points of the human body posture in each of the image sequences includes:
  • the human body pose estimation algorithm is used to extract the preset key points of the human body pose in each of the original images in the image sequence.
  • the extracting the key points of the human body posture preset in each of the original images in the image sequence by using a human posture estimation algorithm includes:
  • the alphapose algorithm is used to extract the key points of the human body posture in each of the original images.
  • the collecting a number of original images corresponding to the target individual, and constructing the multiple image sequences based on the original images includes:
  • the target image is compressed and cropped according to a preset size to obtain a compressed image, and the multiple image sequences are formed based on the compressed image.
  • the deleting the original image that does not meet a preset standard requirement from the original image to obtain the target image includes:
  • the original image in which a plurality of individual instances exist is deleted based on the standard requirement.
  • the acquiring a humanoid contour map and a humanoid stick figure corresponding to each of the image sequences based on a plurality of the image sequences and corresponding key points of the human body posture includes:
  • the human-shaped stick figure corresponding to each of the original images is obtained based on the key points of the human body posture.
  • the extracting the human body posture key points in each of the original images, and obtaining the human figure contour map corresponding to each of the original images based on the human body posture key points includes:
  • the Pose2Seg network is used to obtain the contour map of the human figure.
  • the fusion of the humanoid contour map and the humanoid stick figure to obtain a dual-channel fusion gait map sequence corresponding to the image sequence includes:
  • Fusion processing is performed on the humanoid stick figure and the corresponding humanoid contour figure to obtain the dual-channel fusion gait figure sequence.
  • the extracting the target gait feature corresponding to the target individual in the dual-channel fusion gait graph sequence includes:
  • the target gait feature is determined based on the first gait feature and the second gait feature.
  • said extracting the first step state features corresponding to the humanoid stick figure in each of the two-channel fusion gait sequence;
  • the second gait feature corresponding to the outline of the human figure includes:
  • the GaitSet network is used to extract the second gait feature.
  • the recognizing based on the target gait features based on a plurality of the dual-channel fusion gait graph sequences to obtain a gait recognition result includes:
  • the K-nearest neighbor algorithm is used to obtain the gait recognition result based on the target gait feature.
  • an embodiment of the present invention provides a gait recognition device, including:
  • the image acquisition module is used to acquire multiple image sequences corresponding to the target individual
  • the first extraction module is used to extract the key points of the human body posture in each of the image sequences
  • An image generation module which is used to generate a corresponding human-shaped contour map and a human-shaped stick figure based on the image sequence and the key points of the human body posture;
  • An image fusion module for fusing the humanoid contour map and the humanoid stick figure to obtain a dual-channel fusion gait map
  • the second extraction module is used to extract the target gait feature corresponding to the target individual in the dual-channel fusion gait graph
  • the gait recognition module is used to realize the recognition operation of the target individual based on the target gait feature.
  • an embodiment of the present invention provides a terminal device, including a memory, a processor, and a computer program stored on the memory and running on the processor, wherein the processor executes the The computer program realizes the gait recognition method as described in any one of the above.
  • an embodiment of the present invention also provides a computer-readable storage medium, including computer instructions, which when run on a computer, cause the computer to execute the gait recognition method described in any one of the above.
  • the corresponding key points of the human body posture are respectively extracted, and then the human figure contour map is generated based on the key points of the human body posture and the image sequence And the humanoid stick figure; further fusion of the humanoid contour map and the humanoid stick figure to obtain a dual-channel fusion gait diagram, and extract the gait features of the dual-channel fusion gait diagram, and then the gait can be realized based on the gait feature Identify the operation.
  • the key points of the human body posture of the target individual are processed corresponding to different image sequences to obtain the humanoid contour map and the humanoid stick figure used to generate the dual-channel fusion gait map, and then extract the dual-channel fusion step
  • the corresponding gait characteristics in the state diagram can realize the recognition operation of the target individual according to the preset algorithm.
  • Fig. 1 is a schematic flow chart of the gait recognition method in an embodiment
  • FIG. 2 is a schematic diagram of the extraction process of the key points of the human body posture in an embodiment
  • FIG. 3 is a schematic diagram of the generation process of the image sequence in an embodiment
  • FIG. 4 is a schematic diagram of the generation process of the human-shaped contour diagram and the human-shaped stick diagram in an embodiment
  • Fig. 5 is a schematic diagram of the outline of the human figure in an embodiment
  • Figure 6 is a schematic diagram of the human-shaped stick figure in an embodiment
  • Fig. 7 is a schematic diagram of the generation process of the dual-channel fusion gait diagram in an embodiment
  • FIG. 8 is a schematic diagram of the extraction process of the target gait feature in an embodiment
  • Figure 9 is a schematic structural diagram of the gait recognition device in an embodiment
  • FIG. 10 is a schematic diagram of the structure of the gait recognition device in another embodiment.
  • Fig. 11 is a schematic diagram of the internal structure of a computer device running the above gait recognition method in an embodiment.
  • the gait recognition method performs a series of processing and analysis on the image corresponding to the target individual, and realizes the recognition result through the gait based on the result of the processing and analysis.
  • the gait recognition method of this embodiment includes the steps:
  • Step S10 Acquire multiple image sequences corresponding to the target individual, each of the image sequences includes multiple images, and extract the key points of the human body posture in each of the image sequences.
  • the target individual refers to the object to be identified, and specifically, the gait recognition operation of the target individual is realized by obtaining an image corresponding to the target individual.
  • this embodiment constitutes the image sequence by acquiring images of the target individual at different angles and in different environmental conditions.
  • the key points of the human body posture are extracted from the image sequences of the target individual under different angles and different environmental conditions; this ensures that the key points of the human body posture extracted based on the image sequence can more fully reflect the gait characteristics of the target individual. In turn, it is beneficial to improve the accuracy and precision value of gait recognition.
  • performing the key points of the human body posture based on the image sequences specifically includes the following steps:
  • Step S20 Collect a number of original images corresponding to the target individual, and construct the multiple image sequences based on the original images.
  • each image sequence includes multiple images of the target individual, and the target individual is collected under different angles and different environmental conditions. ; It can collect a number of images corresponding to different angles and different environmental conditions and record them as original images.
  • the image sequence can be classified corresponding to different angles and different environmental conditions.
  • Different environmental conditions can mean that the target individual is in a different environment, or the target individual’s clothing is different; for example,
  • the original image corresponding to the front of the target individual is taken as an image sequence
  • the original image corresponding to the side of the target individual is taken as an image sequence
  • the original image corresponding to the target individual wearing a suit is taken as an image sequence
  • the corresponding target individual is in a dense crowd
  • the original image in the environment of a small number of people is used as an image sequence and so on.
  • Step S22 Extract the pre-set key points of the human body posture in each of the original images in the image sequence by using a human body posture estimation algorithm.
  • this embodiment performs the extraction operation of the key points of the human posture on each original image, so as to ensure that the extracted key points of the human posture can fully reflect the target individual’s Features, to achieve the effect of improving the accuracy of gait recognition.
  • the extraction operation of the key points of the human body posture in each original image can be realized based on the human body posture estimation algorithm, for example, the extraction operation of the key points of the human body posture can be realized by using the alphapose algorithm, and the alphapose algorithm can be used to obtain 17 key points of human posture.
  • the alphapose algorithm is used to extract the key points of human body posture, in the process of realizing gait recognition, several of the above-mentioned 17 key points of human body posture can be selected for the recognition operation according to actual needs.
  • the specific values of the human body posture key points can be expressed in the coordinates in the image numerical matrix.
  • x[k] is used to represent the abscissa of the key points of the human body posture
  • y[k] is used to represent the ordinate of the key points of the human body posture.
  • the value range of k is (1,2,...,17); x[k ], y[k] can determine the specific range size according to the actual size of the original image, for example, x[k] takes (1,2,...,320), y[k] takes (1,2,...,240) , It means that the size of the original image is 320*240.
  • Step S30 Delete the original image that does not meet the preset standard requirements from the original image to obtain a target image; and Step S32: Compress and crop the target image according to a preset size to obtain a compressed image, based on The compressed image constitutes the plurality of image sequences.
  • this embodiment performs preprocessing operations on all original images.
  • corresponding standard requirements can be set to eliminate images that do not meet the requirements of the standard in the original image; for example, the standard requirements can be set to include the complete outline of the target individual, there are no multiple individual instances, and so on. In this way, a target image that can be used to accurately reflect the target individual can be obtained, thereby helping to improve the accuracy of recognition.
  • the size of each target image may also be different.
  • only the target image is required to accurately reflect the contour of the target individual and the human body can be performed based on the target object.
  • the key points of the posture can be extracted.
  • this embodiment performs processing such as compression and cropping on the target image. That is, after the target image is compressed and cropped, a target image of a preset size is obtained. For example, the image can be compressed and cropped to a size of 320*240*3, and then multiple image sequences can be constructed based on the compressed original image.
  • the preprocessing can ensure that the image sequence can accurately reflect the contours of the target individual and other features, and then extract the key points of the human body posture At the same time, a more comprehensive human body posture key point that can reflect the characteristics of the target individual is obtained; and the compression processing can reduce the storage space of the image sequence, which is beneficial to increase the processing speed of the image sequence, that is, to improve the efficiency of gait recognition.
  • Step S12 Based on a plurality of the image sequences and the corresponding key points of the human body posture, obtain a humanoid contour map and a humanoid stick figure corresponding to each of the image sequences.
  • the process of generating a humanoid contour map and a humanoid stick figure based on the image sequence and the key points of the human body posture includes:
  • Step S40 Extract the human body posture key points in each of the original images, and obtain the human figure contour map corresponding to each of the original images based on the human body posture key points; and Step S42: Based on the human body posture The key point obtains the human-shaped stick figure corresponding to each of the original images.
  • each original image and the key points of the human body posture corresponding to the original image are input into the Pose2Seg network to realize the acquisition process of the human figure contour map; specifically, the binary value shown in FIG. 5 can be obtained through the Pose2Seg network
  • the contour map of the human figure is the contour map of the human figure corresponding to each original image and the key points of the human body posture.
  • the size of the binarized humanoid contour map is consistent with the original image and/or the original size obtained after compression. For example, if the pixel size of the compressed image is 320*240, then the pixels of the corresponding binarized humanoid contour map The size is also 320*240.
  • each compressed image and/or the original image and the corresponding key points of the human body posture can be binarized, and then each corresponding human-shaped stick figure can be represented by a matrix, for example, A[i,j] can be used to represent the human stick figure obtained from each compressed image and/or original image and the corresponding key points of the human body posture; specifically, the following formula can be used
  • the value range of i, j is related to the size of the compressed image. For example, when the pixel size of the compressed image and/or the original image is 320*240, the value range of i is (1, 2, ..., 320) , The value range of j is (1, 2, ..., 240), and the initial value of a[i, j] is 0.
  • the human-shaped stick figure shown in Figure 6 can be obtained.
  • the key points of the human body posture are mapped to the corresponding positions, and the coordinates representing the shoulders, left and right hands, hip bones, legs, knees, and left and right feet are connected to the line.
  • the coordinate value of is set to 1, and the final a[i,j] is the numerical representation of the binarized stick figure.
  • the corresponding human figure contour image and the human figure stick figure can be obtained, and the human body posture key points can be combined with the image sequence, and the human body posture key points can be displayed in the corresponding image sequence, which can be combined from Images and key points are used to characterize the target individual, which is conducive to the identification of the target individual.
  • Step S14 For each of the image sequences, fuse the humanoid contour map and the humanoid stick figure to obtain a dual-channel fusion gait map sequence corresponding to the image sequence.
  • the fusion of the humanoid contour map and the humanoid stick figure includes the following steps:
  • Step S50 Use the humanoid stick figure corresponding to each of the original images as the second dimension corresponding to the humanoid contour map; and Step S52: For each of the image sequences, fuse the humanoid contour map, The humanoid stick figure obtains a two-channel fusion gait figure sequence corresponding to the image sequence.
  • the corresponding humanoid stick figure can be used as the second dimension of the humanoid contour map, and the fusion operation of the humanoid stick figure and the humanoid contour map can be realized.
  • a certain humanoid stick figure is represented by a matrix a[i,j]
  • a certain humanoid contour figure is represented by a matrix b[m,n]
  • the humanoid stick figure is taken as the first figure of the humanoid contour figure.
  • the two dimensions can be expressed as c[i*j,m*n].
  • the corresponding dual-channel fusion gait sequence can be generated corresponding to multiple image sequences, that is, corresponding to different angles and different environmental conditions, the corresponding dual-channel fusion gait diagrams are constructed corresponding sequence sets, so that it can be based on the angle, Different environmental conditions can realize the recognition operation of target individuals, which is beneficial to improve the accuracy of recognition.
  • the humanoid stick figure and the humanoid contour map can be merged to obtain the corresponding fusion gait diagram, and then the fusion gait diagram can be further processed to realize the step The operation of state recognition.
  • Step S16 Extract the target gait features corresponding to the target individual in the dual-channel fusion gait chart sequence, and perform recognition based on the target gait features of a plurality of the dual-channel fusion gait chart sequences to obtain Gait recognition results.
  • the process of achieving target gait feature extraction and achieving the recognition operation of the target individual includes:
  • Step S60 Extract the first step state feature corresponding to the humanoid stick figure in each of the two-channel fusion gait sequence; The second gait feature corresponding to the outline of the human figure is described.
  • each dual-channel fusion gait sequence contains a certain number of dual-channel fusion gait diagrams, and every humanoid stick figure and every humanoid contour map have features that can be used for gait recognition; based on this, Extract the first step feature corresponding to the humanoid stick figure on the dual channel fusion gait map in each dual channel fusion gait sequence, and obtain the second gait feature corresponding to the humanoid contour map, so that it can be based on the human shape
  • the features included in the stick figure and the outline figure of the human figure realize the recognition operation of the target individual.
  • the first step feature and the second step feature can be extracted through the GaitSet network.
  • Using the GaitSet network has an efficient extraction speed and can improve the efficiency of gait recognition.
  • Step S64 Determine the target gait feature based on the first gait feature and the second gait feature.
  • the first step feature and the second gait feature can be merged to obtain the target gait feature for the target individual recognition operation. So as to realize the recognition operation of gait characteristics.
  • Step S66 Based on the target gait feature, the K-nearest neighbor algorithm is used to obtain the gait recognition result.
  • the gait recognition operation can be realized based on the target gait characteristics; wherein, the recognition operation can be realized by using a classifier, that is, the target gait characteristics are input into the classifier , You can get the corresponding recognition result, and then realize the recognition operation of the target individual.
  • the K-nearest neighbor algorithm can be used as a classifier to realize the identification groove of the target gait feature. That is, the classification idea based on the K-nearest neighbor algorithm: If most of the preset number (such as k) of the most similar samples in the feature space of a sample belong to a certain category, the sample also belongs to this category.
  • the preset number such as k
  • a preset number of target individuals are collected from one and/or multiple angles, for example, from 0°, 30°, 45°, 60°, 90°, 120° , 145°, 150° and 180° total 9 gait viewing angles for verification.
  • the number of target individuals can be 100 or 200.
  • the collected environmental data requires the target individuals to wear coats and trousers.
  • the 9 gait perspectives are divided into several groups, for example, two groups, each group of 100 people, and each target individual is identified by at least 2 target gait features, and the steps are performed based on the above steps.
  • the final average rank can be obtained as 74%.
  • the gait recognition method of this embodiment is based on the image sequence corresponding to the target individual, and after corresponding processing operations are performed on the image sequence, the accuracy and efficiency of gait recognition can be effectively improved.
  • an embodiment of the present invention provides a gait recognition device 100, which implements a gait recognition operation based on the gait recognition method described in any of the above embodiments.
  • the gait recognition device 100 includes: an image acquisition module 110 for acquiring a plurality of image sequences corresponding to the target individual; a first extraction module 120 for extracting each image Human body posture key points in the sequence; the image generation module 130 is used to generate corresponding human figure contour maps and human figure stick figures based on the image sequence and human body posture key points; the image fusion module 140 is used to fuse human figure contour figures and human figure stick figures The figure obtains the dual-channel fusion gait diagram; the second extraction module 150 is used to extract the gait features corresponding to the target individual in the dual-channel fusion gait diagram; the gait recognition module 160 is used to achieve the target individual based on the gait features Recognition operation.
  • the gait recognition device 100 is further provided with an image processing module 170 for preprocessing and compressing the original image; specifically, the image processing module 170 deletes the collected target There is a phenomenon of missing contours and/or multiple individual instances in the image of an individual, and the target image obtained through the above steps is compressed and cropped to reduce the storage space of the image sequence.
  • the realization of the gait recognition device 100 is consistent with the realization idea of the above-mentioned gait recognition method, and the principle of realizing gait recognition will not be repeated here.
  • the principle of realizing gait recognition will not be repeated here.
  • the corresponding key points of the human body posture are respectively extracted, and then the human figure contour map is generated based on the key points of the human body posture and the image sequence And the humanoid stick figure; further fusion of the humanoid contour map and the humanoid stick figure to obtain a dual-channel fusion gait diagram, and extract the gait features of the dual-channel fusion gait diagram, and then the gait can be realized based on the gait feature Identify the operation.
  • the key points of the human body posture of the target individual are processed corresponding to different image sequences to obtain the humanoid contour map and the humanoid stick figure used to generate the dual-channel fusion gait map, and then extract the dual-channel fusion step
  • the corresponding gait characteristics in the state diagram can realize the recognition operation of the target individual according to the preset algorithm.
  • Fig. 11 shows an internal structure diagram of a computer device in an embodiment.
  • the computer device may specifically be a server or a terminal.
  • the computer device includes a processor, a memory, and a network interface connected through a system bus.
  • the memory includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium of the computer device stores an operating system and may also store a computer program.
  • the processor can realize the gait recognition method.
  • a computer program can also be stored in the internal memory, and when the computer program is executed by the processor, the processor can execute the method for gait recognition.
  • FIG. 11 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied.
  • the specific computer device may It includes more or fewer components than shown in FIG. 11, or combines certain components, or has a different component arrangement.
  • the method for gait recognition provided in the present application can be implemented in the form of a computer program, and the computer program can run on the computer device as shown in FIG. 11.
  • the memory of the computer equipment can store various program modules that make up the gait recognition device. For example, the image acquisition module 110 and so on.
  • a computer device including a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the following steps: Multiple image sequences corresponding to the target individual, each image sequence contains multiple images, and the human body posture key points in each image sequence are extracted; based on the multiple image sequences and the corresponding human body posture key points, each image is obtained The humanoid contour map and the humanoid stick figure corresponding to the sequence; for each of the image sequences, the humanoid contour map and the humanoid stick figure are merged to obtain a dual-channel fusion gait sequence corresponding to the image sequence; the dual-channel fusion gait is extracted The target gait features corresponding to the target individual in the graph sequence are identified based on the target gait features of multiple dual-channel fusion gait graph sequences to obtain the gait recognition result.
  • Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
  • SRAM static RAM
  • DRAM dynamic RAM
  • SDRAM synchronous DRAM
  • DDRSDRAM double data rate SDRAM
  • ESDRAM enhanced SDRAM
  • SLDRAM synchronous chain Channel
  • memory bus Radbus direct RAM
  • RDRAM direct memory bus dynamic RAM
  • RDRAM memory bus dynamic RAM

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Image Analysis (AREA)

Abstract

一种步态识别方法，该方法包括：获取与目标个体对应的多个图像序列，每个图像序列包含多个图像，提取每一个图像序列中的人体姿态关键点（S10）；基于多个图像序列和对应的所述人体姿态关键点，获取与每一个图像序列对应的人形轮廓图和人形棍状图（S12）；针对每一个所述图像序列，融合人形轮廓图、人形棍状图得到与该图像序列对应的双通道融合步态图序列（S14）；提取双通道融合步态图序列中与目标个体对应的目标步态特征，基于多个双通道融合步态图序列的目标步态特征进行识别，以得到步态识别结果（S16），能够提升提高步态识别的精准性。

Description

步态识别方法、装置、终端及存储介质 技术领域
本发明涉及生物信息认证领域，尤其涉及一种步态识别方法、装置、终端及存储介质。
背景技术
在生物认证领域，人的各种外在和内在特征如人脸、虹膜、掌纹、步态等可用来识别身份，例如将人脸、虹膜、掌纹、步态应用在上班打卡、门禁***、手机解锁等各方面中。其中，现有的人脸识别、虹膜识别和掌纹识别都受限于距离，这样，可能在一些情况下无法实现识别操作，而步态识别则不存在这种限制。
但是传统的的步态识别技术主要将对应的所有步态轮廓图压缩成一幅图像，在识别过程中，仅仅通过图像匹配实现识别操作，而忽略了步态中包括的时序信息，从而导致最终的步态识别结果精度较差。
由此可知，如何提升步态识别的精度值是现有技术中亟待解决的问题。
发明内容
有鉴于此，本发明提供了一种步态识别方法、装置、终端及存储介质，用于解决步态识别精度较差的问题。
本发明实施例的具体技术方案为：
第一方面，本发明实施例提供一种步态识别方法，包括：
获取与目标个体对应的多个图像序列，每个所述图像序列包含多个图像，提取每一个所述图像序列中的人体姿态关键点；
基于多个所述图像序列和对应的所述人体姿态关键点，获取与每一个所述 图像序列对应的人形轮廓图和人形棍状图；
针对每一个所述图像序列，融合所述人形轮廓图、所述人形棍状图得到与该图像序列对应的双通道融合步态图序列；
提取所述双通道融合步态图序列中与所述目标个体对应的目标步态特征，基于多个所述双通道融合步态图序列的所述目标步态特征进行识别，以得到步态识别结果。
可选地，所述获取与目标个体对应的多个图像序列，每个所述图像序列包含多个图像，提取每一个所述图像序列中的人体姿态关键点，包括：
采集与所述目标个体对应的若干数量的原始图像，基于所述原始图像构建所述多个图像序列；
利用人体姿态估计算法提取所述图像序列中每一个所述原始图像中预设的所述人体姿态关键点。
可选地，所述利用人体姿态估计算法提取所述图像序列中每一个所述原始图像中预设的所述人体姿态关键点，包括：
利用alphapose算法提取每一个所述原始图像中的所述人体姿态关键点。
可选地，所述采集与所述目标个体对应的若干数量的原始图像，基于所述原始图像构建所述多个图像序列，包括：
删除所述原始图像中不符合预设的标准要求的所述原始图像，得到目标图像；
将所述目标图像按照预设的尺寸大小进行压缩裁剪，得到压缩图像，基于所述压缩图像构成所述多个图像序列。
可选地，所述删除所述原始图像中不符合预设的标准要求的所述原始图像，得到目标图像，包括：
基于所述标准要求删除存在轮廓缺失的所述原始图像；以及
基于所述标准要求删除存在多个个体实例的所述原始图像。
可选地，所述基于多个所述图像序列和对应的所述人体姿态关键点，获取 与每一个所述图像序列对应的人形轮廓图和人形棍状图，包括：
提取每一所述原始图像中的所述人体姿态关键点，基于所述人体姿态关键点获取对应每一所述原始图像的所述人形轮廓图；以及
基于所述人体姿态关键点获取对应每一所述原始图像的所述人形棍状图。
可选地，所述提取每一所述原始图像中的所述人体姿态关键点，基于所述人体姿态关键点获取对应每一所述原始图像的所述人形轮廓图，包括：
利用Pose2Seg网络获取所述人形轮廓图。
可选地，所述针对每一个所述图像序列，融合所述人形轮廓图、所述人形棍状图得到与该图像序列对应的双通道融合步态图序列，包括：
将每一所述原始图像对应的所述人形棍状图作为对应所述人形轮廓图的第二个维度；
对所述人形棍状图与对应的所述人形轮廓图进行融合处理，以得到所述双通道融合步态图序列。
可选地，所述提取所述双通道融合步态图序列中与所述目标个体对应的目标步态特征，包括：
提取每一所述双通道融合步态图序列中与所述人形棍状图对应的第一步态特征；以及
提取每一所述双通道融合步态图序列中与所述人形轮廓图对应的第二步态特征；
基于所述第一步态特征、所述第二步态特征确定所述目标步态特征。
可选地，所述提取每一所述双通道融合步态图序列中与所述人形棍状图对应的第一步态特征；以及提取每一所述双通道融合步态图序列中与所述人形轮廓图对应的第二步态特征，包括：
利用GaitSet网络提取所述第一步态特征；以及
利用GaitSet网络提取所述第二步态特征。
可选地，所述基于基于多个所述双通道融合步态图序列的所述目标步态特 征进行识别，以得到步态识别结果，包括：
基于所述目标步态特征实现利用K近邻算法获取所述步态识别结果。
第二方面，本发明实施例提供一种步态识别装置，包括：
图像获取模块，用于获取与目标个体对应的多个图像序列；
第一提取模块，用于提取每一个所述图像序列中的人体姿态关键点；
图像生成模块，用于基于所述图像序列、所述人体姿态关键点生成对应的人形轮廓图和人形棍状图；
图像融合模块，用于融合所述人形轮廓图、所述人形棍状图得到双通道融合步态图；
第二提取模块，用于提取所述双通道融合步态图中与所述目标个体对应的目标步态特征；
步态识别模块，用于基于所述目标步态特征实现对所述目标个体的识别操作。
第三方面，本发明实施例提供一种终端设备，包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现如上任一项所述的步态识别方法。
第四方面，本发明实施例还提供一种计算机可读存储介质，包括计算机指令，当所述计算机指令在计算机上运行时，使得计算机执行如上任一项所述的步态识别方法。
实施本发明实施例，将具有如下有益效果：
采用了上述步态识别方法、装置、终端及存储介质之后，基于对应目标个体的多个图像采集序列，分别提取对应的人体姿态关键点，再基于该人体姿态关键点和图像序列生成人形轮廓图和人形棍状图；进一步将人形轮廓图和人形棍状图进行融合得到双通道融合步态图，并提取该双通道融合步态图的步态特征，基于该步态特征即可实现步态识别的操作。本实施例目标个体存在的人体姿态关键点，对应不同的图像序列进行一系列的处理，进而得到用于生成双通 道融合步态图的人形轮廓图和人形棍状图，再提取双通道融合步态图中对应的步态特征，即可根据预设的算法实现对目标个体的识别操作。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
其中：
图1为一个实施例中所述步态识别方法的流程示意图；
图2为一个实施例中所述人体姿态关键点的提取流程示意图；
图3为一个实施例中所述图像序列的生成流程示意图；
图4为一个实施例中所述人形轮廓图和所述人形棍状图的生成流程示意图；
图5为一个实施例中所述人形轮廓图示意；
图6为一个实施例中所述人形棍状图示意；
图7为一个实施例中所述双通道融合步态图的生成流程示意图；
图8为一个实施例中所述目标步态特征的提取流程示意图；
图9为一个实施例中所述步态识别装置的结构示意图；
图10为另一个实施例中所述步态识别装置的结构示意图；
图11为一个实施例中运行上述步态识别方法的计算机设备的内部结构示意图。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是 全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
为解决现有技术中在步态识别过程中，对于步态中的时序信息无法考虑，即其实现步态识别仅仅通过简单的图像匹配操作，进而导致识别的精度比较差的问题，在本实施例中，特提出了一种步态识别方法。该步态识别方法通过对目标个体对应的图像进行一系列的处理和分析，基于该处理和分析的结果实现通过步态的识别结果。
在一个实施例中，如图1所示，本实施例的步态识别方法包括步骤：
步骤S10：获取与目标个体对应的多个图像序列，每个所述图像序列包含多个图像，提取每一个所述图像序列中的人体姿态关键点。
目标个体指待识别的对象，具体的，通过获取与该目标个体对应的图像，以实现对该目标个体的步态识别操作。其中，为了保证图像序列能够表征目标个体，即为了提升步态识别的精度值，本实施例通过获取目标个体在不同角度、不同环境状态下的图像构成该图像序列。
进一步的，对目标个体在不同角度、不同环境状态下的图像序列进行人体姿态关键点的提取操作；这样就可保证基于图像序列提取的人体姿态关键点能够更加全面反映目标个体的步态特征，进而有利于提升步态识别的准确性和精度值。
在一个实施例中，如图2所示，获取与目标个体对应的多个图像序列后，基于该图像序列进行人体姿态关键点的具体包括如下步骤：
步骤S20：采集与所述目标个体对应的若干数量的原始图像，基于所述原始图像构建所述多个图像序列。
即获取多个图像序列，然后提取多个图像序列对应的人体姿态关键点；而由于每一图像序列包括多张目标个体的图像，且在不同角度、不同环境状态下对目标个体进行图像的采集；则可对应不同角度、不同环境状态采集若干数量的图像，记为原始图像。
基于采集的若干数量的原始图像，可对应不同角度、不同环境状态进行该图像序列的分类，其中，不同环境状态可以是指目标个体在不同的环境下，也可以指目标个体的着装不同；例如将对应目标个体正面的该原始图像作为一个图像序列，对应目标个体侧面的该原始图像作为一个图像序列，将对应目标个体穿着西装的该原始图像作为一个图像序列，以及将对应目标个体在密集人群或少量人群的环境下的该原始图像作为一个图像序列等等。
步骤S22：利用人体姿态估计算法提取所述图像序列中每一个所述原始图像中预设的所述人体姿态关键点。
具体的，为了尽可能地提升步态识别的准确性，本实施例对每一个原始图像均进行该人体姿态关键点的提取操作，这样就可保证提取的人体姿态关键点能够全面体现目标个体的特征，达到提升步态识别准确性的效果。
其中，在一个实施例中，可基于人体姿态估计算法实现对每一原始图像中人体姿态关键点的提取操作，例如，使用alphapose算法实现人体姿态关键点的提取操作等，且使用alphapose算法可以得到17个人体姿态关键点。
若使用alphapose算法进行人体姿态关键点的提取，则在实现步态识别的过程中，可根据实际的需求选择上述17个人体姿态关键点中的若干个进行识别操作。
示例性地，假设使用其中的14个人体姿态关键点进行不同识别操作，此外，基于图像可通过矩阵进行表示，则可将该人体姿态关键点的具体值在图像数值矩阵中的坐标中表示，如，用x[k]表示人体姿态关键点的横坐标，用y[k]表示人体姿态关键点的纵坐标，其中，k取值范围为(1，2，…，17)；x[k]、y[k]可根据原始图像的实际大小确定具体的范围大小，例如x[k]取(1，2，…，320)，y[k]取值(1，2，…，240)，则表示原始图像的大小为320*240。
为了保证构成的上述多个图像序列能够用于目标个体的识别操作，在一个实施例中，如图3所示，还需要对上述获取的原始图像进行如下步骤的处理：
步骤S30：删除所述原始图像中不符合预设的标准要求的所述原始图像， 得到目标图像；以及步骤S32：将所述目标图像按照预设的尺寸大小进行压缩裁剪，得到压缩图像，基于所述压缩图像构成所述多个图像序列。
实际地，由于原始图像的采集角度、环境状态等原因不同，可能存在目标个体被遮挡或者由于光线采集得到的原始图像无法反映出目标个体等原因，就会存在部分原始图像存在目标个体的轮廓缺失或者多个个体实例的现象。基于此，本实施例对所有原始图像进行预处理操作。
具体的，可设定对应的标准要求，以剔除原始图像中不符合该标准要求的图像；例如，可将该标准要求设定为包括目标个体完整的轮廓、不存在多个个体实例等等，这样，即可得到可用于准确反映目标个体的目标图像，从而有利于提升识别的准确性。
进一步地，由于原始图像的采集角度、环境状态的不同，每一目标图像的大小同样可能存在差异，而实际地，只需要目标图像能够准确反映目标个体的轮廓、并且可基于该目标对象进行人体姿态关键点的提取即可，基于此，本实施例对目标图像进行压缩、裁剪等处理。即将目标图像压缩、裁剪等处理后，得到预设大小的目标图像，例如可将图像压缩裁剪至320*240*3的大小，然后即可基于压缩后的原始图像构建多个图像序列。
基于对应目标个体获取对应的图像序列，并对构成该图像序列的原始图像进行预处理和压缩处理，通过预处理可以保证图像序列能够准确反映目标个体的轮廓等特征，进而在提取人体姿态关键点时，得到更加全面且能够反映目标个体特征的人体姿态关键点；而通过压缩处理，能够减少图像序列的存储空间，有利于提升对图像序列的处理速度，即有利于提升步态识别的效率。
步骤S12：基于多个所述图像序列和对应的所述人体姿态关键点，获取与每一个所述图像序列对应的人形轮廓图和人形棍状图。
具体的，如图4所示，基于图像序列、人体姿态关键点生成人形轮廓图和人形棍状图的包括过程：
步骤S40：提取每一所述原始图像中的所述人体姿态关键点，基于所述人 体姿态关键点获取对应每一所述原始图像的所述人形轮廓图；以及步骤S42：基于所述人体姿态关键点获取对应每一所述原始图像的所述人形棍状图。
在一个实施例中，将每一原始图像和对应该原始图像的人体姿态关键点输入Pose2Seg网络，实现人形轮廓图的获取过程；具体的，通过该Pose2Seg网络可获取如图5所示的二值化人形轮廓图，即对应每一原始图像和人体姿态关键点的人形轮廓图。其中，该二值化人形轮廓图的大小与压缩后得到的原始图像和/或原始的大小一致，例如，若压缩图像的像素大小为320*240，则对应的二值化人形轮廓图的像素大小同样为320*240。
具体的，对于人形棍状图的获取，可基于每一该压缩图像和/或原始图像与对应的人体姿态关键点采用二值化处理后，通过矩阵表示每一对应的人形棍状图，例如可用a[i,j]表示每一压缩图像和/或原始图像与对应的人体姿态关键点获取得到的人形棍状图；具体的，可通过如下公式
Figure PCTCN2020087966-appb-000001
表示所有人形棍状图。其中，i,j的取值范围与压缩图片的大小相关，例如当压缩图像和/或原始图像的像素大小为320*240时，则i的取值范围为(1，2，…，320)，j的取值范围为(1，2，…，240)，且a[i,j]的初始值为0。
基于上述公式可得到如图6所示的人形棍状图，此时，人体姿态关键点即映射到相应位置，分别将代表双肩、左右手、髋骨、双腿膝盖、左右脚等坐标连接线上的坐标值设置为1，最终a[i,j]即为二值化棍状图的数值表示。
基于图像序列和人体姿态关键点获取对应的人形轮廓图和人形棍状图，能够将人体姿态关键点与图像序列结合，由此将人体姿态关键点在对应的图像序列中展现，即可结合从图像与关键点来表征目标个体，进而有利于对目标个体的识别操作。
步骤S14：针对每一个所述图像序列，融合所述人形轮廓图、所述人形棍状图得到与该图像序列对应的双通道融合步态图序列。
具体的，如图7所示，对人形轮廓图、人形棍状图的融合包括如下步骤：
步骤S50：将每一所述原始图像对应的所述人形棍状图作为对应所述人形轮廓图的第二个维度；以及步骤S52：针对每一个所述图像序列，融合所述人形轮廓图、所述人形棍状图得到与该图像序列对应的双通道融合步态图序列。
基于每一人形棍状图和人形轮廓图均为二值化图像，则可将对应的人形棍状图作为人形轮廓图的第二维度，进而实现人形棍状图与人形轮廓图的融合操作。示例性地，假设某一人形棍状图用矩阵a[i,j]表示，某一人形轮廓图用矩阵b[m,n]表示，则将该人形棍状图作为该人形轮廓图的第二维度可表示为c[i*j,m*n]。
同时，可对应生成的多个图像序列生成对应的双通道融合步态图序列，即对应不同角度、不同环境状态将对应的双通道融合步态图构建对应的序列集，这样就可基于角度、不同环境状态实现对目标个体的识别操作，有利于提升识别的准确性。
通过将人形棍状图作为人形轮廓图的第二维度，能够实现将人形棍状图和人形轮廓图进行融合，得到对应的融合步态图，进而进一步对融合步态图进行处理，以实现步态识别的操作。
步骤S16：提取所述双通道融合步态图序列中与所述目标个体对应的目标步态特征，基于多个所述双通道融合步态图序列的所述目标步态特征进行识别，以得到步态识别结果。
具体的，如图8所示，实现目标步态特征提取及实现对目标个体的识别操作的过程包括：
步骤S60：提取每一所述双通道融合步态图序列中与所述人形棍状图对应的第一步态特征；以及步骤S62：提取每一所述双通道融合步态图序列中与所述人形轮廓图对应的第二步态特征。
具体的，因为每一双通道融合步态序列中包含一定数量的双通道融合步态图，且每一个人形棍状图和每一人形轮廓图均存在可用于进行步态识别的特征；基于此，对每一双通道融合步态序列中的双通道融合步态图进行与人形棍状图 对应的第一步态特征的提取，以及获取对应人形轮廓图的第二步态特征，这样即可基于人形棍状图和人形轮廓图中包括的特征，实现对目标个体的识别操作。
其中，在一个实施例中，第一步态特征和第二步态特征的提取可通过GaitSet网络提取，使用GaitSet网络具有高效的提取速度，能够提升步态识别的效率。
步骤S64：基于所述第一步态特征、所述第二步态特征确定所述目标步态特征。
同样的，基于人形棍状图可作为人形轮廓图的第二维度，则可将该第一步态特征、第二步态特征进行融合，以得到用于目标个体识别操作的目标步态特征，从而实现对步态特征的识别操作。
步骤S66：基于所述目标步态特征实现利用K近邻算法获取所述步态识别结果。
具体的，本实施例基于上述步骤得到的目标步态特征，则可基于目标步态特征实现步态识别操作；其中，可利用分类器实现该识别操作，即将目标步态特征输入至分类器中，即可得到对应的识别结果，进而实现对目标个体的识别操作。
示例性地，可采用K近邻算法作为分类器实现对目标步态特征的识别槽。即基于K近邻算法的分类思路：如果一个样本在特征空间中的预设数量(如k个)个最相似的样本中的大多数属于某一个类别，则该样本也属于这个类别。
基于上述步态识别方法的实现实施例，在此可通过具体的实验进行说明该步态识别方法的识别效果。具体的，基于上述方法，对预设数量的目标个体分别从一个和/或多个角度进行图像序列的采集，例如，可从0°、30°、45°、60°、90°、120°、145°、150°和180°总共9个步态视角进行验证，目标个体数量可以是100或200个，采集的环境数据要求目标个体穿外套长裤等。
具体的，将该9个步态视角分成若干组，例如两组，则每一组100人，且每一个目标个体均通过至少2个的目标步态特征进行识别操作，在基于上述步 骤进行步态识别操作后，可得到最后的平均rank为74％。
综上可知，本实施例的步态识别方法基于与目标个体对应的图像序列，并对图像序列进行相对应的处理操作后，能够有效提升步态识别的精度和效率。
基于同一发明构思，本发明实施例提供一种步态识别装置100，该步态识别装置100基于如上任一项实施例所述的步态识别方法实现步态识别操作。
在一个实施例中，如图9所示，该步态识别装置100包括：图像获取模块110，用于获取与目标个体对应的多个图像序列；第一提取模块120，用于提取每一个图像序列中的人体姿态关键点；图像生成模块130，用于基于图像序列、人体姿态关键点生成对应的人形轮廓图和人形棍状图；图像融合模块140，用于融合人形轮廓图、人形棍状图得到双通道融合步态图；第二提取模块150，用于提取双通道融合步态图中与目标个体对应的步态特征；步态识别模块160，用于基于步态特征实现对目标个体的识别操作。
在另一实施例中，如图10所示，该步态识别装置100还设置有用于对原始图像进行预处理和压缩的图像处理模块170；具体的，通过该图像处理模块170删除采集的目标个体的图像中存在轮廓缺失和/或存在多个个体实例的现象，并对通过上述步骤获取的目标图像进行压缩裁剪处理，以减少图像序列的存储空间。
需要说明的是，步态识别装置100的实现与上述步态识别方法的实现思想一致，其实现步态识别的原理在此不再进行赘述，可具体参阅上述步态识别方法中的对应内容。
采用了上述步态识别方法、装置、终端及存储介质之后，基于对应目标个体的多个图像采集序列，分别提取对应的人体姿态关键点，再基于该人体姿态关键点和图像序列生成人形轮廓图和人形棍状图；进一步将人形轮廓图和人形棍状图进行融合得到双通道融合步态图，并提取该双通道融合步态图的步态特征，基于该步态特征即可实现步态识别的操作。本实施例目标个体存在的人体姿态关键点，对应不同的图像序列进行一系列的处理，进而得到用于生成双通 道融合步态图的人形轮廓图和人形棍状图，再提取双通道融合步态图中对应的步态特征，即可根据预设的算法实现对目标个体的识别操作。
图11示出了一个实施例中计算机设备的内部结构图。该计算机设备具体可以是服务器，也可以是终端。如图11所示，该计算机设备包括通过***总线连接的处理器、存储器和网络接口。其中，存储器包括非易失性存储介质和内存储器。该计算机设备的非易失性存储介质存储有操作***，还可存储有计算机程序，该计算机程序被处理器执行时，可使得处理器实现步态识别方法。该内存储器中也可储存有计算机程序，该计算机程序被处理器执行时，可使得处理器执行步态识别的方法。本领域技术人员可以理解，图11中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，具体的计算机设备可以包括比图11中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
在一个实施例中，本申请提供的步态识别的方法可以实现为一种计算机程序的形式，计算机程序可在如图11所示的计算机设备上运行。计算机设备的存储器中可存储组成该步态识别装置的各个程序模块。比如，图像获取模块110等。
在一个实施例中，提出了一种计算机设备，包括存储器和处理器，所述存储器存储有计算机程序，所述计算机程序被所述处理器执行时，使得所述处理器执行以下步骤：获取与目标个体对应的多个图像序列，每个图像序列包含多个图像，提取每一个图像序列中的人体姿态关键点；基于多个图像序列和对应的所述人体姿态关键点，获取与每一个图像序列对应的人形轮廓图和人形棍状图；针对每一个所述图像序列，融合人形轮廓图、人形棍状图得到与该图像序列对应的双通道融合步态图序列；提取双通道融合步态图序列中与目标个体对应的目标步态特征，基于多个双通道融合步态图序列的目标步态特征进行识别，以得到步态识别结果。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程， 是可以通过计算机程序来指令相关的硬件来完成，所述的程序可存储于一非易失性计算机可读取存储介质中，该程序在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
以上所揭露的仅为本发明较佳实施例而已，当然不能以此来限定本发明之权利范围，因此依本发明权利要求所作的等同变化，仍属本发明所涵盖的范围。

Claims (14)

  1. 一种步态识别方法，其特征在于，包括：
    获取与目标个体对应的多个图像序列，每个所述图像序列包含多个图像，提取每一个所述图像序列中的人体姿态关键点；
    基于多个所述图像序列和对应的所述人体姿态关键点，获取与每一个所述图像序列对应的人形轮廓图和人形棍状图；
    针对每一个所述图像序列，融合所述人形轮廓图、所述人形棍状图得到与该图像序列对应的双通道融合步态图序列；
    提取所述双通道融合步态图序列中与所述目标个体对应的目标步态特征，基于多个所述双通道融合步态图序列的所述目标步态特征进行识别，以得到步态识别结果。
  2. 如权利要求1所述的步态识别方法，其特征在于，所述获取与目标个体对应的多个图像序列，每个所述图像序列包含多个图像，提取每一个所述图像序列中的人体姿态关键点，包括：
    采集与所述目标个体对应的若干数量的原始图像，基于所述原始图像构建所述多个图像序列；
    利用人体姿态估计算法提取所述图像序列中每一个所述原始图像中预设的所述人体姿态关键点。
  3. 如权利要求2所述的步态识别方法，其特征在于，所述利用人体姿态估计算法提取所述图像序列中每一个所述原始图像中预设的所述人体姿态关键点，包括：
    利用alphapose算法提取每一个所述原始图像中的所述人体姿态关键点。
  4. 如权利要求2所述的步态识别方法，其特征在于，所述采集与所述目标个体对应的若干数量的原始图像，基于所述原始图像构建所述多个图像序列，包括：
    删除所述原始图像中不符合预设的标准要求的所述原始图像，得到目标图 像；
    将所述目标图像按照预设的尺寸大小进行压缩裁剪，得到压缩图像，基于所述压缩图像构成所述多个图像序列。
  5. 如权利要求4所述的步态识别方法，其特征在于，所述删除所述原始图像中不符合预设的标准要求的所述原始图像，得到目标图像，包括：
    基于所述标准要求删除存在轮廓缺失的所述原始图像；以及
    基于所述标准要求删除存在多个个体实例的所述原始图像。
  6. 如权利要求4所述的步态识别方法，其特征在于，所述基于多个所述图像序列和对应的所述人体姿态关键点，获取与每一个所述图像序列对应的人形轮廓图和人形棍状图，包括：
    提取每一所述原始图像中的所述人体姿态关键点，基于所述人体姿态关键点获取对应每一所述原始图像的所述人形轮廓图；以及
    基于所述人体姿态关键点获取对应每一所述原始图像的所述人形棍状图。
  7. 如权利要求6所述的步态识别方法，其特征在于，所述提取每一所述原始图像中的所述人体姿态关键点，基于所述人体姿态关键点获取对应每一所述原始图像的所述人形轮廓图，包括：
    利用Pose2Seg网络获取所述人形轮廓图。
  8. 如权利要求6所述的步态识别方法，其特征在于，所述针对每一个所述图像序列，融合所述人形轮廓图、所述人形棍状图得到与该图像序列对应的双通道融合步态图序列，包括：
    将每一所述原始图像对应的所述人形棍状图作为对应所述人形轮廓图的第二个维度；
    对所述人形棍状图与对应的所述人形轮廓图进行融合处理，以得到所述双通道融合步态图序列。
  9. 如权利要求8所述的步态识别方法，其特征在于，所述提取所述双通道融合步态图序列中与所述目标个体对应的目标步态特征，包括：
    提取每一所述双通道融合步态图序列中与所述人形棍状图对应的第一步态特征；以及
    提取每一所述双通道融合步态图序列中与所述人形轮廓图对应的第二步态特征；
    基于所述第一步态特征、所述第二步态特征确定所述目标步态特征。
  10. 如权利要求9所述的步态识别方法，其特征在于，所述提取每一所述双通道融合步态图序列中与所述人形棍状图对应的第一步态特征；以及提取每一所述双通道融合步态图序列中与所述人形轮廓图对应的第二步态特征，包括：
    利用GaitSet网络提取所述第一步态特征；以及
    利用GaitSet网络提取所述第二步态特征。
  11. 如权利要求9所述的步态识别方法，其特征在于，所述基于基于多个所述双通道融合步态图序列的所述目标步态特征进行识别，以得到步态识别结果，包括：
    基于所述目标步态特征实现利用K近邻算法获取所述步态识别结果。
  12. 一种步态识别装置，其特征在于，包括：
    图像获取模块，用于获取与目标个体对应的多个图像序列；
    第一提取模块，用于提取每一个所述图像序列中的人体姿态关键点；
    图像生成模块，用于基于所述图像序列、所述人体姿态关键点生成对应的人形轮廓图和人形棍状图；
    图像融合模块，用于融合所述人形轮廓图、所述人形棍状图得到双通道融合步态图；
    第二提取模块，用于提取所述双通道融合步态图中与所述目标个体对应的目标步态特征；
    步态识别模块，用于基于所述目标步态特征实现对所述目标个体的识别操作。
  13. 一种终端设备，其特征在于，包括存储器、处理器及存储在所述存储 器上并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现权利要求1-11中任一项所述的步态识别方法。
  14. 一种计算机可读存储介质，包括计算机指令，当所述计算机指令在计算机上运行时，使得计算机执行如权利要求1-11中任一项所述的步态识别方法。
PCT/CN2020/087966 2020-04-30 2020-04-30 步态识别方法、装置、终端及存储介质 WO2021217542A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/087966 WO2021217542A1 (zh) 2020-04-30 2020-04-30 步态识别方法、装置、终端及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/087966 WO2021217542A1 (zh) 2020-04-30 2020-04-30 步态识别方法、装置、终端及存储介质

Publications (1)

Publication Number Publication Date
WO2021217542A1 true WO2021217542A1 (zh) 2021-11-04

Family

ID=78331619

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/087966 WO2021217542A1 (zh) 2020-04-30 2020-04-30 步态识别方法、装置、终端及存储介质

Country Status (1)

Country Link
WO (1) WO2021217542A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116503943A (zh) * 2023-04-24 2023-07-28 哈尔滨工程大学 一种基于优选动态几何特征的可解释步态识别方法
CN117238032A (zh) * 2023-09-18 2023-12-15 以萨技术股份有限公司 一种步态特征对比方法、存储介质及电子设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9640186B2 (en) * 2014-05-02 2017-05-02 International Business Machines Corporation Deep scattering spectrum in acoustic modeling for speech recognition
CN108280398A (zh) * 2017-12-26 2018-07-13 北京建筑大学 一种基于多视角融合的步态识别方法及***
CN108921062A (zh) * 2018-06-21 2018-11-30 暨南大学 一种联合多步态特征协同字典的步态识别方法
CN110046675A (zh) * 2019-04-25 2019-07-23 河北工业大学 一种基于改进的卷积神经网络的下肢运动能力评估方法
CN110796100A (zh) * 2019-10-31 2020-02-14 浙江大华技术股份有限公司 步态识别方法、装置、终端及存储装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9640186B2 (en) * 2014-05-02 2017-05-02 International Business Machines Corporation Deep scattering spectrum in acoustic modeling for speech recognition
CN108280398A (zh) * 2017-12-26 2018-07-13 北京建筑大学 一种基于多视角融合的步态识别方法及***
CN108921062A (zh) * 2018-06-21 2018-11-30 暨南大学 一种联合多步态特征协同字典的步态识别方法
CN110046675A (zh) * 2019-04-25 2019-07-23 河北工业大学 一种基于改进的卷积神经网络的下肢运动能力评估方法
CN110796100A (zh) * 2019-10-31 2020-02-14 浙江大华技术股份有限公司 步态识别方法、装置、终端及存储装置

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116503943A (zh) * 2023-04-24 2023-07-28 哈尔滨工程大学 一种基于优选动态几何特征的可解释步态识别方法
CN116503943B (zh) * 2023-04-24 2024-06-25 哈尔滨工程大学 一种基于优选动态几何特征的可解释步态识别方法
CN117238032A (zh) * 2023-09-18 2023-12-15 以萨技术股份有限公司 一种步态特征对比方法、存储介质及电子设备

Similar Documents

Publication Publication Date Title
Dagnes et al. Occlusion detection and restoration techniques for 3D face recognition: a literature review
Hassanat et al. Victory sign biometrie for terrorists identification: Preliminary results
WO2021217542A1 (zh) 步态识别方法、装置、终端及存储介质
Caya et al. Recognition of tongue print biometrie using binary robust independent elementary features
Mokni et al. Combining shape analysis and texture pattern for palmprint identification
Ilankumaran et al. Multi-biometric authentication system using finger vein and iris in cloud computing
CN110688950B (zh) 基于深度信息的人脸活体检测方法、装置
Saeed A framework for recognition of facial expression using HOG features
CN114581962A (zh) 一种基于多输入卷积神经网络的高分辨率掌纹识别的方法
Jayaraman et al. Use of geometric features of principal components for indexing a biometric database
Bhargava et al. Fingerprint matching of normalized image based on Euclidean distance
Farmanbar et al. A hybrid approach for person identification using palmprint and face biometrics
JP2013218605A (ja) 画像認識装置、画像認識方法及びプログラム
Kannan et al. VLSI architecture for LGXP texture for face recognition
CN112348008A (zh) 证件信息的识别方法、装置、终端设备及存储介质
Dahea et al. Finger vein recognition system based on multi-algorithm of fusion of Gabor filter and local binary pattern
Mesecan et al. Sift-based iris recognition using sub-segments
KR20160042646A (ko) 얼굴 인식 방법
Abdulbaqi et al. Biometrics detection and recognition based-on geometrical features extraction
Fei et al. Combining enhanced competitive code with compacted ST for 3D palmprint recognition
Babalola et al. Wrist vein recognition by fusion of multiple handcrafted methods
Wang et al. Eyebrow shape analysis by using a modified functional curve procrustes distance
Su et al. Complete Region of Interest for Unconstrained Palmprint Recognition
Kakarash et al. Biometric Iris recognition approach based on filtering techniques
Attia et al. ENSEMBLE OF PREPROCESSING TECHNIQUES FOR 3D PALMPRINT RECOGNITION WITH COLLABORATIVE REPRESENTATION BASED CLASSIFICATION.

Legal Events

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

Ref document number: 20934144

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20934144

Country of ref document: EP

Kind code of ref document: A1