Human body posture estimation algorithm based on deviation prediction in bus compartment scene

The human posture estimation algorithm uses YOLO-like detection and posture templates to address high missed detection rates in bus compartments by accurately capturing human body parts and adjusting for occlusions, enhancing detection speed and accuracy.

CN119068511APending Publication Date: 2024-12-03NORTH CHINA UNIVERSITY OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411101856.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing human posture detection methods in bus compartment scenarios suffer from high missed detection rates due to occlusion, as they rely on bottom-up algorithms that discard insufficient key points, leading to incomplete human skeleton recognition.

Method used

A human posture estimation algorithm using a YOLO-like target detection structure and posture templates to represent human body parts with target rectangular frames, including an inclined angle for the trunk, to accurately capture and adjust for partial occlusions and posture changes, reducing missed detections by refining output results and incorporating offset coefficients.

Benefits of technology

The algorithm significantly improves detection speed and accuracy by dynamically adjusting to posture changes, reducing missed detections and enhancing the network's ability to recognize complete human postures in complex scenes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068511A_ABST
    Figure CN119068511A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of posture detection, and aims to provide a human body posture estimation algorithm based on deviation prediction in a bus compartment scene, comprising the following steps: establishing a posture object set representation form; creating a posture template set; constructing a network of a similar YOLO target detection structure to obtain an estimation network; mapping the attitude template set to an estimation network; obtaining an input picture, and inputting the input picture into the estimation network; the estimation network executes the following steps: segmenting an input picture into a plurality of cells to obtain a cell array; obtaining a target rectangular frame representation of a corresponding attitude object from the input picture and the attitude template set, and determining an offset coefficient of the attitude object relative to each attitude template; obtaining a target rectangular frame representation of the human body target from the input picture and the posture template set, and determining the offset of the human body target relative to the human body target in each posture template; and establishing a representation form of the attitude object set in estimation network output. And the omission ratio of the human body image is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of posture detection, and in particular to a human posture estimation algorithm based on deviation prediction in a bus compartment scenario. Background Art

[0002] At present, human posture detection methods are usually equipped with a bottom-up human posture estimation algorithm. The principle of the bottom-up human posture estimation algorithm is: first detect all the key points in the image, and then use auxiliary information in the post-processing stage to regress and connect the key points belonging to the same person to obtain a complete human skeleton. If reliable regression connection is to be achieved, enough key points of the human target must be detected. If the number of key points judged to belong to a person is less than the specified threshold, these key points are considered to be false detections and are directly discarded without regression connection. This is mainly to reduce network sensitivity and overfitting and improve accuracy. However, in scenarios with high occlusion of human images, many human images will be discarded due to occlusion, resulting in insufficient key points being detected, and the missed detection rate will increase significantly. Summary of the invention

[0003] The purpose of the present invention is to provide a human posture estimation algorithm based on deviation prediction in a bus compartment scene, and the technical problem to be solved is to reduce the missed detection rate of human body images.

[0004] The present invention is achieved through the following technical solutions:

[0005] A human posture estimation algorithm based on deviation prediction in a bus compartment scenario includes the following steps:

[0006] S100, establishing a representation form of a posture object set, wherein the posture object set includes a human body target, a head target, a trunk target, a right hand target, a left hand target, a right foot target, a left foot target, a right elbow target, a left elbow target, a right knee target and a left knee target; the human body target, the head target, the trunk target, the right hand target, the left hand target, the right foot target, the left foot target, the right elbow target, the left elbow target, the right knee target and the left knee target are all represented by a target rectangular frame; wherein the trunk target is represented by a target rectangular frame with an inclined angle;

[0007] S200, creating a posture template set;

[0008] S300, constructing a network with a YOLO-like target detection structure to obtain an estimation network;

[0009] S400, mapping the posture template set to the estimation network;

[0010] S500, obtaining an input image, and inputting the input image into an estimation network;

[0011] The above estimation network performs the following steps:

[0012] S600, dividing the input image into a plurality of cells to obtain a cell array;

[0013] S700, obtaining a target rectangular frame representation of a corresponding posture object from the above-mentioned input image and posture template set, and determining an offset coefficient of the posture object relative to each posture template;

[0014] S800, obtaining a target rectangular frame representation of a human target from the input image and the posture template set, and determining an offset of the human target relative to the human target in each posture template;

[0015] S900. Establish the representation form of the above-mentioned posture object set in the estimation network output. The above-mentioned human body target uses the offset as the output result. The above-mentioned head target, torso target, right hand target, left hand target, right foot target, left foot target, right elbow target, left elbow target, right knee target and left knee target all use the offset coefficient represented by the target rectangular box as the output result.

[0016] By subdividing the human body into multiple specific posture objects (such as the head, torso, limbs and key joints) and using target rectangular boxes (especially the torso target uses a target rectangular box with an inclined angle) to represent it, this method can more accurately capture the various parts of the human body and their relative positions. This detailed division enables the network to more comprehensively detect the key parts of the human body and reduce missed detections due to partial occlusion or posture changes; the posture template set provides a reference framework for the network, allowing the network to learn the typical positional relationship of various parts of the human body under different postures, which helps the network to more accurately match the detected features with the template when recognizing new images, thereby identifying the complete posture of the human body and reducing missed detections due to excessive posture changes; YOLO is a fast and accurate target detection algorithm that can simultaneously predict the positions and categories of multiple targets in a single forward propagation. Applying this algorithm to human posture estimation can significantly improve the detection speed, and due to its end-to-end characteristics, it reduces The error accumulation caused by step-by-step processing helps to reduce the missed detection rate; by calculating the offset coefficient of the posture object relative to the posture template, and the offset of the human target relative to the human target in the template, this method can more finely adjust the detection results to adapt to changes in different postures and scales; this dynamic adjustment capability enables the network to better cope with posture changes in complex scenes, reducing missed detections caused by posture differences; in the output representation form, the human target uses the offset as the output result, while other posture objects use the offset coefficient represented by the target rectangular box as the output result. This design makes the output result contain both position information and posture information, providing a rich information basis for subsequent posture analysis and processing. At the same time, due to the refinement of the output results, the accuracy of posture recognition is also improved and the missed detection rate is reduced.

[0017] Furthermore, the above-mentioned posture object set is represented by the following formula:

[0018]

[0019] Among them, P i Indicates the representation form of posture object set; represents a human target; Indicates the head target; represents the torso target; Indicates right hand target; indicates a left-hand target; Indicates right foot target; Indicates left foot target; indicates the right elbow target; indicates the left elbow target; Indicates right knee target; Indicates left knee target; The target rectangular box representing the human target represents the horizontal coordinate of the center point; The target rectangular box representing the human target represents the vertical coordinate of the center point; The width value of the target rectangular box representation representing the human target; The height value of the target rectangular box representing the human target; The target rectangle representing the head target represents the horizontal coordinate of the center point; The target rectangular box representing the head target represents the vertical coordinate of the center point; The width of the target rectangle representing the head target; The height value of the target rectangle representing the head target; The target rectangle representing the torso target represents the horizontal coordinate of the center point; The target rectangle representing the torso target represents the ordinate of the center point; The width of the target rectangle representing the torso target; The height value of the target rectangle representing the torso target; The tilt angle of the target rectangle representing the torso target; The target rectangle box representing the right-hand target represents the horizontal coordinate of the center point; The target rectangle representing the right-hand target represents the ordinate of the center point; The width of the target rectangle representing the right-hand target; The height value of the target rectangle representing the right-hand target; The target rectangle representing the left-hand target represents the horizontal coordinate of the center point; The target rectangle representing the left-hand target represents the ordinate of the center point; The width of the target rectangle representing the left-hand target; The height value of the target rectangle representation representing the left-hand target; The target rectangle representing the right foot target represents the horizontal coordinate of the center point; The target rectangle representing the right foot target represents the ordinate of the center point; The width value of the target rectangle representing the right foot target; The height value represented by the target rectangle box representing the right foot target; The target rectangle representing the left foot target represents the horizontal coordinate of the center point; The target rectangle representing the left foot target represents the ordinate of the center point; The width value of the target rectangle representing the left foot target; The height value represented by the target rectangle representing the left foot target; The target rectangle representing the right elbow target represents the horizontal coordinate of the center point; The target rectangle representing the right elbow target represents the ordinate of the center point; ε represents a fixed constant; The target rectangle representing the left elbow target represents the horizontal coordinate of the center point; The target rectangle box representing the left elbow target represents the ordinate of the center point; The target rectangle representing the right knee target represents the horizontal coordinate of the center point; The vertical coordinate of the center point of the target rectangle representing the right knee target The target rectangle representing the left knee target represents the horizontal coordinate of the center point; The target rectangle representing the left knee target represents the ordinate of the center point.

[0020] Furthermore, the specific steps of S200 are as follows:

[0021] S210, setting 8 initial posture templates, the 8 initial posture templates are respectively forward standing posture, back standing posture, right standing posture, left standing posture, upright sitting posture, right side sitting posture, left side sitting posture, and incomplete posture;

[0022] S220, acquiring a scene data set, extracting posture objects and posture data from the scene data in the scene data set, and obtaining a posture object set and a posture data set;

[0023] S230, normalizing the posture data using the human body target in the posture object set;

[0024] S240, taking the trunk target, knee target and foot target in the above posture object set as matching objects;

[0025] Setting matching indexes in the eight initial posture templates, the matching indexes including two vertex parameters representing the hips in the torso target, a knee target parameter, and a foot target parameter;

[0026] S250, dividing the posture data set into 8 groups according to the above matching objects and matching indexes, and obtaining parameters of 8 groups of posture object sets;

[0027] S260, respectively calculating the parameter average values ​​of the eight groups of the above posture object sets, and using the parameter average values ​​of the eight groups of posture object sets as posture templates to obtain a posture template set.

[0028] Furthermore, the head target, torso target, right hand target, left hand target, right foot target and left foot target are all represented by two-point target rectangular frames;

[0029] The above-mentioned right elbow target, left elbow target, right knee target and left knee target are all represented by a central target rectangular frame.

[0030] Furthermore, the offset coefficient of the above-mentioned posture object relative to each posture template is determined by the following formula:

[0031]

[0032]

[0033]

[0034]

[0035] Where (cx, cy) represents the center coordinates of the target rectangular box representation of the pose object in the input image; Represents the center point coordinates of the target rectangular box representation of the posture object in the posture template; t cx Represents the estimated value of the horizontal coordinate offset coefficient of the target rectangular box representation of the pose object in the input image by the estimation network; t cy It represents the estimated value of the ordinate offset coefficient of the target rectangular box representation of the pose object in the input image by the estimation network; σ represents the Sigmoid activation function; ρ represents the inverse function of the Sigmoid activation function.

[0036] Furthermore, the offset of the above human target relative to the human target in each posture template is determined by the following formula:

[0037] xx human =2σ(T cx )-0.5+x grid

[0038] cy human =2σ(T cy )-0.5+y grid

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] Among them, (x grid ,y grid ) represents the coordinates of the upper left vertex of the cell array; (cx human ,cy human) represents the center coordinates of the target rectangular box representation of the human target in the input image; w human Indicates the width of the target rectangle representing the human target in the input image; h human Represents the height value of the target rectangular box representation of the human target in the input image; Indicates the width value of the target rectangular box representation of the human target in the posture template; Indicates the height value of the target rectangular box representation of the human target in the posture template; T cx T represents the estimated value of the horizontal coordinate offset of the target rectangular box representing the human target in the input image by the estimation network; cy T represents the estimated value of the vertical coordinate offset of the target rectangular box representing the human target in the input image by the estimation network; w T represents the estimated value of the width ratio of the target rectangular box representation of the human target in the input image by the estimation network; h Represents the estimated value of the height ratio of the target rectangular box representing the human target in the input image by the estimation network.

[0046] Furthermore, the representation form of the above posture object set in the estimation network output is as follows:

[0047]

[0048] in, Represents the representation form of the pose object set in the estimation network output; and They represent human target confidence, head target confidence, torso target confidence, right elbow target confidence, right hand target confidence, left elbow target confidence, left hand target confidence, right knee target confidence, right foot target confidence, left knee target confidence, and left foot target confidence respectively; and They respectively represent the estimated values ​​of the horizontal coordinate offset coefficients of the target rectangular boxes representing the right elbow target, the left elbow target, the right knee target, and the left knee target, which are estimated by the estimation network; and They respectively represent the estimated values ​​of the vertical coordinate offset coefficients of the target rectangular boxes representing the right elbow target, the left elbow target, the right knee target, and the left knee target, which are estimated by the estimation network; and They respectively represent the estimated values ​​of the horizontal and vertical coordinate offsets of the target rectangular box representing the human target, as well as the width and height multiples; and Respectively represent the estimated values ​​of the horizontal coordinate offset coefficient of the upper left vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network; and They respectively represent the estimated values ​​of the vertical coordinate offset coefficient of the upper left vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network; and They respectively represent the estimated values ​​of the horizontal coordinate offset coefficient of the lower right vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network; and They respectively represent the estimated values ​​of the vertical coordinate offset coefficient of the lower right vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network.

[0049] Furthermore, it is necessary to establish isolated object representation in the estimation network;

[0050] The isolated object representation form is as follows:

[0051]

[0052] in, Indicates isolated object representation form; The coordinates of the center point of the target rectangular box representation representing the isolated object; When the isolated object is a head target, a right hand target, a left hand target, a right foot target, or a left foot target, the width of the target rectangle is represented by the target; When the isolated object is a head target, a right hand target, a left hand target, a right foot target, or a left foot target, the height value represented by the target rectangle; When the isolated object is a torso target, the coordinates of the center point represented by the target rectangle; Indicates that when the isolated object is a torso target, the width value represented by the target rectangle; When the isolated object is a torso target, the height value represented by the target rectangle; Indicates the tilt angle of the target rectangle when the isolated object is a torso target.

[0053] Furthermore, the estimated value of the isolated object is calculated as follows:

[0054]

[0055] in, represents the estimated value of isolated objects; represents the target confidence of isolated objects; and The target rectangular box representing the isolated object represents the horizontal coordinate, vertical coordinate, width value and height value of the center point respectively; The target rectangle representing the isolated object represents the tilt angle. This value is meaningful when the isolated object is a torso target, otherwise the value is 0; and They respectively represent the estimated values ​​of each classification score when the isolated objects are head targets, torso targets, right elbow targets, right hand targets, left elbow targets, left hand targets, right knee targets, right foot targets, left knee targets and left foot targets.

[0056] Furthermore, the pose object is corrected by using the estimated value of the isolated object through the above estimation network.

[0057] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0058] By subdividing the human body into multiple specific posture objects (such as the head, torso, limbs and key joints) and using target rectangular boxes (especially the torso target uses a target rectangular box with an inclined angle) to represent it, this method can more accurately capture the various parts of the human body and their relative positions. This detailed division enables the network to more comprehensively detect the key parts of the human body and reduce missed detections due to partial occlusion or posture changes; the posture template set provides a reference framework for the network, allowing the network to learn the typical positional relationship of various parts of the human body under different postures, which helps the network to more accurately match the detected features with the template when recognizing new images, thereby identifying the complete posture of the human body and reducing missed detections due to excessive posture changes; YOLO is a fast and accurate target detection algorithm that can simultaneously predict the positions and categories of multiple targets in a single forward propagation. Applying this algorithm to human posture estimation can significantly improve the detection speed, and due to its end-to-end characteristics, it reduces The error accumulation caused by step-by-step processing helps to reduce the missed detection rate; by calculating the offset coefficient of the posture object relative to the posture template, and the offset of the human target relative to the human target in the template, this method can more finely adjust the detection results to adapt to changes in different postures and scales; this dynamic adjustment capability enables the network to better cope with posture changes in complex scenes, reducing missed detections caused by posture differences; in the output representation form, the human target uses the offset as the output result, while other posture objects use the offset coefficient represented by the target rectangular box as the output result. This design makes the output result contain both position information and posture information, providing a rich information basis for subsequent posture analysis and processing. At the same time, due to the refinement of the output results, the accuracy of posture recognition is also improved and the missed detection rate is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0060] Figure 1 As the main flow chart;

[0061] Figure 2 Schematic diagram of 8 posture templates;

[0062] Figure 3 Schematic diagram of mapping original coordinates to offset coefficients provided for Example 3;

[0063] Figure 4 A schematic diagram of mapping the original coordinates to the offset provided in Example 6;

[0064] Figure 5 Schematic diagram of the estimated network structure. DETAILED DESCRIPTION

[0065] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0066] Example 1

[0067] Combination Figure 1 , this embodiment 1 provides a human posture estimation algorithm based on deviation prediction in a bus compartment scenario, comprising the following steps:

[0068] S100, establishing a representation form of a posture object set, wherein the posture object set includes a human body target, a head target, a trunk target, a right hand target, a left hand target, a right foot target, a left foot target, a right elbow target, a left elbow target, a right knee target and a left knee target; the human body target, the head target, the trunk target, the right hand target, the left hand target, the right foot target, the left foot target, the right elbow target, the left elbow target, the right knee target and the left knee target are all represented by a target rectangular frame; wherein the trunk target is represented by a target rectangular frame with an inclined angle;

[0069] S200, creating a posture template set;

[0070] S300, constructing a network with a YOLO-like target detection structure to obtain an estimation network;

[0071] S400, mapping the posture template set to the estimation network;

[0072] S500, obtaining an input image, and inputting the input image into an estimation network;

[0073] The above estimation network performs the following steps:

[0074] S600, dividing the input image into a plurality of cells to obtain a cell array;

[0075] S700, obtaining a target rectangular frame representation of a corresponding posture object from the above-mentioned input image and posture template set, and determining an offset coefficient of the posture object relative to each posture template;

[0076] S800, obtaining a target rectangular frame representation of a human target from the input image and the posture template set, and determining an offset of the human target relative to the human target in each posture template;

[0077] S900. Establish the representation form of the above-mentioned posture object set in the estimation network output. The above-mentioned human body target uses the offset as the output result. The above-mentioned head target, torso target, right hand target, left hand target, right foot target, left foot target, right elbow target, left elbow target, right knee target and left knee target all use the offset coefficient represented by the target rectangular box as the output result.

[0078] First, the posture features are represented in the form of targets, and the detection method is similar to target detection. Compared with the estimation network that uses key points to represent information, key part targets have stronger local information than key points and are easier to detect. Because the human body target is an indivisible basic element, compared with the key points as indivisible basic elements, the representation form of the posture object represented by the human body target and its key part target and its connection relationship does not need to consider the abandonment problem during regression connection, and has a stronger ability to obtain complete posture information, a higher recall rate, and a lower missed detection rate.

[0079] Secondly, the network of this algorithm estimates the offset of the target relative to the template. Its estimation form is similar to direct coordinate estimation, but after introducing the posture template, it can effectively pass the spatial constraint information of the key parts of the template to the network to make up for the weak global information understanding ability of direct coordinate estimation. At the same time, the network of the YOLO-like detection structure does not have a large number of parameters like the detection form of heat map estimation.

[0080] Finally, since the posture features in the algorithm already include the range information of key parts, when human posture features, human target features, and part target features are needed simultaneously in the future, there is no need to use an additional target detection algorithm to obtain accurate, non-regression fitting prediction bounding boxes, which can achieve more information detection.

[0081] Example 1: Abnormal behavior detection in bus compartment scenes is of great significance for public safety and danger warning, and the premise of abnormal behavior recognition is to first extract human posture feature information based on the collected images or videos; when investigating bus compartment scenes, the human target images are often severely occluded. At this time, the traditional human posture estimation algorithm is used, and even partially occluded human targets are often missed, resulting in a high missed detection rate. To solve this problem, the method of the present invention is adopted: a posture template is created according to the characteristics of the bus compartment scene, and the offset of the target posture compared to the template is estimated using a YOLO-like detection architecture network, and then regressed into an estimated posture object through an inverse mapping method. The end-to-end one-step detection of the YOLO-like detection architecture network can avoid missed detections caused by artificial screening strategies, and can obtain the spatial constraint relationship between key parts through the template, avoiding the weak global information perception ability of direct coordinate estimation or the excessive amount of operation parameters of heat map regression. At the same time, the algorithm can realize the synchronous detection of posture features and key parts and human target features, enriching the detection information of the algorithm.

[0082] Example 2

[0083] On the basis of Example 1, a human posture representation form based on a key target set (i.e., a posture object set representation form) is used to represent posture objects, so that posture objects can represent more information and are convenient for estimation using a YOLO-like architecture. To supplement the inclination information of the human trunk, the trunk target is represented by a target rectangular box with an inclination angle; since the local information of the right elbow target, the left elbow target, the right knee target, and the left knee target is not strong, their range information does not play a big role in expressing the posture information, therefore, the width and height values ​​of their target bounding boxes take a fixed small-scale positive value, that is, a fixed small square box is used to represent these four targets.

[0084] The above posture object set is represented by the following formula:

[0085]

[0086] Among them, P i Indicates the representation form of posture object set; represents a human target; Indicates the head target; represents the torso target; Indicates right hand target; indicates a left-hand target; Indicates right foot target; Indicates left foot target; indicates the right elbow target; indicates the left elbow target; Indicates right knee target; Indicates left knee target; The target rectangular box representing the human target represents the horizontal coordinate of the center point; The target rectangular box representing the human target represents the vertical coordinate of the center point; The width value of the target rectangular box representation representing the human target; The height value of the target rectangular box representing the human target; The target rectangle representing the head target represents the horizontal coordinate of the center point; The target rectangular box representing the head target represents the vertical coordinate of the center point; The width of the target rectangle representing the head target; The height value of the target rectangle representing the head target; The target rectangle representing the torso target represents the horizontal coordinate of the center point; The target rectangle representing the torso target represents the ordinate of the center point; The width of the target rectangle representing the torso target; The height value of the target rectangle representing the torso target; The tilt angle of the target rectangle representing the torso target; The target rectangle box representing the right-hand target represents the horizontal coordinate of the center point; The target rectangle representing the right-hand target represents the ordinate of the center point; The width of the target rectangle representing the right-hand target; The height value of the target rectangle representing the right-hand target; The target rectangle representing the left-hand target represents the horizontal coordinate of the center point; The target rectangle representing the left-hand target represents the ordinate of the center point; The width of the target rectangle representing the left-hand target; The height value of the target rectangle representation representing the left-hand target; The target rectangle representing the right foot target represents the horizontal coordinate of the center point; The target rectangle representing the right foot target represents the ordinate of the center point; The width value of the target rectangle representing the right foot target; The height value represented by the target rectangle box representing the right foot target; The target rectangle representing the left foot target represents the horizontal coordinate of the center point; The target rectangle representing the left foot target represents the ordinate of the center point; The width value of the target rectangle representing the left foot target; The height value represented by the target rectangle representing the left foot target; The target rectangle representing the right elbow target represents the horizontal coordinate of the center point; The target rectangle representing the right elbow target represents the ordinate of the center point; ε represents a fixed constant; The target rectangle representing the left elbow target represents the horizontal coordinate of the center point; The target rectangle box representing the left elbow target represents the ordinate of the center point; The target rectangle representing the right knee target represents the horizontal coordinate of the center point; The vertical coordinate of the center point of the target rectangle representing the right knee target The target rectangle representing the left knee target represents the horizontal coordinate of the center point; The target rectangle representing the left knee target represents the ordinate of the center point.

[0087] Example 3

[0088] Based on any of the above embodiments, the specific steps of S200 are as follows:

[0089] S210, setting 8 initial posture templates, the 8 initial posture templates are respectively forward standing posture, back standing posture, right standing posture, left standing posture, upright sitting posture, right side sitting posture, left side sitting posture, and incomplete posture;

[0090] S220, acquiring a scene data set, extracting posture objects and posture data from the scene data in the scene data set, and obtaining a posture object set and a posture data set;

[0091] S230, normalizing the posture data using the human body target in the posture object set;

[0092] S240, taking the trunk target, knee target and foot target in the above posture object set as matching objects;

[0093] Setting matching indexes in the eight initial posture templates, the matching indexes including two vertex parameters representing the hips in the torso target, a knee target parameter, and a foot target parameter;

[0094] S250, dividing the posture data set into 8 groups according to the above matching objects and matching indexes, and obtaining parameters of 8 groups of posture object sets;

[0095] S260, respectively calculating the parameter average values ​​of the eight groups of the above posture object sets, and using the parameter average values ​​of the eight groups of posture object sets as posture templates to obtain a posture template set.

[0096] Example 2: In a bus compartment scenario, passengers often sit or stand when riding a bus; therefore, consider using common postures in the statistical scene as posture templates, and using the network to estimate the offset coefficient between the target posture and the posture template to estimate the posture object. The template information can be passed to the estimation network through the subsequent template deviation mapping-based YOLO-like detection architecture to provide initial spatial constraints for the estimated posture object.

[0097] According to the common human postures in the bus scene, 8 initial posture templates are set: forward standing posture, back standing posture, right standing posture, left standing posture, upright sitting posture, right side sitting posture, left side sitting posture, and incomplete posture. The incomplete posture is used to represent the human posture that has not fully entered the picture.

[0098] In order to make the posture template fully express the constraint relationship between human targets, the initial posture template is counted through the bus scene data set. The posture data in the data set is normalized relative to the human target to avoid the influence of inconsistent human image scale; the two vertices representing the hips in the torso target, the knee target and the foot target are taken as matching objects, and the OKS (class target key point score) of these five key points is used as the matching index to divide the posture data in the data set into eight corresponding groups. Finally, the average parameter value of the eight groups of posture objects is counted as the posture template, such as Figure 2 shown.

[0099] Example 4

[0100] Based on any of the above embodiments, the head target, torso target, right hand target, left hand target, right foot target and left foot target are all represented by a two-point target rectangular frame, and the two-point type uses the upper left vertex and the lower right vertex to express the position and range of the rectangular frame;

[0101] The above-mentioned right elbow target, left elbow target, right knee target and left knee target are all represented by a central target rectangular frame, and the central frame uses the center point and width and height values ​​to express the position and range of the rectangular frame.

[0102] Example 5

[0103] On the basis of Example 4, in order to realize the template offset prediction of the posture object, the YOLO target detection structure is used to construct the estimation network. The estimation network divides the input image into S×S cells, and each cell is responsible for predicting the offset coefficient of the posture object whose center falls within the cell relative to the 8 posture templates. Here, the estimation of the offset coefficient is used instead of the estimation of the offset, and the two-point rectangular frame representation is used to replace the commonly used central rectangular frame representation. These two changes can enable the network to naturally constrain the estimated human body part target within the human body target range, increase the initial perception ability of the network to accelerate the training convergence.

[0104] Determine the offset coefficient of the above posture object relative to each posture template, the formula is as follows:

[0105]

[0106]

[0107]

[0108]

[0109] Where (cx, cy) represents the center coordinates of the target rectangular box representation of the pose object in the input image; Represents the center point coordinates of the target rectangular box representation of the posture object in the posture template; t cx Represents the estimated value of the horizontal coordinate offset coefficient of the target rectangular box representation of the pose object in the input image by the estimation network; t cy It represents the estimated value of the vertical coordinate offset coefficient of the target rectangular box representation of the pose object in the input image by the estimation network; σ represents the Sigmoid activation function; ρ represents the inverse function of the Sigmoid activation function. cx and t cy is the output of the estimation network, with a value range of (-∞,+∞); 2σ(t cx )-1 and 2σ(t cy )-1 is the offset coefficient after the value range is mapped to the (-1, +1) interval.

[0110] Example 3: Mapping method of original coordinates to offset coefficients: Figure 3 As shown, is the template target, and its parameters are the values ​​normalized relative to the human target; and They are The upper left vertex, lower right vertex and center point of ; O is the corresponding target (that is, the posture object corresponding to the template target on the input image, the input image can be either a posture template or a scene image), and its parameters are also the values ​​normalized relative to the human target; (x1, y1), (x2, y2) and (cx, xy) are the upper left vertex, lower right vertex and center point of O respectively.

[0111] Example 6

[0112] On the basis of Example 5, for human targets that cannot be represented by offset coefficients, an estimation network is used to estimate the offset of the human target in the posture template. The mapping method is as follows: Figure 4 And the following formula shows;

[0113] Determine the offset of the above human target relative to the human target in each posture template. The formula is as follows:

[0114] xx human =2σ(T cx )-0.5+x grid

[0115] cy human =2σ(T cy )-0.5+y grid

[0116]

[0117]

[0118]

[0119]

[0120]

[0121]

[0122] Among them, (x grid ,y grid ) represents the coordinates of the upper left vertex of the cell array; (cx human ,cy human ) represents the center coordinates of the target rectangular box representation of the human target in the input image; w human Indicates the width of the target rectangle representing the human target in the input image; h human Represents the height value of the target rectangular box representation of the human target in the input image; Indicates the width value of the target rectangular box representation of the human target in the posture template; Indicates the height value of the target rectangular box representation of the human target in the posture template; T cx T represents the estimated value of the horizontal coordinate offset of the target rectangular box representing the human target in the input image by the estimation network; cy T represents the estimated value of the vertical coordinate offset of the target rectangular box representing the human target in the input image by the estimation network; w T represents the estimated value of the width ratio of the target rectangular box representation of the human target in the input image by the estimation network; h Represents the estimated value of the height ratio of the target rectangular box representing the human target in the input image by the estimation network.

[0123] T cx With T cy is the output of the estimation network, with a value range of (-∞,+∞); 2σ(T cx)-0.5 and 2σ(T cy )-0.5 is the offset after the value range is mapped to the interval (-0.5, +1.5), that is, the center point range of the target estimated by the cell is within the area twice the cell range; T w With T h are the estimated values ​​of the width and height ratios of the estimation network, respectively, and are the outputs of the estimation network, with a value range of (-∞,+∞); and is the ratio of the width and height after the value range is mapped to the interval (0,4); considering that the spatial range of the bus scene is limited and the target scale difference is not large, the 4-fold scale of the target rectangle of the human target can fully express the target in the scene.

[0124] Example 7

[0125] Based on any of the above embodiments, the representation form of the above-mentioned posture object set in the estimation network output includes information of 11 targets in total, including human body target, head target, torso target, right elbow target, right hand target, left elbow target, left hand target, right knee target, right foot target, left knee target, and left foot target; among which the human body target uses the offset as output information, and the head target, torso target, right hand target, left hand target, right foot target, and left foot target use the offset coefficient represented by a two-point target rectangular box as output information; the right elbow target, left elbow target, right knee target, and left knee target use the offset coefficient represented by a central target rectangular box as output information. The local information of these targets is not obvious, so it is meaningless to estimate the range of their target rectangular boxes. This article uses a fixed-size square to represent them, and does not estimate the width and height values ​​during network estimation.

[0126] The representation form of the above posture object set in the estimation network output is as follows:

[0127]

[0128] in, Represents the representation form of the pose object set in the estimation network output; and They represent human target confidence, head target confidence, torso target confidence, right elbow target confidence, right hand target confidence, left elbow target confidence, left hand target confidence, right knee target confidence, right foot target confidence, left knee target confidence, and left foot target confidence respectively; and They respectively represent the estimated values ​​of the horizontal coordinate offset coefficients of the target rectangular boxes representing the right elbow target, the left elbow target, the right knee target, and the left knee target, which are estimated by the estimation network; and They respectively represent the estimated values ​​of the vertical coordinate offset coefficients of the target rectangular boxes representing the right elbow target, the left elbow target, the right knee target, and the left knee target, which are estimated by the estimation network; and They respectively represent the estimated values ​​of the horizontal and vertical coordinate offsets of the target rectangular box representing the human target, as well as the width and height multiples; and Respectively represent the estimated values ​​of the horizontal coordinate offset coefficient of the upper left vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network; and They respectively represent the estimated values ​​of the vertical coordinate offset coefficient of the upper left vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network; and They respectively represent the estimated values ​​of the horizontal coordinate offset coefficient of the lower right vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network; and They respectively represent the estimated values ​​of the vertical coordinate offset coefficient of the lower right vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target of the estimation network.

[0129] Example 8

[0130] On the basis of any of the above embodiments, in order to improve the algorithm's perception of local information, the estimation network is responsible for estimating n isolated objects used to correct the posture object while estimating the posture object; when the isolated object is a head target, a right hand target, a left hand target, a right foot target, or a left foot target, the isolated object representation will include the horizontal coordinate and vertical coordinate of the center point of the target rectangular box and the width and height value of the target rectangular box; when the isolated target object is a torso target, the isolated object representation will additionally include the inclination angle of the target rectangular box; when the isolated target object is a right elbow target, a left elbow target, a right knee target, or a left knee target, the width and height values ​​of its target rectangular box will be fixed values ​​ε.

[0131] The isolated object representation form is as follows:

[0132]

[0133] in, Indicates isolated object representation form; The coordinates of the center point of the target rectangular box representation representing the isolated object; When the isolated object is a head target, a right hand target, a left hand target, a right foot target, or a left foot target, the width of the target rectangle is represented by the target; When the isolated object is a head target, a right hand target, a left hand target, a right foot target, or a left foot target, the height value represented by the target rectangle; When the isolated object is a torso target, the coordinates of the center point represented by the target rectangle; Indicates that when the isolated object is a torso target, the width value represented by the target rectangle; Indicates that when the isolated object is a torso target, the height value represented by the target rectangle; Indicates the tilt angle of the target rectangle when the isolated object is a torso target.

[0134] In a specific embodiment, in order to supplement the estimation network's perception ability of local information, the estimation network is responsible for estimating n isolated target objects that do not participate in the expression of the posture object while estimating the posture object, so as to correct the part targets in the posture object estimated by the network.

[0135] Calculate the estimated value of the above isolated objects, the formula is as follows:

[0136]

[0137] in, represents the estimated value of isolated objects; represents the target confidence of isolated objects; and The target rectangular box representing the isolated object represents the horizontal coordinate, vertical coordinate, width value and height value of the center point respectively; The target rectangle representing the isolated object represents the tilt angle. This value is meaningful when the isolated object is a torso target, otherwise the value is 0; and They respectively represent the estimated values ​​of each classification score when the isolated objects are head targets, torso targets, right elbow targets, right hand targets, left elbow targets, left hand targets, right knee targets, right foot targets, left knee targets and left foot targets.

[0138] In a specific embodiment, the pose object is corrected by using the estimated value of the isolated object through the above-mentioned estimation network.

[0139] The above estimation network structure is divided into three parts (such as Figure 5 As shown): pre-processing part α, estimation network and post-processing

[0140] Pre-processing part: In order to reduce the negative impact of the uneven scale of the input image and the changes in light and shadow on the network estimation, the network performs preliminary processing on the input image I through the pre-processing part α(I) so that the subsequent network can better estimate. This part mainly includes two operations: the first is the scale normalization processing a1(I), which is responsible for normalizing the input image I to a three-channel RGB image I′ of 448×448×3 to align the subsequent network input scale. The second is the histogram equalization processing a2(I′), which uses an adaptive histogram equalization algorithm to balance the image contrast to reduce the negative impact of light and shadow on subsequent network estimation.

[0141] Estimation network: Human pose estimation network based on template offset prediction It is a template offset coefficient prediction network based on a YOLOv5-like structure, such as Figure 5 As shown. The network adopts a three-stage network structure of backbone network-neck network-detection head network; the backbone network adopts a Focus module to achieve low-cost downsampling, and performs feature sampling through CBS convolution module and CSP residual module. The network in this chapter sets the input image scale to 448×448×3, and obtains the minimum 7×7 scale feature after six downsamplings; the neck network adopts FPN+PAN architecture, the FPN structure conveys strong semantic features through a top-down upsampling path, and the PAN structure conveys strong positioning features through a bottom-up downsampling path, realizing deep perception of image features. Feature sampling is performed through the CBS convolution module, upsampling is performed through the Upsample module, and Concat enhances feature depth; the detection head network uses 1×1 convolution to achieve fully connected mapping, and the network output is reshaped into S×S×(48·8+16·n inde ). S×S is the scale of the detection head, that is, the scale of the segmentation cell. The estimation network is responsible for detecting posture objects and isolated objects with small scale changes. There are no targets with relatively small scales among these objects. Therefore, the estimation network contains three detection head networks of 28×28, 14×14, and 7×7. The 7×7 and 14×14 detection head networks are responsible for detecting larger posture objects, human targets, torso targets, etc., and the 28×28 detection head network is responsible for detecting smaller hand targets and foot targets.

[0142] Post-processing part: responsible for further processing of the network output, including competition module φ1 and correction module Competition Module Eliminate the estimated result S by competitive screening pose With S obj The redundant detection results in mainly include confidence competition and non-maximum suppression competition, and finally obtain the filtered estimation result S pose′ With S obj′ . Correction module The isolated objects S estimated by the network obj′ To correct the estimated pose object S pose′ , to supplement the algorithm's detection capabilities for key parts with strong local information.

[0143] The estimation network is trained as follows:

[0144] The estimation network is trained on the bus compartment dataset to acquire the ability to estimate human targets and human posture objects. Its loss function As shown in formula (1), it is divided into two parts: pose object loss Object loss with isolated objects

[0145]

[0146]

[0147]

[0148]

[0149]

[0150]

[0151]

[0152] Among them, the pose object loss As shown in formula (2), it includes the human target positioning loss within the pose object Human Target Confidence Loss Key location target loss and key part target confidence loss and The CIOU loss form is used to train the network's ability to estimate the target rectangle in the pose object, as shown in formula (4) and formula (5). Since the human target in the pose object is responsible for the normalization reference of other target coordinates mapped into offset coefficients, its rectangular box positioning is more important, so it is independently assigned weights to make it converge first. and The binary cross entropy loss is used to balance the influence of positive and negative samples and is responsible for training the network's ability to estimate the target confidence, as shown in formula (7) and formula (8) respectively.

[0153] Isolated Object Loss As shown in formula (3), it includes the isolated object localization loss Isolated Object Confidence Loss With isolated object classification score loss As shown in formula (9), the CIOU loss form is adopted to train the network's ability to estimate the rectangular box of isolated objects. and As shown in formula (10) and formula (11), a binary cross entropy loss is used to balance the influence of positive and negative samples, which is responsible for training the network's ability to estimate the confidence and classification scores of isolated objects.

[0154] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A human posture estimation algorithm based on deviation prediction in bus compartment scenarios, It is characterized in that The following steps are involved: S100, establishing a representation form of a posture object set, wherein the posture object set includes a human body target, a head target, a trunk target, a right hand target, a left hand target, a right foot target, a left foot target, a right elbow target, a left elbow target, a right knee target and a left knee target; the human body target, the head target, the trunk target, the right hand target, the left hand target, the right foot target, the left foot target, the right elbow target, the left elbow target, the right knee target and the left knee target are all represented by a target rectangular frame; wherein the trunk target is represented by a target rectangular frame with an inclined angle; S200, creating a posture template set; S300, constructing a network with a YOLO-like target detection structure to obtain an estimation network; S400, mapping the posture template set to the estimation network; S500, obtaining an input image, and inputting the input image into an estimation network; The estimation network performs the following steps: S600, dividing the input image into a plurality of cells to obtain a cell array; S700, obtaining a target rectangular frame representation of a corresponding posture object from the input image and the posture template set, and determining an offset coefficient of the posture object relative to each posture template; S800, obtaining a target rectangular frame representation of a human target from the input image and the posture template set, and determining an offset of the human target relative to the human target in each posture template; S900. Establish a representation form of the posture object set in the estimation network output, the human body target uses an offset as an output result, and the head target, torso target, right hand target, left hand target, right foot target, left foot target, right elbow target, left elbow target, right knee target and left knee target all use an offset coefficient represented by a target rectangular box as an output result.

2. According to claim 1, a human posture estimation algorithm based on deviation prediction in a bus compartment scene, It is characterized in that The representation form of the posture object set is as follows: Among them, P i Indicates the representation form of posture object set; represents a human target; Indicates the head target; represents the torso target; Indicates right hand target; indicates a left-hand target; Indicates right foot target; Indicates left foot target; indicates the right elbow target; indicates the left elbow target; Indicates right knee target; Indicates left knee target; The target rectangular box representing the human target represents the horizontal coordinate of the center point; The target rectangular box representing the human target represents the vertical coordinate of the center point; The width value of the target rectangular box representation representing the human target; The height value of the target rectangular box representing the human target; The target rectangle representing the head target represents the horizontal coordinate of the center point; The target rectangular box representing the head target represents the vertical coordinate of the center point; The width of the target rectangle representing the head target; The height value of the target rectangle representing the head target; The target rectangle representing the torso target represents the horizontal coordinate of the center point; The target rectangle representing the torso target represents the ordinate of the center point; The width of the target rectangle representing the torso target; The height value of the target rectangle representing the torso target; The tilt angle of the target rectangle representing the torso target; The target rectangle box representing the right-hand target represents the horizontal coordinate of the center point; The target rectangle representing the right-hand target represents the vertical coordinate of the center point; The width of the target rectangle representing the right-hand target; The height value of the target rectangle representing the right-hand target; The target rectangle representing the left-hand target represents the horizontal coordinate of the center point; The target rectangle representing the left-hand target represents the ordinate of the center point; The width of the target rectangle representing the left-hand target; The height value of the target rectangle representation representing the left-hand target; The target rectangle representing the right foot target represents the horizontal coordinate of the center point; The target rectangle representing the right foot target represents the ordinate of the center point; The width value of the target rectangle representing the right foot target; The height value represented by the target rectangle box representing the right foot target; The target rectangle representing the left foot target represents the horizontal coordinate of the center point; The target rectangle representing the left foot target represents the ordinate of the center point; The width value of the target rectangle representing the left foot target; The height value represented by the target rectangle representing the left foot target; The target rectangle representing the right elbow target represents the horizontal coordinate of the center point; The target rectangle representing the right elbow target represents the ordinate of the center point; ε represents a fixed constant; The target rectangle representing the left elbow target represents the horizontal coordinate of the center point; The target rectangle box representing the left elbow target represents the ordinate of the center point; The target rectangle representing the right knee target represents the horizontal coordinate of the center point; The vertical coordinate of the center point of the target rectangle representing the right knee target The target rectangle representing the left knee target represents the horizontal coordinate of the center point; The target rectangle representing the left knee target represents the ordinate of the center point.

3. According to claim 1, a human posture estimation algorithm based on deviation prediction in a bus compartment scene, It is characterized in that The specific steps of S200 are as follows: S210, setting 8 initial posture templates, wherein the 8 initial posture templates are respectively a forward standing posture, a back-facing standing posture, a right-facing standing posture, a left-facing standing posture, a sitting posture, a right-facing side-sitting posture, a left-facing side-sitting posture, and a deformed posture; S220, acquiring a scene data set, extracting gesture objects and gesture data from the scene data in the scene data set, and obtaining a gesture object set and a gesture data set; S230, normalizing the posture data using the human body target in the posture object set; S240, taking the trunk target, knee target and foot target in the posture object set as matching objects; Setting matching indexes in the eight initial posture templates, the matching indexes including two vertex parameters representing the hips in the torso target, a knee target parameter, and a foot target parameter; S250, dividing the posture data set into 8 groups according to the matching objects and matching indexes, and obtaining parameters of 8 groups of posture object sets; S260, respectively calculating the parameter average values ​​of the eight groups of posture object sets, and using the parameter average values ​​of the eight groups of posture object sets as posture templates to obtain a posture template set.

4. According to claim 1, a human posture estimation algorithm based on deviation prediction in a bus compartment scene, It is characterized in that The head target, torso target, right hand target, left hand target, right foot target and left foot target are all represented by two-point target rectangular frames; The right elbow target, left elbow target, right knee target and left knee target are all represented by a central target rectangular frame.

5. According to claim 4, a human posture estimation algorithm based on deviation prediction in a bus compartment scene, It is characterized in that Determine the offset coefficient of the posture object relative to each posture template, the formula is as follows: Where (cx, cy) represents the center coordinates of the target rectangular box representation of the pose object in the input image; Represents the center point coordinates of the target rectangular box representation of the posture object in the posture template; t cx Represents the estimated value of the horizontal coordinate offset coefficient of the target rectangular box representation of the pose object in the input image by the estimation network; t cy It represents the estimated value of the ordinate offset coefficient of the target rectangular box representation of the pose object in the input image by the estimation network; σ represents the Sigmoid activation function; ρ represents the inverse function of the Sigmoid activation function.

6. According to claim 4, a human posture estimation algorithm based on deviation prediction in a bus compartment scene, It is characterized in that Determine the offset of the human target relative to the human target in each posture template, the formula is as follows: cx human =2σ(T cx )-0.5+x grid cy human =2σ(T cy )-0.5+y grid Among them, (x grid ,y grid ) represents the coordinates of the upper left vertex of the cell array; (cx human ,cy human ) represents the center coordinates of the target rectangular box representation of the human target in the input image; w human Indicates the width of the target rectangle representing the human target in the input image; h human Represents the height value of the target rectangular box representation of the human target in the input image; Indicates the width value of the target rectangular box representation of the human target in the posture template; Indicates the height value of the target rectangular box representation of the human target in the posture template; T cx T represents the estimated value of the horizontal coordinate offset of the target rectangular box representing the human target in the input image by the estimation network; cy T represents the estimated value of the vertical coordinate offset of the target rectangular box representing the human target in the input image by the estimation network; w T represents the estimated value of the width ratio of the target rectangular box representation of the human target in the input image by the estimation network; h Represents the estimated value of the height ratio of the target rectangular box representing the human target in the input image by the estimation network.

7. According to claim 1, a human posture estimation algorithm based on deviation prediction in a bus compartment scene, It is characterized in that The representation form of the posture object set in the estimation network output is as follows: in, Represents the representation form of the pose object set in the estimation network output; and They represent human target confidence, head target confidence, torso target confidence, right elbow target confidence, right hand target confidence, left elbow target confidence, left hand target confidence, right knee target confidence, right foot target confidence, left knee target confidence, and left foot target confidence respectively; and They respectively represent the estimated values ​​of the horizontal coordinate offset coefficients of the target rectangular boxes representing the right elbow target, the left elbow target, the right knee target, and the left knee target, which are estimated by the estimation network; and They respectively represent the estimated values ​​of the vertical coordinate offset coefficients of the target rectangular boxes representing the right elbow target, the left elbow target, the right knee target, and the left knee target, which are estimated by the estimation network; and They respectively represent the estimated values ​​of the horizontal and vertical coordinate offsets of the target rectangular box representing the human target, as well as the width and height multiples; and Respectively represent the estimated values ​​of the horizontal coordinate offset coefficient of the upper left vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network; and They respectively represent the estimated values ​​of the vertical coordinate offset coefficient of the upper left vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network; and They respectively represent the estimated values ​​of the horizontal coordinate offset coefficient of the lower right vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target by the estimation network; and They respectively represent the estimated values ​​of the vertical coordinate offset coefficient of the lower right vertex of the target rectangular box representation of the head target, torso target, right hand target, left hand target, right foot target and left foot target of the estimation network.

8. According to claim 1, a human posture estimation algorithm based on deviation prediction in a bus compartment scene, It is characterized in that It is also necessary to establish isolated object representation in the estimation network; The isolated object representation form is as follows: in, Indicates isolated object representation form; The coordinates of the center point of the target rectangular box representation representing the isolated object; When the isolated object is a head target, a right hand target, a left hand target, a right foot target, or a left foot target, the width of the target rectangle is represented by the target; When the isolated object is a head target, a right hand target, a left hand target, a right foot target, or a left foot target, the height value represented by the target rectangle; When the isolated object is a torso target, the coordinates of the center point represented by the target rectangle; Indicates that when the isolated object is a torso target, the width value represented by the target rectangle; Indicates that when the isolated object is a torso target, the height value represented by the target rectangle; Indicates the tilt angle of the target rectangle when the isolated object is a torso target.

9. According to claim 8, a human posture estimation algorithm based on deviation prediction in a bus compartment scene, It is characterized in that The estimated value of the isolated object is calculated as follows: in, represents the estimated value of isolated objects; represents the target confidence of isolated objects; and The target rectangular box representing the isolated object represents the horizontal coordinate, vertical coordinate, width value and height value of the center point respectively; The target rectangle representing the isolated object represents the tilt angle. This value is meaningful when the isolated object is a torso target, otherwise the value is 0; and They respectively represent the estimated values ​​of each classification score when the isolated objects are head targets, torso targets, right elbow targets, right hand targets, left elbow targets, left hand targets, right knee targets, right foot targets, left knee targets and left foot targets.

10. The human body posture estimation algorithm based on deviation prediction in a bus compartment scenario according to claim 9, It is characterized in that The pose object is corrected using the estimated values ​​of the isolated objects through the estimation network.