Training-free action recognition method and system for skeleton point features based on AI recognition

Abstract

The invention discloses a training-free action recognition method and system for skeleton point features based on AI recognition, and the method comprises the steps: obtaining an action video which needs to be recognized, and the action video comprises a standard action video and a to-be-detected action video; recognizing the action video and the action video to be detected by using an AI model for human skeleton point recognition, obtaining three-dimensional coordinate data of each skeleton point, and obtaining posture and track data of the standard action and posture and track data of the action video to be detected according to three-dimensional coordinates; and comparing the posture and track data of the standard action with the posture and track data of the action video to be detected, and if a comparison result is within a set error range, identifying a response action. According to the method, one-by-one action AI training is not needed, only a standard video is recorded, actions are recognized through digital comparison, software generalization is high, and the method is suitable for multiple scenes; cost and time are saved, and industrial application is efficient; customized parameters are supported, different requirements are flexibly met, and the recognition precision and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recognition technology, and in particular to a training-free action recognition method and system based on skeleton point features recognized by AI. Background Art

[0002] Object recognition based on AI features has been widely used in industry solutions. For example, by extracting features from images to identify objects as large as vehicles and pedestrians, as small as helmets and fireworks; related research on using AI features to identify an action that lasts for a period of time from a video is also underway. The current research methods are to extract the image features of each frame in the video, convert the video into a continuous feature sequence, and put the feature sequence into a recurrent neural network or a similar network, such as LSTM, for training. The result is labeled as one of a series of action names. By training a large number of video samples (converted into feature sequences), the accuracy of action recognition can be improved. For example, the MS-TCN model "https: / / arxiv.org / abs / 1903.01945" currently has the highest accuracy of action recognition of 91.4% (CVPR2023 ASPnet@50Salads).

[0003] However, this end-to-end (from video to action name) training and recognition method requires a large number of video samples, including positive and negative samples, and action annotation for each sample (labeling the corresponding action name for each video segment). Furthermore, to improve accuracy, the network depth continues to increase, increasing the computing power required for training and inference. Furthermore, algorithms based on image RGB data features or optical flow variation are significantly affected by changes in background objects, lighting conditions, and other factors. Deployment in manufacturing scenarios, such as detecting whether operators on production lines are performing standard actions or missing certain steps, is difficult. This is because materials, tools, and operator clothing on the production line frequently change, affecting the RGB features in the video image and, in turn, the accuracy of action recognition. Furthermore, as production products change, so too do operational actions, rendering previously trained weights ineffective. Retraining is required to recognize new action steps, which is time-consuming and labor-intensive, making it impractical. Summary of the Invention

[0004] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of this application to avoid obscuring the purpose of this section, the abstract and the title of the invention, and such simplifications or omissions should not be used to limit the scope of the present invention.

[0005] In view of the above existing problems, the present invention is proposed. Therefore, the present invention provides a training-free action recognition method based on AI-recognized skeleton point features to solve the problems of existing video action recognition methods in terms of data requirements, computing power requirements and generalization ability.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a training-free action recognition method based on AI-recognized skeleton point features, comprising:

[0008] Obtaining an action video to be identified, wherein the action video includes a standard action video and a to-be-detected action video;

[0009] The standard action video and the action video to be detected are identified using an AI model for human skeleton point recognition, and the three-dimensional coordinate data of each skeleton point is obtained. Based on the three-dimensional coordinates, the posture and trajectory data of the standard action and the posture and trajectory data of the action video to be detected are obtained;

[0010] The posture and trajectory data of the standard action are compared with the posture and trajectory data of the action video to be detected. If the comparison result is within the set error range, the response action is recognized.

[0011] As a preferred solution of the training-free action recognition method based on AI-recognized skeleton point features described in the present invention, obtaining the action video to be recognized includes:

[0012] When capturing the standard action video and the action video to be detected, the camera position is fixed; if the camera positions are different, the action video to be detected is projected and transformed.

[0013] As a preferred solution of the training-free action recognition method based on AI-recognized skeleton point features described in the present invention, the three-dimensional coordinate data includes:

[0014] Save the three-dimensional coordinates of the skeleton points as structured sequence data;

[0015] By obtaining the relative position between at least two skeleton points in the structured sequence data as the posture data of the action;

[0016] The change of the coordinates of at least one skeleton point in the structured sequence data is obtained as the trajectory data of the action.

[0017] As a preferred solution of the training-free action recognition method based on AI-recognized skeleton point features described in the present invention, the posture and trajectory data of the standard action include:

[0018] In the posture and trajectory data of the standard action, the posture formed by some skeleton points is designated as the key posture data, and the trajectory of the coordinate change of some skeleton points is designated as the key trajectory data;

[0019] Recording and saving the key postures and key trajectories of the standard movements;

[0020] When comparing the action video to be detected with the standard action video, only the key posture and key trajectory data formed by some skeleton points are compared.

[0021] As a preferred solution of the training-free action recognition method based on AI-recognized skeleton point features described in the present invention, the error range of the posture data includes:

[0022] Compare the key posture data of multiple groups of qualified action videos to be detected and standard action videos, calculate the posture error data between each group of action videos to be detected and standard action videos, and select the maximum posture error data as the upper limit of the posture data error range.

[0023] As a preferred solution of the training-free action recognition method based on AI-recognized skeleton point features described in the present invention, the trajectory data error range includes:

[0024] Compare the key trajectory data of multiple groups of qualified action videos to be detected and standard action videos, calculate the trajectory similarity between each group of action videos to be detected and standard action videos, and select the trajectory with the largest similarity as the upper limit of the trajectory data error range.

[0025] As a preferred solution of the training-free action recognition method based on AI-recognized skeleton point features described in the present invention, the posture and trajectory data include:

[0026] For any frame in the action video, extract the position data of each pair of specified skeleton points in the three-dimensional coordinate system, calculate the coordinate difference between the two points, and aggregate the coordinate differences to obtain the posture data;

[0027] The original three-dimensional position coordinates of the skeleton points in the starting frame of the action video and the three-dimensional coordinates of the current frame of the action video after the position of the skeleton points has changed are recorded. By comparing the position coordinates of the same skeleton points in the starting frame and the current frame, the change amount of each skeleton point in the three-dimensional space is calculated, and the said change amounts are aggregated to obtain trajectory data.

[0028] In a second aspect, the present invention provides a system for training-free action recognition based on AI-recognized skeleton point features, comprising:

[0029] A data acquisition module is used to acquire action videos that need to be identified, wherein the action videos include standard action videos and action videos to be detected;

[0030] A data processing module is used to identify the standard action video and the action video to be detected using an AI model for human skeleton point recognition, obtain the three-dimensional coordinate data of each skeleton point, and obtain the posture and trajectory data of the standard action and the posture and trajectory data of the action video to be detected based on the three-dimensional coordinates;

[0031] The data analysis module is used to compare the posture and trajectory data of the standard action with the posture and trajectory data of the action video to be detected. If the comparison result is within a set error range, the response action is recognized.

[0032] In a third aspect, the present invention provides a computing device, comprising:

[0033] memory and processor;

[0034] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the training-free action recognition method based on AI-recognized skeleton point features are implemented.

[0035] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the training-free action recognition method based on AI-recognized skeletal point features.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. There is no need to train the AI ​​model for each action. You only need to record the "standard action video". Then, by using the steps described in this patent to digitally process and compare the "action video to be detected" and the "standard action video", you can identify the action. A set of software can be reused in various scenarios and has stronger generalization capabilities.

[0038] 2. This patent avoids model training for each action, greatly reducing the cost and time of deployment, and has very high practicality in industrial applications.

[0039] 3. This patent can customize auxiliary parameters for motion recognition, such as methods and parameters for specifying postures and trajectories, to facilitate custom settings in various scenarios. Users can make settings based on actual conditions to improve recognition accuracy and efficiency, and have relatively high flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort. Among them:

[0041] Figure 1 This is a schematic diagram of the overall process of a training-free action recognition method based on AI-recognized skeleton point features according to an embodiment of the present invention;

[0042] Figure 2 A schematic diagram of hand skeleton points of a training-free action recognition method based on AI-recognized skeleton point features according to an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of the posture of a "V" gesture represented by the relative positions of the six joint points in the green frame according to a training-free action recognition method based on AI-recognized skeleton point features according to an embodiment of the present invention;

[0044] Figure 4 A schematic diagram of the trajectory of skeleton point position changes in a training-free action recognition method based on AI-recognized skeleton point features according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0046] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0048] The present invention is described in detail with reference to schematic diagrams. For ease of illustration, cross-sectional views of device structures may be partially enlarged and not to scale when describing embodiments of the present invention. Furthermore, the schematic diagrams are merely illustrative and should not limit the scope of the present invention. Furthermore, in actual production, the three-dimensional dimensions of length, width, and depth should be included.

[0049] Furthermore, in the description of the present invention, it should be noted that the terms "upper, lower, inner, and outer" and other references to orientations or positional relationships are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the systems or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first, second, or third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0050] In this disclosure, unless otherwise specified or limited, the terms "mounted," "connected," and "connected" should be interpreted broadly. For example, they may refer to fixed, removable, or integral connections. They may also refer to mechanical, electrical, or direct connections, indirect connections through an intermediary, or internal communication between two components. Those skilled in the art will understand the specific meanings of these terms in this disclosure.

[0051] Example 1

[0052] Reference Figure 1-4 , as one embodiment of the present invention, provides a training-free action recognition method based on AI-recognized skeleton point features, comprising:

[0053] S100: Obtaining an action video to be identified, where the action video includes a standard action video and a to-be-detected action video;

[0054] Specifically, a video of standard movements is captured. The movements in the standard video must be complete and standardized, and the video can include multiple movements. The standard movements can be completed by an instructor or skilled personnel to ensure the completeness and standardization of the movements.

[0055] In another possible embodiment, the standard action video may be obtained through a mobile phone via social media, video sharing platforms, and the like;

[0056] After obtaining the standard action video, preprocess the video, including video editing, noise removal, and labeling. If the standard action video contains multiple actions, you can choose to split it into multiple small videos or not split it;

[0057] Preferably, when capturing the standard action video and the action video to be detected, the camera position is fixed; if the camera positions are different, the action video to be detected is projected and transformed;

[0058] Projection transformations include, but are not limited to, affine transformations, perspective transformations, and rotation-scaling transformations. Affine transformations handle basic transformations like translation, rotation, and scaling; perspective transformations more accurately simulate the effects of camera perspective changes on the video image; and rotation-scaling transformations focus on rotating and scaling the video image. Choosing the appropriate projection transformation method based on the specific video content and camera position can improve the accuracy and efficiency of video processing.

[0059] It should be noted that when shooting and acquiring standard action videos and videos of actions to be detected, fixing the camera position is an important measure to ensure video consistency and reduce the complexity of subsequent processing. However, if the camera positions are different due to actual conditions, the video of the action to be detected needs to be projected to adjust its perspective and proportion to keep it consistent with the standard action video.

[0060] S102: Using an AI model for human skeleton point recognition to identify the standard action video and the action video to be detected, obtain the three-dimensional coordinate data of each skeleton point, and obtain the posture and trajectory data of the standard action and the posture and trajectory data of the action video to be detected based on the three-dimensional coordinates;

[0061] AI models for human skeleton point recognition include OpenPose, MediaPipe, etc.

[0062] For example, taking OpenPose as an example, the process of using OpenPose to recognize action videos and obtain the three-dimensional coordinate data of skeleton points is as follows:

[0063] ① Install and configure the OpenPose environment and prepare to process the input standard action video and the action video to be detected. OpenPose uses deep learning technology, especially convolutional neural network, to predict the key points of the human body in the video frame, including joints such as head, shoulders, elbows, wrists, hips, knees and ankles, etc. Figure 2 Shown are the hand bone points;

[0064] ② Import the video file into the OpenPose system. OpenPose will process the video frame by frame, use its pre-trained model to detect human posture for each frame, identify multiple key points of the human body, and output the two-dimensional coordinates of each key point (i.e., the x, y position on the image plane).

[0065] ③ Since OpenPose directly outputs two-dimensional coordinates, the two-dimensional coordinates should be converted into three-dimensional coordinates. The conversion can use multi-view stereo vision technology or depth estimation model to estimate the depth information of the key points, that is, the z-axis coordinates;

[0066] The result of OpenPose processing is a series of data sets containing the three-dimensional coordinates of human skeleton points. These data sets can be used for subsequent tasks such as action recognition, action classification, and posture evaluation. In the embodiment of this application, they are mainly used for action recognition. By comparing the motion trajectories of the skeleton points in the standard action video and the action video to be detected, it is possible to evaluate whether the action to be detected meets the standards, thereby realizing automatic detection and scoring of the action.

[0067] It should be noted that in some advanced applications, if shooting conditions permit, such as using multiple cameras to shoot from different angles simultaneously, more accurate three-dimensional coordinates can be obtained directly through three-dimensional reconstruction technology; in single-view video processing, deep learning algorithms can be relied upon to predict the depth of key points. These algorithms usually use the continuity and contextual information of the video to assist in estimation; for example, by training an additional depth estimation network, the network can receive the two-dimensional coordinates output by OpenPose as input and output the corresponding depth value, thereby completing the conversion from two-dimensional to three-dimensional.

[0068] Preferably, the three-dimensional coordinates of the skeleton points are saved as structured sequence data; the relative position between at least two skeleton points in the structured sequence data is obtained as the posture data of the action; and the change of the coordinates of at least one skeleton point in the structured sequence data is obtained as the trajectory data of the action.

[0069] Preferably, for any frame in the action video, the position data of each pair of specified skeletal points in the three-dimensional coordinate system are extracted, and the coordinate difference between the two points is calculated, and the coordinate difference is aggregated to obtain the posture data;

[0070] Preferably, the original three-dimensional position coordinates of the skeleton points in the starting frame of the action video and the three-dimensional coordinates of the current frame of the action video after the positions of the skeleton points have changed are recorded, and the position coordinates of the same skeleton points in the starting frame and the current frame are compared to calculate the change of each skeleton point in the three-dimensional space, and the change amounts are aggregated to obtain trajectory data;

[0071] In the embodiments of this application, Figure 3 As shown, the formula for obtaining posture data is:

[0072] G0={(b1-c1),(b2-c2)…(b n -c n )}

[0073] Among them, G0 is the digital representation of the human body posture in the current frame of the action video frame, b1 represents the three-dimensional coordinate data of the b1 skeleton point, b2 represents the three-dimensional coordinate data of the b2 skeleton point, c1 represents the three-dimensional coordinate data of the c1 skeleton point, b n Representative b n The three-dimensional coordinate data of the skeleton point, c n Represents c n The three-dimensional coordinate data of the skeleton points, n is the number of skeleton points;

[0074] In the embodiments of this application, Figure 4 As shown, the trajectory data is expressed as follows:

[0075] D0={(b i '-b i ),(c j '=c j )}

[0076] Among them, D0 is the digital representation of the motion trajectory from the starting frame to the current frame in the action video, b i and c j is the three-dimensional coordinate data of the i-th and j-th bone points, b i ' and c j ' is the three-dimensional coordinate data of the i-th and j-th bone points after the change;

[0077] Preferably, in the posture and trajectory data of the standard action, the posture formed by some skeletal points is designated as the key posture data, and the trajectory of the coordinate change of some skeletal points is designated as the key trajectory data; the key posture and key trajectory of the standard action are recorded and saved; when comparing the action video to be detected with the standard action video, only the key posture and key trajectory data formed by some skeletal points are compared;

[0078] In the embodiment of the present application, when specifying the trajectory of the change of the coordinates of some skeletal points as the key trajectory data, the start and end positions of the trajectory, whether to include the direction of movement, the weight of the vertical or horizontal component, etc. can be specified, and the selection is made according to the actual situation;

[0079] In the embodiment of the present application, although some key skeleton point posture and trajectory data are selected as the comparison benchmark to improve the comparison efficiency and accuracy, this does not mean that there is an absolute restriction on the selection of data;

[0080] In another possible embodiment, when faced with a relatively small amount of video data to be detected, or when the system has sufficient computing resources and external conditions permit, it is entirely possible and feasible to select the posture and trajectory data of all skeletal points for comparison. The core of the present invention is to provide a flexible and efficient data processing and comparison method, and the selection of key data as the preferred solution is mainly based on efficiency and accuracy considerations, rather than a rigid regulation of the data processing range. Therefore, in specific applications, according to actual needs and resource conditions, the data selection strategy can be flexibly adjusted, including but not limited to the comparison method of all data or only key data.

[0081] S104: comparing the posture and trajectory data of the standard action with the posture and trajectory data of the action video to be detected. If the comparison result is within a set error range, the response action is recognized;

[0082] Preferably, key posture data of multiple groups of qualified action videos to be detected and standard action videos are compared, and posture error data between each group of action videos to be detected and standard action videos are calculated, and the maximum posture error data is selected as the upper limit value of the posture data error range;

[0083] In an embodiment of the present application, if the posture of the Kth frame of the standard action video is G0 and the posture of the Kth frame of the action video to be detected is G1, the posture error data obtained is expressed as: G1-G0.

[0084] It should be noted that when comparing the key posture data of multiple sets of qualified action videos to be detected with the standard action videos, a set of key skeleton points in the standard action is first determined. Assume that the coordinates of one of the key skeleton points in a key frame of the standard action are (1, 2, 3);

[0085] Randomly select two groups of action videos to be detected, and extract the coordinates of the skeleton points in their corresponding key frames. Suppose the coordinates of the corresponding skeleton points in the first group of videos to be detected are (2, 3, 4), and those in the second group are (3, 4, 5);

[0086] The coordinates of the skeleton points in each set of action videos to be detected are subtracted dimension by dimension from the coordinates in the standard action, and the absolute value is taken to eliminate the influence of positive and negative directions. The first set of errors is (|2-1|,|3-2|,|4-3|)=(1,1,1), and the second set of errors is (|3-1|,|4-2|,|5-3|)=(2,2,2);

[0087] Compare these two sets of errors and select the maximum value, (2,2,2), as the upper limit of the posture error data range of the skeleton point in the current key frame;

[0088] In future comparisons, as long as the absolute value of the error between the coordinates of the skeleton point and the standard coordinates in the same key frame of a set of action videos to be detected does not exceed any component of (2,2,2), the posture of the skeleton point in the current key frame can be considered qualified;

[0089] It should also be noted that the minimum value of the error is the error-free state, that is, the error is (0,0,0).

[0090] Preferably, key trajectory data of multiple groups of qualified action videos to be detected and standard action videos are compared, and the trajectory similarity between each group of action videos to be detected and the standard action video is calculated, and the trajectory with the largest similarity is selected as the upper limit of the trajectory data error range;

[0091] In the embodiment of the present application, if the trajectory of the Kth frame of the standard action video is D0, and the trajectory of the Kth frame of the action video to be detected is D1, then the trajectory error data is expressed as: D1-D0; D1-D0 can be eliminated in the software algorithm by using dynamic time normalization DTW, calculating Euclidean distance, etc. to eliminate the influence of different action speeds in different videos;

[0092] The method for obtaining the upper limit of the trajectory data error range is essentially similar to the method for obtaining the upper limit of the posture data error range. Both are determined by comparing key data between multiple sets of qualified action videos to be detected and standard action videos, and calculating the corresponding errors or similarities. Specifically, for key trajectory data, a set of key trajectory points in the standard action video will be selected first, and the similarity differences between multiple sets of action videos to be detected and these standard trajectory points at the same time frame or key frame will be calculated. By comparing these difference values, the maximum value is selected as the upper limit of the trajectory data error range.

[0093] It should be noted that the present invention uses a method of comparing the posture and trajectory data of multiple qualified action videos with the standard action video to obtain multiple errors, and taking the maximum error as the upper limit of the error range to illustrate how to complete the recognition response. However, this does not mean that there is an absolute limit on the acquisition of the error range. Using other methods to obtain the error range that meets actual needs, such as hyperparameters, is not essentially different from this method.

[0094] It should also be noted that the detection sequence of the present invention first involves the accurate identification of standard action videos, and the extraction and preservation of their key posture and trajectory data as a comparison benchmark. Subsequently, the same digital processing process is used to identify the action video to be detected, and the corresponding posture and trajectory data are obtained. By carefully comparing these data with the data of the standard action video and analyzing whether the error between the two is within a preset range, it is possible to quickly and accurately determine whether the action to be detected is successfully completed. This process does not require separate training of AI models for each action, greatly improving the generalization ability and practicality of the system, reducing deployment costs and time, and supporting customized setting of auxiliary parameters to ensure efficient and flexible action recognition in different scenarios.

[0095] The above is a schematic scheme of a training-free action recognition method based on AI-recognized skeletal point features in this embodiment. It should be noted that the technical solution of the system for training-free action recognition based on AI-recognized skeletal point features and the technical solution of the above-mentioned training-free action recognition method based on AI-recognized skeletal point features belong to the same concept. For details not described in detail in the technical solution of the training-free action recognition system based on AI-recognized skeletal point features in this embodiment, please refer to the description of the technical solution of the above-mentioned training-free action recognition method based on AI-recognized skeletal point features.

[0096] The training-free action recognition system based on AI-recognized skeleton features in this embodiment includes:

[0097] The data acquisition module is used to acquire the action video to be identified, and the action video includes the standard action video and the action video to be detected;

[0098] The data processing module is used to use the AI ​​model of human skeleton point recognition to identify the standard action video and the action video to be detected, obtain the three-dimensional coordinate data of each skeleton point, and obtain the posture and trajectory data of the standard action and the posture and trajectory data of the action video to be detected based on the three-dimensional coordinates;

[0099] The data analysis module is used to compare the posture and trajectory data of the standard action with the posture and trajectory data of the action video to be detected. If the comparison result is within the set error range, the response action is recognized.

[0100] This embodiment further provides a computing device suitable for training-free action recognition based on AI-recognized skeleton point features, including:

[0101] Memory and processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the training-free action recognition method based on AI-recognized skeleton point features proposed in the above embodiment.

[0102] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the training-free action recognition method based on AI-recognized skeletal point features as proposed in the above embodiment.

[0103] The storage medium proposed in this embodiment and the training-free action recognition method based on AI-recognized skeletal point features proposed in the above embodiment belong to the same inventive concept. For technical details not fully described in this embodiment, please refer to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0104] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented with the help of software and necessary general-purpose hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a computer's floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0105] Example 2

[0106] Referring to Table 1, an embodiment of the present invention provides a training-free action recognition method based on AI-recognized skeleton point features. In order to verify its beneficial effects, the experimental verification results of the present invention are provided.

[0107] To verify the effectiveness of the skeleton point recognition model and motion digitization method mentioned in this patent in real-world scenarios, experiments were conducted in multiple production line environments. Operation videos were captured using high-precision cameras and processed using the motion digitization technology of the present invention. The relevant experimental data was recorded, as shown in Table 1.

[0108] Table 1 Actual experimental results on the production line

[0109]

[0110]

[0111] As shown in Table 1, the accuracy of motion recognition reached 98.38% (182 / 185), meaning that almost all correct motions were accurately identified. In particular, the system achieved a 100% recognition rate for incorrect motions, significantly improving production line safety and efficiency.

[0112] Compared to traditional AI models for end-to-end action recognition, our approach achieves significant improvements in accuracy, meeting the high standards required for industrial applications. These experimental results fully demonstrate the effectiveness and reliability of our technology in complex production environments.

[0113] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A training-free action recognition method based on AI-recognized skeleton point features, characterized in that: include: Obtaining an action video to be identified, wherein the action video includes a standard action video and an action video to be detected; The standard action video and the action video to be detected are identified by using an AI model for human skeleton point recognition, and the three-dimensional coordinate data of each skeleton point is obtained. According to the three-dimensional coordinates, the posture and trajectory data of the standard action and the posture and trajectory data of the action video to be detected are obtained; The posture and trajectory data of the standard action are compared with the posture and trajectory data of the action video to be detected. If the comparison result is within a set error range, the response action is recognized.

2. The training-free action recognition method based on AI-recognized skeleton point features as claimed in claim 1, characterized in that: Obtaining the action video that needs to be identified includes: When capturing and acquiring the standard action video and the action video to be detected, the camera position is fixed; if the camera positions are different, the action video to be detected is projected and transformed.

3. The training-free action recognition method based on AI-recognized skeleton point features as claimed in claim 1 or 2, characterized in that: The three-dimensional coordinate data includes: Save the three-dimensional coordinates of the bone points as structured sequence data; By acquiring the relative position between at least two skeleton points in the structured sequence data as the posture data of the action; The change of the coordinates of at least one skeleton point in the structured sequence data is obtained as the trajectory data of the action.

4. The training-free action recognition method based on AI-recognized skeleton point features as claimed in claim 3, characterized in that: The posture and trajectory data of standard actions include: In the posture and trajectory data of the standard action, the posture formed by some skeleton points is designated as the key posture data, and the trajectory of the coordinate change of some skeleton points is designated as the key trajectory data; Record and save the key postures and key trajectories of the standard action; When comparing the action video to be detected with the standard action video, only the key posture and key trajectory data formed by some skeleton points are compared.

5. The training-free action recognition method based on AI-recognized skeleton point features as claimed in claim 4, characterized in that: The error range of attitude data includes: Compare the key posture data of multiple groups of qualified action videos to be detected and standard action videos, calculate the posture error data between each group of action videos to be detected and the standard action video, and select the maximum posture error data as the upper limit of the posture data error range.

6. The training-free action recognition method based on AI-recognized skeleton point features as claimed in claim 4, characterized in that: The error range of trajectory data includes: Compare the key trajectory data of multiple groups of qualified action videos to be detected and standard action videos, calculate the trajectory similarity between each group of action videos to be detected and standard action videos, and select the trajectory with the largest similarity as the upper limit of the trajectory data error range.

7. The training-free action recognition method based on AI-recognized skeleton point features as claimed in claim 4, characterized in that: Posture and trajectory data include, For any frame in the action video, extract the position data of each pair of specified skeleton points in the three-dimensional coordinate system, calculate the coordinate difference between the two points, and collect the coordinate difference to obtain the posture data; The original three-dimensional position coordinates of the skeleton points in the starting frame of the action video and the three-dimensional coordinates of the current frame of the action video after the positions of the skeleton points have changed are recorded. By comparing the position coordinates of the same skeleton points in the starting frame and the current frame, the change amount of each skeleton point in the three-dimensional space is calculated, and the change amounts are aggregated to obtain trajectory data.

8. A system for training-free action recognition based on AI-recognized skeleton point features, characterized in that: include, A data acquisition module is used to acquire action videos that need to be identified, wherein the action videos include standard action videos and action videos to be detected; A data processing module is used to identify the standard action video and the action video to be detected by using an AI model for human skeleton point recognition, obtain the three-dimensional coordinate data of each skeleton point, and obtain the posture and trajectory data of the standard action and the posture and trajectory data of the action video to be detected according to the three-dimensional coordinates; The data analysis module is used to compare the posture and trajectory data of the standard action with the posture and trajectory data of the action video to be detected. If the comparison result is within a set error range, the response action is identified.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the training-free action recognition method based on AI-recognized skeleton point features are implemented as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the training-free action recognition method based on AI-recognized skeleton point features as described in any one of claims 1 to 7.

Images