50 results about "Point sequence" patented technology

The invention discloses a method and a system for generating 3D (three-dimensional) human motion under text driving by utilizing a 2D (two-dimensional) video. The method comprises the following steps of: acquiring the video and preprocessing to obtain a two-dimensional key point sequence and text description; the two-dimensional key point sequence passes through a spatiotemporal feature adapter to obtain a potential spatiotemporal feature sequence, a residual vector quantizer quantizes and outputs a three-dimensional SMP L parameter sequence, and meanwhile, potential spatiotemporal features and a discrete Token sequence are mapped; preprocessing a text to extract a semantic vector, partially covering a Token sequence of a basic quantization layer, reconstructing a prediction sequence through a predictor in combination with the semantic vector, and obtaining a complete sequence through a refiner; constructing a total loss function and a text-to-action loss function to train the module; and inputting the text description and the basic quantization layer Token to a trained module, outputting a three-dimensional SMPL parameter sequence, and rendering to generate a three-dimensional human body grid and animation. According to the method, the end-to-end generation from the text to the three-dimensional SMPL action is realized only by two-dimensional key points and text description.

The invention relates to a fitness action error correction method and device for bone key point detection, equipment and a medium. The method comprises the following steps: acquiring video data in a multi-person fitness scene, and performing skeleton key point detection on each frame of image in the video data to obtain a multi-person skeleton key point sequence; performing spatial clustering processing on the multi-person skeleton key point sequence to obtain a plurality of individual skeleton key point sequences; according to the skeleton key point sequence of each individual, generating an action track of each individual; based on a preset standard action sequence, identifying an action matching result corresponding to each action track; according to the action matching result, the action deviation value of each individual is calculated, and action error correction feedback information is generated based on the action deviation values; the error correction feedback information is used for indicating each individual to adjust the fitness action. The method can adapt to complex scenes, solves the problem that key points of scenes of multiple persons are confused, and provides effective support for fitness guidance and the like.

The invention discloses a synchronous analysis method, device and equipment based on multi-person postures and dynamic time warping and a computer readable storage medium, and the method comprises the steps: carrying out the processing of a video stream containing a plurality of individuals, so as to detect and output a face bounding box of each individual in a scene; for each face bounding box, generating an independent skeleton point sequence uniquely corresponding to each individual; based on each independent skeleton point sequence, constructing an independent action sequence of a plurality of individuals; calculating and outputting an individual similarity score of each individual according to each independent action sequence and the corresponding standard action template; performing statistical variance calculation on the set of the similarity scores of all the individuals to obtain a group synchronism index representing group action synchronism; and generating an analysis report containing the evaluation result of the individual performance and the group synchronism. The method has the advantages that the key points of the individuals in the multi-target scene are accurately associated, and the collaboration of group actions is objectively analyzed.

The invention relates to a human body posture estimation key point correction method based on skeleton proportion constraint, and belongs to the technical field of human body posture estimation key point detection. The method comprises the following steps: extracting key points of a human body in a human body motion video, forming a skeleton based on the key points, calculating a skeleton proportion value of each skeleton length and a thoracolumbar spine length in the skeleton, and defining a frame corresponding to the skeleton proportion value not within a preset threshold as an invalid frame; iterating the positions of the key points in the invalid frame until the bone proportion values formed by all key point sequences in the invalid frame are within a preset threshold value or reach a preset maximum number of iterations; and detecting the widths of the four limbs based on the iterated key points, adjusting the positions of the key points to the axis positions corresponding to the width centers of the four limbs to obtain the adjusted key points, and reconstructing a key point sequence to realize key point correction. The objective of the invention is to solve the technical problem of low key point detection accuracy caused by deviation in key point positioning in the prior art.

The invention relates to a universal decoding method for an annular coding mark, and the method comprises the steps: obtaining a standard coding image which comprises a to-be-decoded annular coding mark, the annular coding mark comprises a central circle and an annular belt which is concentric with the central circle, and the position of the central circle is used as the position of the annular coding mark; setting a constraint condition based on the annular coding mark, and screening the standard coding image to obtain an effective coding image; taking the pixel transformation position of the annular belt in the annular coding mark as a transformation point, and generating a transformation point sequence according to the effective coding image; and identifying a transition point sequence, and distributing code values to complete decoding. The method comprises the following steps: screening a standard coded image to obtain an effective coded image suitable for decoding; the pixel transformation position of the annular belt in the annular coding mark is used as a transformation point to generate a transformation point sequence, so that the processing efficiency is higher; and whether the codes are similar codes or not can be judged according to whether the transition point sequences meet consistency conditions or not, and different code values are allocated according to different conditions, so that universal decoding is realized.

The invention provides a method and system for pre-training a sign language understanding framework based on semantic enhancement of skeletons, and relates to the technical field of sign language recognizing.The method comprises the steps that sign language video data and skeleton sequences and text data matched with the sign language video data are obtained; extracting a skeleton key point sequence from the sign language video data, modeling to form skeleton features, and performing word segmentation processing on a text; in a pre-training stage, skeleton features and word segmentation texts are input into a fusion network to generate bidirectional enhanced features, and global and local similarities are obtained through double-layer semantic alignment; calculating the comparison loss based on the similarity, and coordinating the weight through the balance parameter to obtain the hierarchical loss; executing the matching task and the language modeling task to obtain corresponding loss, weighting and combining the three types of loss into pre-training total loss, and adjusting parameters to complete training; and finally, part of parameters are optimized in combination with specific task types in a fine tuning stage, and enhanced understanding of sign language semantics is realized. The sign language understanding accuracy is improved.

The invention relates to the technical field of computers, and discloses an anthropomorphic sliding track generation method and system, and the method comprises the steps: obtaining video data containing human sliding operation, and carrying out the preprocessing, thereby obtaining a static image and scalar time; inputting the static image into an image encoder, inputting scalar time into a time encoder, respectively extracting space task and time constraint features, and fusing the space task and time constraint features into fusion condition features; the method comprises the following steps of: inputting a track point into a layered generator, outputting a current track point coordinate by a track generation head at each time step of the generator, and generating a track point sequence and a video frame sequence by a video generation head in combination with the coordinate, a fusion feature and a frame generated in a previous time step; and inputting the track and the video frame into a discriminator, and returning an authenticity result to adjust parameters of the generator after combined judgment by the discriminator until the output reaches the standard, thereby obtaining the anthropomorphic sliding track. According to the method, anthropomorphic similarity and sample diversity of track generation are improved, and a richer data set close to real human behaviors can be provided for downstream applications.

The invention discloses a computational efficient point cloud analysis method based on a grouping selective state space, and the method comprises the steps: firstly, a sequence extension module carries out the serialization of points along each axis, and enables a disordered point cloud to more stably adapt to the causal characteristics of Mamba without parameters; secondly, sorting prompt and position embedding are adopted to provide sorting and position information for the point sequence respectively, so that geometric semantics are better captured; thirdly, the chain type bidirectional Mama enables the forward process and the reverse process in the parallel bidirectional Mama to be connected in series, a global receptive field on a point sequence is provided, and meanwhile high-order geometric information is captured in the scanning process. And fourthly, the grouping selective state space model introduces parameter sharing among multiple dimensions into the selective state space model, so that overfitting caused by a calculation mode in the selective state space model is relieved. And 5, a sequence merging module fuses corresponding high-order interaction features obtained through causal reasoning on different sequences. And finally, packaging the flow into a basic block, and embedding the basic block into a standard codec architecture for hierarchical feature aggregation. According to the method, the accuracy of point cloud analysis based on the state space model is improved while the calculation efficiency is ensured.

The embodiment of the application provides a personnel fall judgment method, device, storage medium and system, the method comprises the following steps: acquiring the latest personnel image of the target personnel, acquiring the key point information corresponding to each key point of the target personnel in each personnel image; judging whether the target personnel falls based on the acquired key point information, and generating first judgment information; inputting all the acquired key point information into a personnel key point sequence classifier, judging whether the target personnel falls through the personnel key point sequence classifier, and generating second judgment information; acquiring the personnel image information of each personnel image, inputting all the acquired personnel image information into a personnel image sequence classifier, judging whether the target personnel falls through the personnel image sequence classifier, and generating third judgment information; judging whether the target personnel really falls based on the first judgment information, the second judgment information and the third judgment information, so that the accuracy of judging whether the personnel falls can be improved.

The motion synthesis method provided by the embodiment of the invention comprises the following steps: detecting a video sequence to obtain a key point sequence, and carrying out visualization processing to align with a human skeleton to obtain a preliminary skeleton sequence; a diffusion probability model is adopted, Gaussian noise of different degrees is overlaid on the key point sequence according to time, and the longer the time is, the more the added Gaussian noise is; and using the noise-added key point sequence as a condition, adopting the diffusion probability model to predict noise under the guidance of semantic features, and obtaining a final skeleton sequence after denoising. According to the motion synthesis method provided by the embodiment of the invention, the robustness of human body motion synthesis can be improved through introduction of the diffusion probability model and embedding of conditions such as voice semantics. The embodiment of the invention further provides an action synthesis device and electronic equipment.

Method for providing a sequence of measurement points, in particular for carrying out a test procedure for measuring a physical unit, wherein each measurement point is defined by a combination of one input value from several input variables according to the sequence of input values, wherein the sequence of input values ​​of each input variable corresponds to a step function with steps having a step height (H), a jump height (S) between the steps, and a step length (L); comprising the following steps: - Providing (S3, S4, S5; S11, S21) a step height value sequence for each of the input variables, wherein the step height value sequence corresponds to a uniformly distributed value sequence and specifies a sequence of step heights (H) along the course of the step function of the input variable in question; - Providing (S2, S12, S22) a step length value sequence for each of the input variables, wherein the step length value sequence corresponds to a uniformly distributed value sequence and specifies a sequence of step lengths (L) along the course of the step function; - Generating (S3, S4, S5; S13, S14, S15, S16; S23, S24, S25, S26) sequences of input values ​​for each of the inputs depending on the sequence of step heights (H) and depending on the sequence of step lengths (L), wherein generating sequences of input values ​​for each of the inputs comprises generating the input values ​​of each of the inputs in a sequence resulting from concatenating the step heights according to their order in the step height value sequence assigned to the input in question, wherein each of the step heights is provided successively with a number resulting from the sequence of step lengths assigned to the input in question in the step length value sequence; - Generating (S6, S18, S28) the sequence of measurement points by combining the corresponding input values ​​of the sequences of input values ​​of the multiple input variables.

The present disclosure relates to a motion scoring method, device, equipment and storage medium, and relates to the technical field of fitness mirrors. The method comprises the following steps: first, acquiring a human body key point sequence corresponding to a to-be-scored motion on a target image, a human body key point template and a human body joint angle template; determining a key point Pearson matching degree, a joint angle Pearson matching degree and a joint angle difference value similarity; and determining a motion score of the to-be-scored motion on the target image according to the key point Pearson matching degree, the joint angle Pearson matching degree and the joint angle difference value similarity. It can be seen that, by means of the key point Pearson matching degree, the joint angle Pearson matching degree and the joint angle difference value similarity, the motion score of the to-be-scored motion on the target image is determined, and the accuracy of the motion score is improved.

The application relates to a method, system and computer readable medium for finding a target user based on a signaling trajectory. The method comprises: recalling candidate users meeting a target trajectory point sequence T(p1, p2,..., p n ); obtaining a candidate trajectory point sequence Ts(ps1, ps2,..., ps m ); comparing the candidate trajectory point sequence with the target trajectory point sequence to generate a distance matrix d; calculating a shortest path D between trajectories based on the distance matrix by using dynamic programming; and finding the target user according to the shortest path.

This application discloses a method, system, device, and medium for handling violation scenarios in large-scale model logs, mainly relating to the field of violation handling technology. It aims to address the problems of rigid rules, limited coverage, knowledge silos, lack of correlation, reliance on experts, and high iteration costs in existing solutions. The method includes: tracking behavioral nodes corresponding to several consecutive raw logs to form a behavioral node sequence; comparing the behavioral node sequence with preset legal path sequences in a behavior tree model library to determine if a preset abnormal node sequence exists; when an abnormal node sequence exists, determining whether it conforms to an event sequence corresponding to any rule in the event analysis rule library; if it does, entering the preset alarm scheme corresponding to the event sequence; if it does not, entering the unknown scenario mining process, inputting the abnormal node sequence into a trained second large-scale language model to obtain an initial processing scheme, and inputting the initial processing scheme into a preset verification terminal to obtain the final processing scheme.

The invention relates to a video alignment method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining a jump point sequence of a reference code stream and the first frame offset of the reference code stream, and obtaining the first frame offset of a candidate code stream based on the jump point sequence of the reference code stream, the first frame offset of the reference code stream and the first frame offset of the candidate code stream; correcting and aligning original jump point sequences of the candidate code streams to obtain current jump point sequences of the candidate code streams, the reference code stream and all the candidate code streams being code streams of the same video, and the reference code stream and all the candidate code streams being code streams of the same video; the plot jump points at the same positions in the jump point sequence of the reference code stream and the original jump point sequence of the candidate code streams are used for representing the jump points of the same plot, and the picture contents corresponding to the same positions in the current jump point sequences of any two candidate code streams are the same, and issuing the current jump point sequence of the target code stream to a video playing terminal. Calibration is achieved through calculation, and therefore the problem that pictures of different code streams are inconsistent at the same progress point, and consequently the pictures flicker in the jumping process is solved.

Embodiments of the present application disclose a data processing method and device, computer equipment and a storage medium, wherein the method comprises the following steps: obtaining a question answered by a target user and a target question not answered; generating a question sequence, a knowledge point sequence and an answer record sequence of the target user according to the answered question and the target question; performing feature extraction on the question sequence, the knowledge point sequence and the answer record sequence to generate a question vector corresponding to the question sequence, a knowledge point vector corresponding to the knowledge point sequence and an answer record vector corresponding to the answer record sequence; performing graph feature extraction on the question vector and the knowledge point vector to generate a corresponding graph vector; and generating a correct probability of the target user answering the target question based on the graph vector, the question vector, the knowledge point vector and the answer record vector through a question answering prediction model. The present application can accurately grasp the degree of user's mastery of knowledge points and improve the accuracy of predicting user's answers.

The application discloses a kind of two-dimensional space to three-dimensional space key point mapping conversion method and system, including sequence preprocessing module: the length of two-dimensional space key point sequence is obtained, carries out sparse preprocessing, obtains sparse two-dimensional space key point sequence;Mapping dimension-increasing module: sequence modeling is carried out to sparse two-dimensional space key point sequence, obtains three-dimensional space key point sequence after preliminary mapping;Sequence aggregation module: three-dimensional space key point of center target is obtained by compressing and aggregating three-dimensional space key point sequence after preliminary mapping;Training network module: end-to-end training is carried out to network using supervised learning, and mapping accuracy of three-dimensional space key point estimation network is improved using space-time constraint strategy;Visualization module is used to obtain visual result.The application has good robustness and universality, and can be widely applied in key point detection of a variety of objects.

The invention discloses a human body action video generation method based on angle conditions. Comprising the following steps: acquiring a reference figure image or video frame and a reference posture sequence of a target action, and performing human body key point detection; generating a key point latent variable sequence and a continuous angle feature sequence; constructing a conditional sequence, generating a latent variable-conditional sequence according to the key point latent variable sequence and the conditional sequence, and further obtaining a training data set; then training the key point generation model to obtain a trained key point generation model; and selecting a reference person video frame and a target action and generating a corresponding latent variable-conditional sequence, thereby obtaining a target key point sequence and inputting the target key point sequence into a video generation model, and generating a target human body action video. According to the invention, while the consistency of figure identities is maintained, the cross-body-type high-precision and editable human body video generation is realized, the action connection is natural, the robustness is high, and the method is suitable for scenes such as virtual human production, movie and television animation, human-computer interaction and the like.

This application provides a computer vision-based gesture recognition system, method, device, and medium. For each occluded keypoint, based on a pre-defined spatiotemporal graph model of the target object's hand keypoints, a graph convolutional network is used to extract spatial features of keypoints with high detection confidence in the spatial neighborhood of the occluded keypoint and temporal series features with high detection confidence in historical frames before occlusion. This yields a spatiotemporal collaborative attention map for guiding information completion. Based on the spatiotemporal collaborative attention map, feature propagation and aggregation are performed from the keypoints with high detection confidence to the occluded keypoints to obtain the completed feature representation of the occluded keypoints. The completed coordinates of the occluded keypoints are then output, generating an occlusion-resistant keypoint sequence for the target object's hand region. Based on this occlusion-resistant keypoint sequence, the gesture category of the target object is identified. Using the scheme of this application, robust gesture recognition against occlusion can be achieved.

It is determined whether each of the first motion point sequence and the second motion point sequence contains a first motion point of a first ordinal number and a second motion point of a second ordinal number, and the first motion point sequence and the second motion point sequence contain the first and second multiple motion points of one or more body parts of a person, ordered according to the respective time points at which the first and second multiple motion points were detected from the first and second series of video frames corresponding to the detection region. A method for identifying the start and end of a motion point sequence for a person to perform the task, based on the first motion point and the second motion point, from the first motion point sequence and / or the second motion point sequence.

The invention relates to the field of computer vision, and provides a dynamic gesture detection method and system and electronic equipment, and the method comprises the steps: obtaining to-be-detected target video sequence information; preprocessing the target video sequence information to obtain an appearance frame sequence and a key point sequence; performing data processing on the appearance frame sequence based on a space-time Transform encoder to obtain an appearance feature vector, and performing data processing on the key point sequence based on the space-time Transform encoder to obtain a key point feature vector; performing cross-modal attention fusion on the appearance feature vector and the key point feature vector to generate a fusion feature vector; and performing dynamic gesture detection based on the fused feature vector. The gesture detection method is used for overcoming the defects that in the related technology, when dynamic gesture detection is carried out, time sequence modeling capacity is poor, and performance in a complex scene is poor. The gesture detection method is higher in accuracy.

The invention provides a video compression method and system based on a LivePortrait-GAN video technology, and the method comprises the steps: extracting a coordinate set of key points of a human body of a teacher, carrying out the matching of the coordinate set with a preset teaching action semantic template library, carrying out the coding of a template number and a space offset if the matching succeeds, and otherwise, reserving a residual key point sequence; adjusting the residual key point sequence in combination with the synchronous voice rhythm characteristics, the structure mask and the space offset; fusing the semantic anchoring vector, the residual vector and a gating parameter generated based on the ratio of the residual amplitude to the template offset amplitude, and generating a driving vector with a smooth time sequence; and driving the clipped and fixed-point LivePortrait-GAN generator to reconstruct a portrait picture by using the identity characteristics of the static reference image of the teacher and the driving vector, and outputting a final video frame through frame-level fusion. The key model can be efficiently deployed on a domestic artificial intelligence acceleration chip (such as AX650N), end-to-end real-time coding and rendering are achieved, and the domestic controllable deployment requirement is met.

The embodiment of the present disclosure discloses a passing point prediction method and a passing point prediction model training method and device, the method comprising: obtaining a historical passing point sequence of a navigated object, the historical passing point sequence comprising data records indexed by historical passing points, each data record comprising a historical passing point and a corresponding historical scene feature sequence; obtaining a scene feature of a target navigation route of the navigated object as a current scene feature; inputting the current scene feature and the historical passing point sequence into a model, the model comprising a recommendation representation module, and obtaining a recommended representation of a historical passing point corresponding to a historical scene feature based on at least the current scene feature and the historical scene feature in the historical passing point sequence; the model comprising a first scoring module, and obtaining a score of each historical passing point based on the recommended representation of the historical passing point; and determining a passing point prediction result based on at least the score of each historical passing point. The embodiment of the present disclosure can improve the prediction accuracy.

The invention discloses a client moving target rendering method, device and equipment and a storage medium, a client comprises a moving target layer constructed based on a scene rendering component and used for bearing a moving target object, and the method comprises the steps that moving target data is acquired, and the moving target data comprises a unique identifier, motion state information and attribute information of the moving target object; based on the unique identifier of the moving target object, searching the moving target object corresponding to the unique identifier in the moving target layer, and taking the moving target object as a to-be-rendered moving target object under the condition that the moving target object corresponding to the unique identifier exists in the moving target layer; performing interpolation calculation on the motion state information of the moving target object to be rendered to generate an interpolation point sequence; and based on the interpolation point sequence, updating the motion state information of the to-be-rendered moving target object, and rendering the to-be-rendered moving target object based on the updated motion state information and attribute information. Therefore, the continuity of the moving target object in the moving process can be improved.

The invention discloses a large-model log violation scene processing method, system and device and a medium, mainly relates to the technical field of violation processing, and is used for solving the problems of rule stiffness, limited coverage, knowledge islanding, lack of association, dependency on experts and high iteration cost in the existing scheme. Comprising the steps of tracking behavior nodes corresponding to a plurality of continuous original logs to form a behavior node sequence; comparing the behavior node sequence with a preset legal path sequence in a behavior tree model library, and determining whether a preset abnormal node sequence exists or not; when the abnormal node sequence exists, determining whether the abnormal node sequence accords with an event sequence corresponding to any rule in an event analysis rule base, and when the abnormal node sequence accords with the event sequence corresponding to any rule in the event analysis rule base, entering a preset alarm scheme corresponding to the event sequence; and if not, entering an unknown scene mining process, inputting the abnormal node sequence into a trained second large language model to obtain an initial processing scheme, and inputting the initial processing scheme into a preset verification terminal to obtain a final processing scheme.

This invention discloses a method and apparatus for real-time writing direction determination based on single-character stroke features. The method includes: acquiring a sequence of coordinate points of a stylus on a whiteboard in real time; calculating a feature set representing the writing features of a single character based on the coordinate point sequence; calculating a horizontal comprehensive score by weighting the feature set; calculating a vertical comprehensive score by weighting the reciprocals of each feature in the feature set; and determining the writing direction of the stylus based on the horizontal and vertical comprehensive scores to obtain a determination result. This invention proposes a method and apparatus for real-time writing direction determination based on single-character stroke features. It calculates a feature set and calculates a weighted score based on the single-character stroke features in real time to determine the writing direction, without relying on multi-character layouts, thus avoiding response delays at the source and enabling independent and effective recognition of single characters. Therefore, it can solve the problems of response delays and single-character recognition failures caused by existing technologies that rely on multi-character layouts to determine writing direction.

The invention discloses a game scoring method, device and equipment based on dynamic time warping and a computer readable storage medium. The method comprises the steps of obtaining a player skeleton point sequence corresponding to real-time actions of a player; comparing the player skeleton point sequence with a standard action template by applying a kinematics dynamic time warping algorithm so as to generate a posture evaluation score representing the player posture accuracy; in the game engine, a game event object corresponding to the standard action template is driven to move along a preset track; monitoring the position of the game event object in real time, and judging whether the game event object enters a preset effective judgment interval or not to generate a time effectiveness signal; calculating to obtain a final game score based on the posture evaluation score and the time validity signal; and based on the final game score, updating the continuous click number and score statistical data in the game. The method has the advantage that game scoring is more reasonable.

This application provides a method for reconstructing a three-dimensional model of a target object, comprising: obtaining image feature information corresponding to each video frame image in a video to be detected and a two-dimensional key point sequence of the target object; estimating and obtaining a three-dimensional key point sequence of the target object corresponding to the two-dimensional key point sequence of the target object based on the two-dimensional key point sequence of the target object; concatenating the image feature information, the two-dimensional key point sequence of the target object, and the three-dimensional key point sequence of the target object to obtain a feature sequence corresponding to the target object; and obtaining a three-dimensional model of the target object based on the feature sequence corresponding to the target object. This method uses the key point sequence as part of the model input elements, improving the accuracy of the model key point prediction and making the posture changes between the three-dimensional models of the target object corresponding to each video frame image in the video to be detected smoother and more realistic.

The invention discloses an action correction method, device and equipment based on dynamic time warping and a computer readable storage medium. The method comprises the steps of obtaining a user skeleton point sequence corresponding to a user action and a standard skeleton point sequence corresponding to a standard action; applying a kinematics dynamic time warping algorithm to determine an optimal warping path and time alignment similarity; according to the optimal regular path, space trajectory similarity is obtained through calculation; dividing the user skeleton point sequence into a plurality of sub-fragments according to the optimal regular path, and calculating to obtain a local composite evaluation score for each sub-fragment; determining the sub-fragments of which the local composite evaluation scores are lower than a preset error threshold value as error action fragments in the user skeleton point sequence; and for the error action fragment, calculating to obtain a quantized multi-dimensional deviation vector, and generating a correction instruction according to the multi-dimensional deviation vector. The method has the advantages of accurate error positioning and quantitative correction guidance.

The disclosure provides an image and text based video generation method and device, and relates to the technical field of computer vision. The method comprises the following steps: obtaining fusion encoding, wherein the fusion encoding comprises image encoding of a reference image and text encoding of action text, and the reference image comprises a target object; inputting the fusion encoding into a decoder to obtain a target key point sequence; and obtaining a reconstructed video according to the reference image and the target key point sequence, wherein the video content of the reconstructed video is that the target object performs an action described in the action text; wherein the decoder is obtained by training based on a first key point sequence of an original video sample and a second key point sequence corresponding to a reconstructed video sample, and original video encoding samples and fusion encoding samples, and the fusion encoding samples comprise text encoding of an action text sample and image encoding of a reference image sample.