A method and system for positioning temporal actions based on feature decoupling
Through the feature decoupling method, features suitable for classification and positioning are extracted separately, and a decoupling positioning model is constructed, which solves the problem of poor balance between action classification and positioning tasks in the existing technology, and improves the accuracy and accuracy of action positioning.
Patent Information
- Application Number
- CN202210101251.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-27
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-01-27
AI Technical Summary
The existing timing action positioning methods cannot take into account both action classification and positioning tasks, resulting in reduced classification accuracy and inaccurate positioning.
Using a feature-based decoupling method, the features suitable for classification and positioning are extracted through the autoencoder, the classification network and the positioning network, the decoupling positioning model is constructed, the loss value is calculated and the model parameters are updated, and the pre-trained decoupling positioning model is obtained.
The accuracy of classification and the accuracy of positioning are improved, so that the parts of classification and positioning that each focus on no longer affect each other.
Smart Images

Figure CN114511809B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a temporal action positioning method and system based on feature decoupling. Background Art
[0002] Currently, temporal action localization is attracting increasing research attention due to its wide-ranging applications in action retrieval, video summarization, and intelligent security. Temporal action localization is a fundamental task in video understanding, requiring both the classification of action instances and the localization of their temporal boundaries. Effectively addressing this task requires pursuing two objectives: sufficient visual invariance for action classification and fine-grained differentiation of the start and end moments of an action for localization.
[0003] Existing traditional temporal action localization methods use shared inputs (example explanatory features, such as proposal-level features) for classification and localization tasks, but do not fully consider the above two objectives. It is important to emphasize that classification tasks require features to focus on the most discriminatory points, which can also be understood as the differences between different actions; while localization tasks require features to focus on the start and end times of the current action and whether the action itself is complete. It is well known that the most discriminatory points and the start and end times of the action are often different (for classification and localization tasks, they are noise to each other).
[0004] In summary, the current shared input cannot take into account both classification and localization tasks at the same time, resulting in reduced classification accuracy and inability to accurately locate. A new temporal action localization method or system is urgently needed. Summary of the Invention
[0005] The present invention aims to provide a method and system for localizing temporal actions based on feature decoupling to address one or more of the aforementioned technical issues. The method provided by the present invention eliminates the mutual influence between the respective areas of focus of classification and localization, thereby improving both classification accuracy and localization precision.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions:
[0007] The present invention provides a method for localizing temporal actions based on feature decoupling, comprising the following steps:
[0008] Obtain the video to be temporally positioned and cut it into several video segments, and extract the rough features of each video segment;
[0009] The rough features of each video clip are input into a pre-trained decoupling localization model, and the decoupling localization model outputs action classification results and action localization results;
[0010] Wherein, the decoupling positioning model includes:
[0011] An autoencoder that takes the rough features of a video clip as input, decouples them, and outputs the fine classification features and fine positioning features of the corresponding video clip;
[0012] A classification network, configured to input the detailed classification features and output an action classification result;
[0013] The positioning network is used to input the detailed positioning features and output action positioning results.
[0014] A further improvement of the method of the present invention is that the autoencoder comprises:
[0015] An encoder is configured to input rough features of a video clip and output three sets of latent variables; wherein the three sets of latent variables include latent variables suitable only for classification, latent variables suitable only for positioning, and latent variables suitable for both positioning and classification;
[0016] The decoder is used to input the three parts of latent variables and output restored rough features.
[0017] A further improvement of the method of the present invention is that the step of obtaining the pre-trained decoupled positioning model includes:
[0018] (1) Obtaining a sample training set; wherein each sample includes rough features of the sample video clip, action location labels, and action classification labels;
[0019] (2) Inputting the rough features of each sample video clip into the autoencoder, and outputting three parts of latent variables through the encoder in the autoencoder; inputting the three parts of latent variables into the decoder of the autoencoder to obtain the restored rough features;
[0020] (3) processing the three parts of latent variables and inputting them into the decoder of the autoencoder to obtain fine classification features; inputting the fine classification features into the classification network to obtain the predicted classification results; wherein the processing rule is to replace any part of the three parts of latent variables with a 0 vector;
[0021] (4) The three parts of latent variables are processed and input into the decoder of the autoencoder to obtain fine positioning features; the fine positioning features are input into the positioning network to obtain the predicted positioning results; wherein the processing rule is to replace any part of the two parts of latent variables remaining after replacing the 0 vector in step (3) with a 0 vector;
[0022] (5) Calculate the loss value of the decoupled positioning model, update the model parameters based on the loss value, and obtain the pre-trained decoupled positioning model.
[0023] A further improvement of the method of the present invention is that the calculation expression of the loss value L of the decoupling positioning model is:
[0024] L=l a +l c +l l ;
[0025] Where, l a is the reconstruction loss of the autoencoder, l c is the classification loss, l l is the positioning loss.
[0026] A further improvement of the method of the present invention is that the reconstruction loss of the autoencoder is cosine similarity or Euclidean distance loss;
[0027] The classification loss is cross entropy loss;
[0028] The positioning loss is an absolute distance loss.
[0029] The present invention provides a temporal action positioning system based on feature decoupling, comprising:
[0030] The rough feature acquisition module is used to obtain the video to be temporally positioned and cut it into several video segments, and extract the rough features of each video segment;
[0031] A result acquisition module is used to input the rough features of each video clip into a pre-trained decoupling positioning model, and the decoupling positioning model outputs an action classification result and an action positioning result;
[0032] Wherein, the decoupling positioning model includes:
[0033] An autoencoder that takes the rough features of a video clip as input, decouples them, and outputs the fine classification features and fine positioning features of the corresponding video clip;
[0034] A classification network, configured to input the detailed classification features and output an action classification result;
[0035] The positioning network is used to input the detailed positioning features and output action positioning results.
[0036] A further improvement of the system of the present invention is that the autoencoder includes:
[0037] An encoder is configured to input rough features of a video clip and output three sets of latent variables; wherein the three sets of latent variables include latent variables suitable only for classification, latent variables suitable only for positioning, and latent variables suitable for both positioning and classification;
[0038] The decoder is used to input the three parts of latent variables and output restored rough features.
[0039] A further improvement of the system of the present invention is that the step of obtaining the pre-trained decoupled positioning model includes:
[0040] (1) Obtaining a sample training set; wherein each sample includes rough features of the sample video clip, action location labels, and action classification labels;
[0041] (2) Inputting the rough features of each sample video clip into the autoencoder, and outputting three parts of latent variables through the encoder in the autoencoder; inputting the three parts of latent variables into the decoder of the autoencoder to obtain the restored rough features;
[0042] (3) processing the three parts of latent variables and inputting them into the decoder of the autoencoder to obtain fine classification features; inputting the fine classification features into the classification network to obtain the predicted classification results; wherein the processing rule is to replace any part of the three parts of latent variables with a 0 vector;
[0043] (4) The three parts of latent variables are processed and input into the decoder of the autoencoder to obtain fine positioning features; the fine positioning features are input into the positioning network to obtain the predicted positioning results; wherein the processing rule is to replace any part of the two parts of latent variables remaining after replacing the 0 vector in step (3) with a 0 vector;
[0044] (5) Calculate the loss value of the decoupled positioning model, update the model parameters based on the loss value, and obtain the pre-trained decoupled positioning model.
[0045] A further improvement of the system of the present invention is that the calculation expression of the loss value L of the decoupling positioning model is:
[0046] L=l a +l c +l l ;
[0047] Where, l a is the reconstruction loss of the autoencoder, l c is the classification loss, l l is the positioning loss.
[0048] A further improvement of the system of the present invention is that the reconstruction loss of the autoencoder is cosine similarity or Euclidean distance loss;
[0049] The classification loss is cross entropy loss;
[0050] The positioning loss is an absolute distance loss.
[0051] Compared with the prior art, the present invention has the following beneficial effects:
[0052] In the method specifically provided by the present invention, by decoupling the shared features, features suitable for classification and features suitable for positioning are obtained respectively, so that the parts that classification and positioning respectively focus on no longer affect each other, and the accuracy of classification and the accuracy of positioning can be improved. The core invention of the present invention is that, first, the present invention points out for the first time the problem that shared input cannot take into account classification and positioning at the same time, and provides a new solution direction for the study of time sequence action tasks. Second, the present invention decouples the shared input through the constructed model, and obtains features that are more suitable for classification and positioning respectively, so that the results of classification and positioning are more accurate. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art; obviously, the drawings described below are some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0054] Figure 1 1 is a schematic diagram of an implementation flow of a method for locating sequential actions based on feature decoupling according to an embodiment of the present invention;
[0055] Figure 2 is a schematic diagram of inputting a point of interest in an embodiment of the present invention;
[0056] Figure 3 is a schematic diagram of decoupling feature focus in an embodiment of the present invention;
[0057] Figure 4 Schematic diagram of the overall framework and decoupling network in an embodiment of the present invention; wherein, Figure 4 (a) is a schematic diagram of the overall framework. Figure 4 (b) is a schematic diagram of the decoupling network;
[0058] Figure 5 3 is a schematic diagram of positioning effects of the method in the embodiment of the present invention and the existing method on two popular datasets (THUMOS14 and ActivityNet v1.3) in the embodiment of the present invention. DETAILED DESCRIPTION
[0059] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.
[0060] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0061] The present invention is described in further detail below with reference to the accompanying drawings:
[0062] See also Figure 1 A temporal action positioning method based on decoupling classification features and positioning features according to an embodiment of the present invention includes the following steps:
[0063] Get the video and split it into segments with a fixed number of frames and extract features;
[0064] Based on the constructed autoencoder, the extracted video clip features are decoupled to obtain video clip features suitable for classification and video clip features suitable for positioning;
[0065] Based on the video clip features suitable for classification, obtaining proposal-level aggregated classification features; based on the video clip features suitable for positioning, obtaining proposal-level aggregated positioning features;
[0066] Based on the aggregated classification features at the proposal level, a classification result is obtained; based on the aggregated positioning features at the proposal level, a positioning result is obtained.
[0067] In an embodiment of the present invention, constructing an autoencoder includes:
[0068] (1) Obtain video clips inside the proposal;
[0069] (2) Project the features of these video clips into three different spaces: the part that is only suitable for classification, the part that is only suitable for positioning, and the part that both parties are concerned about (the common part);
[0070] (3) Combine the above three parts and project them back to the original space, requiring that the features of the video clip before projection and the features after projection are completely consistent (including feature dimensions and the data values of the features themselves);
[0071] In an embodiment of the present invention, obtaining features suitable for classification within a proposal includes: (1) merging the portion suitable only for classification with the shared portion; and (2) projecting back to the original space to obtain features of the video segment suitable only for classification.
[0072] In an embodiment of the present invention, obtaining features suitable for positioning within a proposal includes: (1) merging a portion suitable only for positioning with a shared portion; and (2) projecting back to the original space to obtain features of the video segment suitable only for positioning.
[0073] In an embodiment of the present invention, obtaining proposal-level features includes: (1) aggregating features suitable for classification within a proposal to obtain proposal-level classification features; (2) aggregating features suitable for positioning within a proposal to obtain proposal-level positioning features.
[0074] In an embodiment of the present invention, the step of obtaining the classification result specifically includes: scoring the action category of the classification features of the proposal level (assuming there are 20 types of actions, then one proposal has 20 scores), and the action category with the highest score is the predicted action category.
[0075] In this embodiment of the present invention, the steps of obtaining the positioning result specifically include: expanding the original proposal in time sequence by 1 / 2 of its original length in both directions, and treating the expanded portions as two new proposals; then passing these two new proposals through steps 1, 3, and 4 (the network used by the original proposal is the same), and concatenating the positioning features of the original proposal and the two new proposals to obtain the final proposal positioning features;
[0076] In the embodiment of the present invention, the final classification and positioning results are obtained: (1) a proposal completeness score is obtained using the final proposal positioning feature, and a proposal classification score of the corresponding original proposal feature is obtained, and these two scores are multiplied together to obtain the final proposal classification result. (2) The start time and end time of the action are obtained using the final proposal positioning feature.
[0077] See also Figure 2 , the shared input focuses on the "take-off" of the action "high jump" rather than the "run-up", so this shared input is more beneficial for classification, because some other actions such as "long jump" also include "run-up", which makes "run-up" a disturbance in classification; but this representation is not suitable for judging integrity and positioning actions. Judging integrity requires considering both "run-up" and "take-off", and positioning actions requires considering their start and end times.
[0078] See also Figure 3In addition to considering the “run-up” and “take-off” more evenly, the decoupled features of the present invention focus on capturing the start and end frames of the “high jump”, indicating that they are more suitable for the localization task. In addition, the classification features of the present invention still focus on the most discriminative frame of the “high jump” (“take-off”).
[0079] Based on the description of the above technical solutions and the exemplary comparison, it can be seen that the method of the present invention decouples the shared features to obtain features suitable for classification and features suitable for positioning respectively, so that the parts that classification and positioning focus on no longer affect each other, thereby improving the accuracy of classification and the precision of positioning. Specific embodiments
[0081] A method for localizing temporal actions based on decoupling classification features and localization features according to an embodiment of the present invention includes the following steps:
[0082] Step 1: Feature Extraction:
[0083] 1) The video is divided into 64-frame video segments, and then the video segment features are extracted through the I3D network and marked as x j , where j represents the jth video clip and x is a 1024-dimensional feature vector;
[0084] 2) Use the existing theoretical BSN to obtain the initial proposal N represents the number of proposals, p i represents the i-th proposal, t i,s represents the start time of the i-th proposal, t i,e Represents the end time of the i-th proposal;
[0085] 3) Get p i Internal video clip features, marked as K i Represents the number of video clips in the proposal;
[0086] Step 2: Build an autoencoder for decoupling:
[0087] 1) The features Y of these video clips i Projected into three different spaces, respectively, the parts suitable for classification only Only suitable for positioning part The parts that both parties are concerned about (the common parts)
[0088] The calculation formula is: in and is a trainable weight with a dimension of (256*1024); the superscript cls indicates classification, loc indicates positioning, and com indicates common; the subscript e indicates encoder;
[0089] 2) Reconstruct the features Y of the video clips within the proposal i : Splice the above three parts together and reconstruct the original video features through projection. The specific formula is as follows:
[0090] in, It is the reconstructed feature with a dimension of 1024*K i ;W d is a trainable weight with a dimension of (256*3*1024); σ is the ReLU activation function; [·] is the concatenation operation; the subscript d represents the decoder;
[0091] Step 3: Obtain the video segment features suitable for classification within the proposal: Use the all-zero matrix H zero Replace the noise (only the part suitable for positioning), and then obtain the video segment features suitable for classification within the proposal through projection:
[0092] in It is the video segment feature suitable for classification within the proposal, with a dimension of 1024*K i ;
[0093] Step 4: Obtain the video segment features suitable for positioning within the proposal: Use an all-zero matrix to replace the noise (only the part suitable for classification), and then obtain the video segment features suitable for classification within the proposal through projection:
[0094] in It is the video segment feature suitable for positioning within the proposal, with a dimension of 1024*K i ;
[0095] Step 5: and Aggregated into proposal-level classification features through the existing aggregation network ContextLoc and Positioning features, their dimensions are all 1024;
[0096] Step 6: Expand the original proposal to 1 / 2 of its original length in time, and treat the expanded parts as two new proposals to obtain the video clip features within the proposal. and The superscripts L and R indicate left and right expansion respectively. and Obtained through steps 4 and 5 and Will and original proposal positioning features The final positioning features are obtained by splicing;
[0097] Step 7: Convert the original proposal features Get the predicted classification score s through the fully connected layer o , the final positioning features of the corresponding expanded proposal are passed through two fully connected layers to obtain the predicted completeness score s e And the positioning result (boundary regression), the final proposal classification score is the product of the two scores s o ×s e .
[0098] Please refer to Table 1, which shows the performance comparison on the THUMOS14 and ActivityNet v1.3 datasets. The tIoU threshold on THUMOS14 is 0.3 to 0.7, and the mean of 0.3 to 0.7 is also reported. The tIoU threshold on ActivityNetv1.3 is 0.5 to 0.95, and the mean is 0.5 to 0.95 with a step size of 0.05. The middle of Table 1 is the performance comparison of the method of the present invention with the previous method on the THUMOS14 dataset, and the right side of Table 1 is the performance comparison result on the ActivityNet v1.3 dataset. The indicator for evaluating performance is mAP (mean Average Precision) under different tIoU thresholds. It can be seen from Table 1 that the present invention has been greatly improved over the previous method. This is a specific benefit brought about by the use of decoupling features in the present invention.
[0099] Table 1. Performance comparison on HUMOS14 and ActivityNet v1.3 datasets
[0100]
[0101] The following are device embodiments of the present invention, which can be used to perform the method embodiments of the present invention. For details not disclosed in the device embodiments, please refer to the method embodiments of the present invention.
[0102] In yet another embodiment of the present invention, a temporal action positioning system based on feature decoupling is provided, comprising:
[0103] The rough feature acquisition module is used to obtain the video to be temporally positioned and cut it into several video segments, and extract the rough features of each video segment;
[0104] A result acquisition module is used to input the rough features of each video clip into a pre-trained decoupling positioning model, and the decoupling positioning model outputs an action classification result and an action positioning result;
[0105] Wherein, the decoupling positioning model includes:
[0106] An autoencoder that takes the rough features of a video clip as input, decouples them, and outputs the fine classification features and fine positioning features of the corresponding video clip;
[0107] A classification network, configured to input the detailed classification features and output an action classification result;
[0108] The positioning network is used to input the detailed positioning features and output action positioning results.
[0109] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.
[0110] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0111] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.
[0112] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1A step that specifies a function in one or more boxes.
[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.
Claims
1. A temporal action localization method based on feature decoupling, characterized in that: The following steps are involved: Obtain the video to be temporally positioned and cut it into several video clips, and extract the rough features of each video clip; The rough features of each video clip are respectively input into a pre-trained decoupled positioning model, and the decoupled positioning model outputs an action classification result and an action positioning result; Wherein, the decoupled positioning model includes: The autoencoder is used to input the rough features of the video clip, decouple and output the fine classification features and fine positioning features of the corresponding video clip; A classification network, used for inputting the fine classification features and outputting action classification results; A positioning network, used for inputting the fine positioning features and outputting action positioning results; The autoencoder comprises: An encoder is used to input the rough features of the video clip and output three parts of latent variables; wherein the three parts of latent variables include latent variables suitable only for classification, latent variables suitable only for positioning, and latent variables suitable for both positioning and classification; A decoder, used for inputting the three parts of latent variables and outputting restored rough features; The steps of obtaining the pre-trained decoupled positioning model include: (1) Obtaining a sample training set; wherein each sample includes a rough feature of a sample video clip, an action location label, and an action classification label; (2) inputting the rough features of each sample video clip into an autoencoder, and outputting three parts of latent variables through the encoder in the autoencoder; inputting the three parts of latent variables into the decoder of the autoencoder to obtain the restored rough features; (3) processing the three parts of latent variables and inputting them into the decoder in the autoencoder to obtain fine classification features; inputting the fine classification features into the classification network to obtain the predicted classification results; wherein the processing rule is to replace any part of the three parts of latent variables with a 0 vector; (4) The three parts of latent variables are processed and input into the decoder in the autoencoder to obtain fine positioning features; the fine positioning features are input into the positioning network to obtain a predicted positioning result; wherein the processing rule is to replace any part of the two parts of latent variables remaining after replacing the 0 vector in step (3) with a 0 vector; (5) calculating a loss value of the decoupled positioning model, updating model parameters based on the loss value, and obtaining the pre-trained decoupled positioning model; In the step of inputting the rough features of each sample video clip into the autoencoder and outputting three parts of latent variables through the encoder in the autoencoder, the calculation formula of the three parts of latent variables output by the encoder is: In the formula, is a latent variable suitable only for classification, is a hidden variable suitable only for positioning, To locate the latent variables that are suitable for classification; and The encoder is a trainable weight; Y i is the rough feature of the sample video clip; In the step of inputting the three parts of latent variables into the decoder of the autoencoder to obtain the restored rough features, the expression of the restored rough features is: In the formula, is the rough feature of the restoration; W d is the trainable weight of the decoder; σ is the ReLU activation function; [·] is the concatenation operation; The expression for replacing any part of the three parts of hidden variables with a 0 vector is: In the formula, Features of video clips suitable for classification within the proposal; The expression for replacing any one of the two parts of hidden variables remaining after replacing the 0 vector in step (3) with the 0 vector is: In the formula, Features of video clips suitable for positioning within the proposal.
2. The method for locating temporal actions based on feature decoupling according to claim 1, characterized in that: The calculation expression of the loss value L of the decoupled positioning model is: L=l a +l c +l l ; In the formula, l a is the reconstruction loss of the autoencoder, l c is the classification loss, l l is the positioning loss.
3. The method for positioning sequential actions based on feature decoupling according to claim 2, characterized in that: The reconstruction loss of the autoencoder is cosine similarity or Euclidean distance loss; The classification loss is a cross entropy loss; The positioning loss is an absolute distance loss.
4. A temporal action positioning system based on feature decoupling, characterized in that: The method for locating a temporal action based on feature decoupling according to claim 1 comprises: A rough feature acquisition module is used to acquire the video to be temporally positioned and cut it into several video clips, and extract the rough features of each video clip; A result acquisition module, used for inputting the rough features of each video clip into a pre-trained decoupled positioning model, wherein the decoupled positioning model outputs an action classification result and an action positioning result; Wherein, the decoupled positioning model includes: The autoencoder is used to input the rough features of the video clip, decouple and output the fine classification features and fine positioning features of the corresponding video clip; A classification network, used for inputting the fine classification features and outputting action classification results; The positioning network is used to input the fine positioning features and output the action positioning results.
