Moving target intelligent detection method and system based on weak supervision dynamic optimization

Abstract

The invention provides a moving target intelligent detection method and system based on weak supervision dynamic optimization, and the method comprises the steps: respectively extracting video features and text features from an original video and a text, carrying out the fusion, and generating a frame-level semantic similarity score as a pseudo tag; utilizing learnable object query and fusion feature interaction to generate positive and negative proposal masks; guiding feature comparison learning of the positive proposal by using a pseudo tag, so that the positive proposal infinitely fits text features in a semantic space, and the negative proposal infinitely deviates from a related region of the text features; performing text reconstruction based on a mask condition Transform by using positive and negative proposal masks, and performing semantic consistency training on different proposals to obtain a video time domain positioning result; and dynamically optimizing a video time domain positioning result, generating a final positioning result, and completing intelligent detection of the moving target. According to the invention, by constructing a learnable negative proposal and a dynamic pseudo-label constraint mechanism, the time domain positioning precision under a weak supervision condition is significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of video timing target positioning, and in particular to a method and system for intelligent detection of moving targets based on weakly supervised dynamic optimization, as well as a corresponding computer terminal and computer-readable storage medium. Background Art

[0002] With the continuous advancement of artificial intelligence, video coding, and multimodal understanding technologies, video has become one of the primary media for information expression and dissemination in modern society. Compared to static images or text, video integrates multimodal information such as vision, speech, and semantics, offering stronger contextualization and temporal coherence. Consequently, it has been widely used in areas such as social media, content recommendation, and intelligent surveillance. Compared to images and text, it contains richer semantic information. Currently, many tasks related to video understanding exist, such as video action classification, video question answering, and video text retrieval. These tasks focus on understanding the content of short, edited videos. For understanding long, unedited videos, automatically locating a segment within a long video based on a user-provided text description remains a significant challenge.

[0003] The task of video temporal object localization aims to automatically locate semantically related temporal segments in a long, unedited video based on a natural language description. This challenge lies in the fact that models must possess both temporal modeling capabilities and cross-modal semantic alignment capabilities, and accurate temporal boundary annotations are often lacking during training. To reduce the cost of manual annotation, weakly supervised video temporal localization has become a research focus. Its goal is to learn the alignment relationship between text and video segments in the absence of start and end time labels. Existing weakly supervised localization methods can be roughly divided into two categories: one is based on multiple instance learning (MIL) (such as WSLLN), which scores multiple proposals and selects the most relevant proposal as a positive sample; the other is based on reconstruction, which constructs a set of candidate proposals, uses them to reconstruct the text, and optimizes the model training process through reconstruction error (such as WSTAN). Both are designed within this framework. However, methods based on MIL suffer from difficulties in distinguishing action from background and incomplete localization. Methods based on reconstruction also face challenges such as inconsistency between the reconstruction target and the localization task and insufficient negative sample modeling, which compromise the model's semantic alignment and discriminative performance. Therefore, existing methods have shortcomings in terms of the accuracy of proposal generation, the ability to distinguish positive and negative samples, and the consistency of training objectives.

[0004] In recent years, the success of DETR in image object detection has inspired some researchers to apply it to temporal localization scenarios. Inspired by this, some methods incorporate learnable object queries and a Transformer decoder into a proposal generation module, achieving an end-to-end learnable proposal prediction mechanism. These methods reduce reliance on human-designed features while improving the expressiveness and structural flexibility of proposals. Nevertheless, generating discriminative negative proposals in weakly supervised scenarios remains a key challenge. On the one hand, statically generated negative proposals fail to cover complex and diverse video clutter regions, resulting in a lack of supervisory signal during training. On the other hand, introducing "hard negative examples" that are too close to semantic boundaries in the early stages of training can mislead the model and lead to error accumulation. To this end, some studies have proposed using pseudo-label mechanisms as soft supervisory signals to guide the model in learning the semantic alignment between proposals and text. However, most of these methods use pseudo-labels as classification supervision signals or as static scoring criteria, failing to explicitly incorporate the continuity information of pseudo-labels into the feature space for structural optimization. The lack of a pseudo-label-based feature comparison mechanism results in limited structural distinguishability between positive and negative proposals in the semantic space. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method and system for intelligent detection of moving targets based on weakly supervised dynamic optimization, and also provides a corresponding computer terminal and computer-readable storage medium.

[0006] According to one aspect of the present invention, a method for intelligent detection of moving targets based on weakly supervised dynamic optimization is provided, comprising:

[0007] Extracting video features and text features of the original video and text respectively, and fusing the video features and the text features to obtain cross-modal fusion features;

[0008] Based on the fused features, generating a frame-level semantic similarity score as a pseudo label;

[0009] Generate positive and negative proposal masks by using learnable object queries and fused features;

[0010] Using the pseudo labels to guide feature contrast learning of positive proposals, the positive proposals are made to infinitely fit the text features in the semantic space, and the negative proposals are made to infinitely deviate from the relevant areas of the text features;

[0011] Using the positive and negative proposal masks, text reconstruction is performed based on the proposal mask conditional Transformer, and semantic consistency training is performed on different proposals to obtain the video time domain positioning result;

[0012] The video time domain positioning result is dynamically optimized to generate a final positioning result and complete the intelligent detection of moving targets.

[0013] According to another aspect of the present invention, there is provided a moving target intelligent detection system based on weakly supervised dynamic optimization, comprising:

[0014] A feature extraction module is used to extract video features and text features of the original video and text respectively, and fuse the video features and the text features to obtain cross-modal fusion features;

[0015] a pseudo-label construction module, which generates a frame-level semantic similarity score as a pseudo-label based on the fused features;

[0016] A proposal generation module that uses learnable object queries and fused features to generate positive and negative proposal masks;

[0017] A proposal optimization module, which uses the pseudo-labels to guide feature contrast learning of positive proposals, so that the positive proposals infinitely fit the text features in the semantic space and the negative proposals infinitely deviate from the relevant areas of the text features;

[0018] A text reconstruction module, which uses the positive and negative proposal masks to reconstruct text based on the proposal mask conditional Transformer, and performs semantic consistency training on different proposals to obtain video temporal localization results;

[0019] The positioning optimization module is used to dynamically optimize the video time domain positioning result, generate the final positioning result, and complete the intelligent detection of moving targets.

[0020] According to a third aspect of the present invention, a computer terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the terminal can be used to execute the method described above in the present invention, or to execute the system described above in the present invention.

[0021] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it can be used to execute the method described above in the present invention, or to run the system described above in the present invention.

[0022] Due to the adoption of the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art:

[0023] The intelligent moving object detection method and system based on weakly supervised dynamic optimization, designed for weakly supervised temporal object localization in video, is implemented by dividing the alignment process between positive and negative proposals and textual semantics into two stages: pseudo-label generation and proposal guidance. Pseudo-labels are first generated by modeling the correlation between video and textual features. The pseudo-labels are then used to guide positive and negative proposals toward or away from semantic regions. This provides a stable and reliable supervision signal, improving proposal quality and semantic matching accuracy.

[0024] The intelligent detection method and system for moving targets based on weakly supervised dynamic optimization provided by the present invention utilizes weighted reconstruction and semantic comparison mechanisms to achieve dynamic modeling of the difficulty of negative proposals and fine optimization of temporal boundaries. First, semantic modeling is performed on each group of negative proposal masks through a mask-conditional Transformer, and the degree of alignment with the text description is evaluated in combination with the reconstruction error. Weights are assigned accordingly to generate weighted negative proposal masks. Afterwards, a progressive cross-entropy loss is designed. Through the weight coefficient that increases with the training rounds, the model is guided to learn from "semantically irrelevant" to "semantically similar" negative proposals from shallow to deep. This not only improves the model's ability to discriminate against negative samples, but also strengthens the model's semantic perception of the boundary areas of text descriptions, thereby achieving a more refined video temporal positioning effect under weak supervision conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0026] Figure 1 This is a workflow diagram of a moving target intelligent detection method based on weakly supervised dynamic optimization in a preferred embodiment of the present invention.

[0027] Figure 2 Schematic diagram of the components of a moving target intelligent detection system based on weakly supervised dynamic optimization in a preferred embodiment of the present invention.

[0028] Figure 3 The figure is a flow chart of a method for intelligent detection of moving targets based on weakly supervised dynamic optimization in a specific application example of the present invention.

[0029] Figure 4 The figure is a flowchart of weighted negative proposal mask generation and mask reconstruction in a specific application example of the present invention. DETAILED DESCRIPTION

[0030] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. It should be noted that those skilled in the art may make various modifications and improvements without departing from the scope of the present invention, and these modifications and improvements fall within the scope of protection of the present invention.

[0031] In the existing technology, when it comes to video temporal target localization tasks, traditional methods face great difficulties in constructing positive and negative proposals or achieving semantic alignment due to the lack of precise timestamp annotations. On the one hand, statically generated negative proposals cannot cover complex and diverse video interference areas, resulting in a scarcity of supervisory signals during the training process; on the other hand, if difficult negative samples that are too close to the semantic boundary are introduced in the early training stage, it may mislead the model and lead to error accumulation. In addition, existing methods lack an effective semantic comparison mechanism in the proposal construction process, making it difficult to accurately evaluate the semantic consistency between positive and negative proposals and text, resulting in unstable proposal quality; at the same time, in weakly supervised training, pseudo-labels are usually only used as classification supervision signals or static scoring basis, lacking deep integration and comparative modeling with the proposal structure, and failing to fully improve the discrimination ability of positive and negative proposals in the semantic space, thereby limiting the model's ability in fine-grained semantic alignment and boundary discrimination.

[0032] To address the above problems, an embodiment of the present invention provides an intelligent detection method for moving targets based on weakly supervised dynamic optimization. This method designs a new paradigm for solving the task of video time domain target positioning. It obtains the semantic correlation information between text and video by introducing a pseudo-label mechanism, and combines it with a learnable Gaussian proposal structure. Through a dynamic curriculum training strategy, it guides and optimizes positive and negative proposals, constructing a temporal positioning technology with stronger semantic expression and discrimination capabilities.

[0033] Specifically, if Figure 1 As shown, the moving target intelligent detection method based on weakly supervised dynamic optimization provided by this embodiment may include:

[0034] S1, extracts video features and text features from the original video and text respectively, and fuses the video features and text features to obtain cross-modal fusion features;

[0035] S2, based on the fused features, generates frame-level semantic similarity scores as pseudo labels;

[0036] S3, uses learnable object queries to interact with fused features to generate positive and negative proposal masks;

[0037] S4 uses pseudo labels to guide the feature contrast learning of positive proposals, so that positive proposals can infinitely fit the text features in the semantic space and negative proposals can infinitely deviate from the relevant areas of text features;

[0038] S5, using positive and negative proposal masks, performs text reconstruction based on the proposal mask conditional transformer (MCT), and performs semantic consistency training on different proposals to obtain video temporal localization results;

[0039] S6, dynamically optimizes the video time domain positioning results, generates the final positioning results, and completes the intelligent detection of moving targets.

[0040] In some preferred embodiments, the above S1, which extracts video features and text features of the original video and text respectively, and fuses the video features and text features to obtain cross-modal fusion features, may further include:

[0041] S11: For a given raw video and text, a pre-trained 3D convolutional network is used to extract temporal features of the video frames, and a pre-trained GloVe model is used to extract word vector features of the text to obtain basic semantic representations.

[0042] In S12, the time series features and word vector features are mapped to a unified embedding space through a fully connected layer, and the context information of the time series features and word vector features is encoded using Transformer to obtain a cross-modal fusion feature H that contains semantic and time series relationships, expressed as:

[0043] H=Dec(V,Enc(Q))

[0044] Where Enc(·) represents the encoder; Dec(·) represents the decoder.

[0045] In some preferred embodiments, the above S2, generating a frame-level semantic similarity score as a pseudo label based on the fusion feature, may further include:

[0046] In the pseudo-label construction stage, the cross-modal feature H is processed by a feed-forward neural network (FFN) to predict the semantic relevance score S between each time step and the text description c .

[0047] S21, the fusion feature H is input into the feed-forward neural network FFN to obtain the correlation score between each frame of video and text as a pseudo label, which is expressed as:

[0048] S c =σ(FFN c (H))

[0049] Where S c is a pseudo label; σ is a Sigmoid activation function, which indicates the semantic relevance between each frame of video and text; FFN c It is a feed-forward neural network.

[0050] A higher score indicates a stronger semantic relevance between the video clip and the text. This score serves as a pseudo-label during the training phase to help guide feature comparisons between positive and negative proposals.

[0051] In some preferred embodiments, the above S3, using the learnable object query and the fusion feature to interact to generate positive and negative proposal masks, may further include:

[0052] S32, the generation of proposal mask is based on DETR decoder and learnable query mechanism. Introducing K positive query vectors Q p ∈R K×D and mK negative query vectors Q n ∈R K×D , and input it into a proposal generation model PFG (Proposal Feature Generator) together with the fusion feature H to obtain positive and negative proposal features, expressed as:

[0053] F p =PFG(Q p ,H),F n =PFG(Q n ,H)

[0054] Where, F p is a positive proposal feature; F n is the negative proposal feature; R represents the real number space; D is the dimension of each query vector;

[0055] S32, based on F p 、F n Predict the center position c of the corresponding time period k and width ω k , further construct the Gaussian mask function in the time dimension, which is defined as follows:

[0056]

[0057] Where, f (k) (t) is the value of the k-th time proposal at the t-th time position, which is used to construct the Gaussian mask corresponding to the k-th time proposal; N is the total number of frames in the video; t is the index of the current time frame;

[0058] S33, the output of the Gaussian mask function generated based on the positive and negative proposal features at each time position is stacked row by row, and the positive proposal mask set M can be constructed respectively. p ∈R K×N , negative proposal mask set M n ∈R mK×N, which is used for subsequent feature comparison and reconstruction training. To alleviate the problem of lack of precise boundary labels in weakly supervised scenarios, a pseudo-label-guided contrast loss is introduced to achieve soft-supervised learning of semantic boundaries by constructing the relative relationship between positive and negative proposals and text.

[0059] In some preferred embodiments, the above S4, which uses pseudo labels to guide feature contrast learning of positive proposals, so that the positive proposals infinitely fit the text features in the semantic space and the negative proposals infinitely deviate from the relevant areas of the text features, may further include:

[0060] S41, in order to improve the semantic discrimination ability of positive and negative proposals, a pseudo-label guided contrast loss L is constructed p As shown below:

[0061]

[0062] Where, is the i-th negative proposal mask in the negative proposal mask set; is the i-th positive proposal mask in the positive proposal mask set; K is the number of positive query vectors; S c is a pseudo label; α p is a hyperparameter that controls the contrast interval; T is the matrix transpose symbol;

[0063] S42, using pseudo labels to guide contrast loss L p , guiding the feature contrast learning of positive proposals, so that in the semantic space, positive proposals infinitely fit the text features, while negative proposals infinitely deviate from the text features.

[0064] This loss encourages positive proposals to have a higher weighted total score on the pseudo-label score than negative proposals, which means that positive proposals are more consistent with the text semantics, while negative proposals are far away from semantically related areas, thereby improving the distinguishability of positive and negative proposals and providing high-quality candidates for subsequent reconstruction modules.

[0065] In some preferred embodiments, the above S5, which performs text reconstruction based on the proposal mask conditional Transformer and performs semantic consistency training on different proposals, may further include:

[0066] To further improve the model's ability to discriminate easily confused fragments, a progressive negative proposal constraint mechanism is introduced. First, a mask-conditioned Transformer is used to extract the multimodal reconstruction features corresponding to each negative proposal mask and predict the reconstruction error of the text description. To achieve a dynamic training strategy from easy to difficult, a weighting coefficient is introduced that gradually increases with each training round.

[0067] S51, in order to improve the semantic reconstruction capability, the mask conditional Transformer is used to reconstruct the text from the positive proposal mask and the weighted negative proposal mask;

[0068] The weighted negative proposal is expressed as:

[0069]

[0070] Where, is the weighted negative proposal mask corresponding to the k-th positive proposal; is the Gaussian mask of the i-th negative proposal associated with the k-th positive proposal; m is the number of candidate negative proposals corresponding to each positive proposal; ω i is the importance weight of the i-th negative proposal mask, which is obtained as follows:

[0071]

[0072] Where FFN w It is a feedforward neural network; is the negative proposal mask of group i Multimodal features CLS characterization;

[0073] The positive proposal mask, weighted negative proposal mask, and the mask corresponding to the entire original video are input into Transformer along with the masked text, the masked text is predicted, the original text is reconstructed, and the text reconstruction loss L is calculated. rec , perform semantic consistency evaluation on different proposals; wherein, the masked text refers to the masked input formed by randomly masking a set proportion of words from the original text

[0074] Furthermore, the text reconstruction loss L rec , expressed as:

[0075]

[0076] Where, denote the cross entropy loss of positive proposal, reference hard negative proposal and progressive negative proposal, respectively, k * The number of the positive proposal with the smallest current loss.

[0077] S52: Select the sample with the largest reconstruction error in each group of negative proposals as the training target, and perform semantic consistency training on different proposals; let the current training round be e, and the maximum number of rounds be e max , define the dynamic adjustment weight as:

[0078]

[0079] Where β and α are hyperparameters that control the training progress;

[0080] Final progressive negative proposal cross entropy loss Expressed as:

[0081]

[0082] Where, is the cross entropy loss of weighted negative proposals; k * The number of the positive proposal with the smallest current loss.

[0083] This loss mechanism prompts the model to focus on negative samples with a long semantic distance in the early stage, and gradually guides it to identify "hard negative samples" that are close to the text semantics but irrelevant in the later stage, effectively improving the semantic accuracy and generalization ability of the model's time domain positioning.

[0084] In some preferred implementations, the above S6, dynamically optimizing the video time domain positioning result to generate the final positioning result, may further include:

[0085] The prediction output of the final time boundary is achieved by using a multi-loss joint optimization mechanism; among them:

[0086] Multiple losses include: pseudo-label guided contrast loss L p , text reconstruction loss L rec And a video contrast loss L ivc and a proposal diversity loss L div ; The final optimization goal of the joint optimization is:

[0087] L total =λ p L p +λ rec L rec +λ ivc L ivc +λ div L div

[0088] Where λ p ,λ rec ,λ ivc and λ div Respectively.

[0089] Furthermore, the contrast loss L ivc and proposal diversity loss L div Obtained through the following methods:

[0090]

[0091] Where, Denote the cross entropy loss of positive proposals and reference negative proposals, respectively, and k * is the number of the positive proposal with the minimum current loss, and β1 and β2 are the hyperparameters of the comparison interval.

[0092] L div =‖Mp (M p ) T -λI‖ F

[0093] Where M p is the set of positive proposal masks, T is the matrix transpose symbol, λ is the control coefficient for adjusting the overlap, I is the unit matrix, ‖·‖ F is the Frobenius norm.

[0094] The above embodiment of the present invention provides an intelligent moving target detection method based on weakly supervised dynamic optimization, which includes the following technical advantages:

[0095] Given raw video and text, we first use a pre-trained 3D convolutional network to extract temporal features from the video frames, and the GloVe model to extract word vector representations from the text. Then, we map these two features into a unified embedding space using a fully connected layer, and use a Transformer to encode the contextual information of the video and text, obtaining a basic representation that incorporates both semantic and temporal relationships, thus generating cross-modal fusion features.

[0096] A pseudo-label supervision mechanism is introduced, combined with the DETR decoder structure, to generate the center and width of positive and negative proposals and construct a learnable Gaussian mask. Pseudo-labels guide the semantic alignment of proposals with text, and contrast loss is used to optimize the discrimination between positive and negative proposals. While inheriting the weak supervision concept, this approach improves structural expression and boundary discrimination. By fusing video and text features, frame-level semantic similarity scores are generated as pseudo-labels, and contrast loss is introduced to make the distribution of positive and negative proposals in the semantic space more discriminative.

[0097] To alleviate the problem of overly random negative proposals in early training and a lack of supervisory value in later stages, a progressive negative proposal modeling mechanism is introduced. This dynamically adjusts the difficulty of negative proposals according to the training stage, gradually transitioning from semantically irrelevant easy negative proposals to difficult negative proposals that overlap with the text description. This effectively improves the model's ability to discriminate at the boundary between positive and negative samples and enhances training stability. By interactively generating proposal masks using learnable queries and fused features, a progressive negative proposal modeling mechanism is designed to achieve semantic modeling from easy to difficult, improving boundary discrimination and negative sample quality.

[0098] With the help of Mask Conditioned Transformer (MCT), the proposal information is used to reconstruct the masked text, thereby evaluating the semantic coverage of the proposal. During the training process, positive proposals should be able to accurately reconstruct the text and represent semantic segments that are highly consistent with the description. Negative proposals are far away from the text semantics in the early stages of training and difficult to restore the text. As training progresses, content with a certain degree of semantic overlap with the text description is gradually introduced, causing it to produce greater prediction errors during the reconstruction process, thereby strengthening the distinction between positive and negative proposals in semantic modeling, improving boundary judgment capabilities and proposal expression effects. By reconstructing the original text, the semantic expression capabilities of various proposals are evaluated, and the semantic separation effect of the proposals is enhanced.

[0099] The contrastive loss uses frame-level semantic scores to constrain the semantic distinction between positive and negative proposals; the diversity loss encourages differences between positive proposals to prevent them from converging to the same location; the text reconstruction loss measures the proposal's coverage of the descriptive semantics; and the intra-video contrastive loss further emphasizes the relative merits of positive proposals, negative proposals, and reference proposals. Multiple losses are collaboratively optimized to improve positioning accuracy and stability from three perspectives: semantic representation, temporal boundaries, and reconstruction capabilities. By integrating multiple losses, including pseudo-label-guided contrastive loss, reconstruction loss, diversity, and contrastive loss, semantic alignment quality and boundary differentiation are comprehensively optimized, enhancing the model's robustness in positioning under weak supervision.

[0100] In summary, the method provided by the above embodiment of the present invention designs a learnable proposal mechanism based on pseudo-label guidance, which realizes the video time domain target positioning task under weak supervision. First, the original input video and text are extracted to obtain semantic feature representations containing contextual information; secondly, the extracted video features are cross-modally fused with the text features, and the semantic relevance score between each frame and the text is generated through the prediction network, and pseudo labels are constructed for subsequent training supervision; then, pseudo labels are used to guide the generation of learnable positive and negative proposals, and the discrimination ability between positive and negative proposals is optimized through semantic contrast loss; further, mask conditional reconstruction technology is introduced to evaluate the semantic consistency of different proposals; finally, the training difficulty is dynamically adjusted through a progressive training strategy, guiding the model to gradually learn more complex negative sample features, thereby improving boundary discrimination ability and semantic perception ability. Without relying on manual annotation of time boundaries, this method effectively alleviates the problem of insufficient training signals in weak supervision scenarios by constructing a proposal mechanism that combines pseudo-supervisory signals with learnable structures, improves the semantic alignment effect between text and video, and has good positioning accuracy and generalization ability.

[0101] Based on the same inventive concept, an embodiment of the present invention further provides a moving target intelligent detection system based on weakly supervised dynamic optimization.

[0102] Specifically, if Figure 2 As shown, the moving target intelligent detection system based on weakly supervised dynamic optimization provided by this embodiment may include:

[0103] Feature extraction module, which is used to extract video features and text features from the original video and text respectively, and fuse the video features and text features to obtain cross-modal fusion features;

[0104] Pseudo-label construction module, which generates frame-level semantic similarity scores as pseudo-labels based on fused features;

[0105] A proposal generation module that uses learnable object queries and fused features to generate positive and negative proposal masks;

[0106] Proposal optimization module, which uses pseudo-labels to guide feature contrast learning of positive proposals, making positive proposals infinitely fit text features in the semantic space and making negative proposals infinitely deviate from the relevant areas of text features;

[0107] The text reconstruction module uses positive and negative proposal masks to reconstruct text based on the proposal mask conditional Transformer, and performs semantic consistency training on different proposals to obtain video temporal localization results;

[0108] Positioning optimization module, which is used to dynamically optimize the video time domain positioning results, generate the final positioning results, and complete the intelligent detection of moving targets.

[0109] The specific contents of the functional modules constituting the intelligent moving target detection system according to the above embodiment of the present invention are further described below.

[0110] The feature extraction module is used to extract features from the original video and text respectively, obtain basic semantic representation, and fuse the extracted features to obtain cross-modal fusion features. It further includes:

[0111] Given an uncut video sequence and the corresponding natural language description First, use the pre-trained 3D convolutional neural network to extract video frame features and obtain Use the pre-trained GloVe model to extract the semantic embedding of each word and obtain text features

[0112] The text features are then input into the Transformer encoder for context modeling to obtain Enc(Q), which is then interacted with the video features in the Transformer decoder to construct cross-modal fusion features, expressed as: H∈R X×D :

[0113] H=Dec(V,Enc(Q))

[0114] Where X is the number of extracted video frames; D is the hidden representation dimension of the cross-modal fusion feature; Enc(·) represents the encoder, and Dec(·) represents the decoder.

[0115] The generated fusion feature H will serve as the basic feature input for subsequent proposal generation and semantic supervision.

[0116] The pseudo-label construction module generates frame-level semantic similarity scores as pseudo-labels based on fused features. It further includes:

[0117] The fusion feature H is input into the feedforward neural network FFN to obtain the correlation score between each frame of video and text:

[0118] S c =σ(FFN c (H))

[0119] Among them, σ is the Sigmoid activation function, S c As a pseudo-label, it indicates the semantic relevance of each frame to the text.

[0120] The proposal generation module introduces contrast loss to make the distribution of positive and negative proposals in the semantic space more discriminative. The process of constructing positive and negative proposals further includes:

[0121] Introduce K positive query vectors Q p ∈R K×D and mK negative query vectors Q n ∈R K×D , together with H, is input into the proposal generation model PFG (Proposal Feature Generator) to obtain the proposal features:

[0122] F p =PFG(Q p ,H),F n =PFG(Q n ,H)

[0123] Based on F p 、F n Predict the center position c of the corresponding time period k and width ω k , further construct the Gaussian mask function in the time dimension, which is defined as follows:

[0124]

[0125] Where, f (k)(t) is the value of the k-th time proposal at the t-th time position, which is used to construct the Gaussian mask corresponding to the k-th time proposal; N is the total number of frames in the video; t is the index of the current time frame;

[0126] The output of the Gaussian mask function generated based on the positive and negative proposal features at each time position is stacked row by row, and the positive proposal mask set M can be constructed respectively. p ∈R K×N , negative proposal mask set M n ∈R mK×N

[0127] The proposal optimization module uses learnable queries and fused features to interactively generate proposal masks, designs a progressive negative proposal modeling mechanism, implements semantic modeling from easy to difficult, and improves boundary discrimination and negative sample quality. It further includes:

[0128] In order to improve the semantic discrimination ability of positive and negative proposals, a pseudo-label guided contrast loss is constructed as follows:

[0129]

[0130] Among them, α p is a hyperparameter that controls the comparison interval.

[0131] The text reconstruction module evaluates the semantic expression ability of various proposals by reconstructing the original text and enhances the semantic separation effect of proposals. It further includes:

[0132] To improve the semantic reconstruction capability of the model, a mask-conditional Transformer module is used to reconstruct text from positive proposals and weighted negative proposals. The negative proposal mask weighting method is as follows:

[0133]

[0134] Among them, ω i is the importance weight of the i-th negative proposal mask, which is obtained as follows:

[0135]

[0136] The positive proposal mask, weighted negative proposal mask and reference proposal mask are input into Transformer with the masked text to reconstruct the original text and calculate the cross entropy loss L ce as a training signal.

[0137] In order to implement a negative sample training strategy with progressive difficulty, a progressive cross entropy loss is introduced, and its weight is defined as:

[0138]

[0139] Among them, e is the current training round, e max is the maximum number of rounds, and β and α are hyperparameters that control the training progress.

[0140] The final progressive negative proposal constraint loss is expressed as:

[0141]

[0142] Among them, k * The number of the positive proposal with the smallest current loss.

[0143] The positioning optimization module integrates pseudo-label guided contrast loss, reconstruction loss, diversity and contrast loss, and other modules to comprehensively optimize the semantic alignment quality and boundary differentiation ability, and improve the positioning robustness of the model under weak supervision conditions. It further includes:

[0144] Integrate multiple loss functions to build a multi-objective optimization framework. First, the pseudo-label guided positive and negative proposal comparison loss L p Used to improve the semantic discrimination ability of proposals; text reconstruction loss L rec It consists of a positive proposal and a reference proposal, and is used to optimize the alignment between the proposal and the language semantics.

[0145] Furthermore, the intra-video contrast loss L is introduced ivc , encourages positive proposals to have advantages over other fragments in semantic reconstruction; at the same time, the proposal diversity loss L is added div , to prevent multiple proposal masks from highly overlapping and improve the structural expression ability. The final training goal is:

[0146] L total =λ p L p +λ rec L rec +λ cvc L ivc +λ div L div .

[0147] The following is a detailed description of the technical solution provided by the above embodiment of the present invention in conjunction with a specific application example. In this specific application example, the workflow diagram of the moving target intelligent detection method based on weakly supervised dynamic optimization is as follows: Figure 3 shown.

[0148] The first step is to use the feature extractor to extract relevant features from the original input video and text, and then perform feature fusion;

[0149] Given an uncut video sequence and the corresponding natural language description This step first uses the pre-trained 3D convolutional neural network to extract video frame features and obtain Use GloVe to extract the semantic embedding of each word and obtain text features The text features are then input into the Transformer encoder for context modeling to obtain Enc(Q), which is then interacted with the video features in the Transformer decoder to construct a cross-modal fusion feature representation H∈R X×D :

[0150] H=Dec(V,Enc(Q))

[0151] The fused feature H will serve as the basic feature input for subsequent proposal generation and semantic supervision.

[0152] In the second step, frame-level pseudo labels are generated based on the fused features, and Gaussian mask proposals are constructed in combination with learnable queries to guide and distinguish positive and negative proposals.

[0153] When obtaining cross-modal fusion features H∈R X×D After that, this step first predicts the semantic relevance of each frame of video clip through a feedforward neural network to obtain a pseudo-label score sequence S c ,Right now:

[0154] S c =σ(FFN c (H))

[0155] Among them, σ is the Sigmoid function, FFN c Represents the feedforward prediction head. The pseudo-label is used to measure the semantic match between each time step and the text. The pseudo-label will provide a soft guidance signal at the frame level in the subsequent positive and negative proposal supervision.

[0156] The third step is to establish learnable temporal candidate segments under weak supervision. This step introduces the learnable query mechanism in the DETR decoder structure. K positive proposal queries Q are constructed respectively. p With mK negative proposal queries Q n , and interacts with the fusion feature H to generate the corresponding proposal representation:

[0157] F p =PFG(Q p ,H),F n =PFG(Q n ,H)

[0158] Based on F p 、F n Predict the center position c of the corresponding time period k and width ω k , further construct the Gaussian mask function in the time dimension, which is defined as follows:

[0159]

[0160] Where, f (k) (t) is the value of the k-th time proposal at the t-th time position, which is used to construct the Gaussian mask corresponding to the k-th time proposal; N is the total number of frames in the video; t is the index of the current time frame;

[0161] The output of the Gaussian mask function generated based on the positive and negative proposal features at each time position is stacked row by row, and the positive proposal mask set M can be constructed respectively. p ∈R K×N , negative proposal mask set M n ∈R mK×N

[0162] In the fourth step, to improve the semantic discrimination ability between proposals, a pseudo-label-guided contrast loss function is introduced to constrain the difference in semantic scores between positive and negative proposals. It is defined as follows:

[0163]

[0164] Among them, α p To control the hyperparameters of the interval, we ensure that positive proposals are closer to the semantic region of the text, while negative proposals remain away from it. Through this mechanism, the model can obtain stable semantic guidance in weakly supervised scenarios, thereby optimizing proposal quality and matching accuracy.

[0165] The fifth step is to introduce a mask-conditional text reconstruction mechanism, which guides the model to learn more refined semantic alignment and context restoration capabilities by weighted fusion of multiple negative proposal masks, such as Figure 4 shown.

[0166] To further improve the semantic alignment between proposals and text, this step introduces a mask-conditional Transformer module to assist in optimizing the semantic representation of proposals using the text reconstruction task. In particular, to address the diversity of negative proposals, a weighted mask fusion mechanism is designed to dynamically select representative masks from multiple negative proposals.

[0167] First, for each set of negative proposal masks Extracting multimodal features And take its CLS characterization Input the feedforward network to calculate the weights and perform Softmax normalization:

[0168]

[0169] Based on this, we get the weighted fused negative proposal mask:

[0170]

[0171] Combine this mask with the positive proposal mask, reference mask, and mask text Input the reconstruction module together, predict the obscured text, and calculate the reconstruction loss L rec , thereby enhancing the model's perception of the semantic coverage of the proposal.

[0172] By using loss weights that change dynamically with each training round, we perform progressive difficulty modeling on negative proposals, thereby gradually improving the model's ability to perceive and discriminate semantic boundary areas.

[0173] To improve the model's ability to discern semantic boundaries, this step introduces a progressive training strategy into the text reconstruction module to dynamically control the difficulty of negative proposal supervision. By adjusting the influence coefficient of the negative proposal cross-entropy loss, the training process gradually transitions from negative examples with clear semantic distinctions to more challenging negative examples with blurred boundaries.

[0174] Specifically, we first select the sample with the largest reconstruction error in each group of negative proposals as the training target. Let the current training round be e and the maximum number of rounds be e max , define the dynamic adjustment weight as:

[0175]

[0176] Among them, β and α are hyperparameters that control the intensity and rate of change. The final negative proposal cross entropy loss is:

[0177]

[0178] where k * The number of the positive proposal with the lowest current loss is assigned. This mechanism ensures that the model initially focuses on learning easily distinguishable negative samples, and gradually focuses on samples with blurred boundaries in the later stages, achieving dynamic refinement of the semantic understanding granularity.

[0179] In the sixth step, a multi-loss joint optimization mechanism and learnable proposal parameter regression are used to achieve the prediction output of the final time boundary;

[0180] In the model training phase, this step integrates multiple loss functions to build a multi-objective optimization framework. First, the pseudo-label guided positive and negative proposal comparison loss L p Used to improve the semantic discrimination ability of proposals; text reconstruction loss L rec It consists of a positive proposal and a reference proposal, and is used to optimize the alignment between the proposal and the language semantics.

[0181] Furthermore, the intra-video contrast loss L is introduced ivc , encourages positive proposals to have advantages over other fragments in semantic reconstruction; at the same time, the proposal diversity loss L is added div, to prevent multiple proposal masks from highly overlapping and improve the structural expression ability. The final training goal is:

[0182] L total =λ p L p +λ rec L rec +λ ivc L ivc +λ div L div

[0183] The technical effects of the technical solutions provided by the above embodiments of the present invention are evaluated below in conjunction with a specific verification example.

[0184] In this specific verification example:

[0185] 1. Data composition, including:

[0186] The evaluation datasets used were Charades-STA and ActivityNet Captions. For video, Charades-STA used a pre-trained I3D model for feature extraction, while ActivityNet Captions used a C3D model. For text, word embedding was uniformly implemented using the GloVe embedding method, with a 300-dimensional word embedding vector extracted for each word in the text description. These features served as input for multimodal fusion and were subsequently used for cross-modal alignment and proposal modeling.

[0187] 2. Evaluation criteria include:

[0188] R@a,IoU≥b and mIoU are used as evaluation metrics. r@a,IoU≥b represents the percentage of test samples among the top-a prediction results where at least one prediction has an IoU value greater than or equal to b with the ground truth. mIoU represents the average IoU of all test samples. Here, a∈{1,5} and b∈{0.1,0.3,0.5,0.7} are set.

[0189] 3. Comparison of algorithms, including:

[0190] In comparative experiments, we compared the performance of several state-of-the-art algorithms in recent years, including CTF, RTBPN, VCA, LCNet, CPL, and CPL-R, on the Charades-STA and ActivityNet Captions datasets. It is worth noting that during the experiment, the performance results of the CPL method reproduced differed significantly from those given in the original paper, especially on the ActivityNet Captions dataset. Therefore, for a fairer comparison, we named the CPL method's reproduction results CPL-R and included them in the comparison.

[0191] Table 1 compares the performance of our proposed method with several existing methods on the Charades-STA validation set. As can be seen from the table, the proposed temporal target localization algorithm achieves a 4.59% improvement over the current best baseline method, CPL-R, under the evaluation metric R@1, IoU = 0.7, demonstrating a strong positioning accuracy advantage. Furthermore, our proposed method surpasses CPL-R by 3.17% and 1.01% in R@1mIoU and R@5mIoU, respectively, demonstrating the proposed mechanism's superiority in overall positioning accuracy.

[0192] Table 2 compares the proposed method with existing representative methods on the ActivityNet Captions dataset. As can be seen from the table, under the R@1 metric, the proposed method's overall performance is slightly better than the baseline methods CPL and CPL-R. In terms of the R@5 metric, the proposed method significantly outperforms CPL in multiple settings, demonstrating strong localization and recall capabilities. The proposed method demonstrates good performance under fully weak supervision, validating its feasibility in practical applications.

[0193] Therefore, as verified by comparative experiments, the algorithm proposed in this invention, under the synergy of the learnable proposal mechanism and pseudo-label supervision, can more effectively capture time segment features that are highly relevant to text semantics, strengthen the cross-modal alignment between video and text, and improve the accuracy and robustness of positioning boundaries.

[0194] Table 1 Performance comparison of the proposed algorithm on the Charades-STA validation set

[0195]

[0196] Table 2 Performance comparison of the proposed algorithm on the ActivityNet Captions validation set

[0197]

[0198]

[0199] An embodiment of the present invention further provides a computer terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the processor can be used to execute any one of the methods described in the foregoing embodiments of the present invention, or to execute any one of the systems described in the foregoing embodiments of the present invention.

[0200] Optionally, the memory is used to store programs; the memory may include volatile memory (English: volatile memory), such as random-access memory (English: random-access memory, abbreviated: RAM), such as static random-access memory (English: static random-access memory, abbreviated: SRAM), double data rate synchronous dynamic random access memory (English: Double Data Rate Synchronous Dynamic Random Access Memory, abbreviated: DDR SDRAM), etc.; the memory may also include non-volatile memory (English: non-volatile memory), such as flash memory (English: flash memory). The memory is used to store computer programs (such as applications, functional modules, etc. that implement the above-mentioned methods), computer instructions, etc., and the above-mentioned computer programs, computer instructions, etc. can be partitioned and stored in one or more memories. In addition, the above-mentioned computer programs, computer instructions, data, etc. can be called by the processor.

[0201] The processor is configured to execute the computer program stored in the memory to implement the various steps of the method or various modules of the system involved in the above embodiments. For details, please refer to the relevant descriptions in the above method and system embodiments.

[0202] The processor and memory can be independent structures or integrated structures. When the processor and memory are independent structures, the memory and processor can be coupled via a bus to transmit data signals. This embodiment of the present invention will not be described in detail.

[0203] The execution subjects of the above-mentioned processor and memory can be computer terminal devices with computing functions such as computers, single-chip microcomputers, and microcontrollers. In specific implementation, the embodiment of the present invention does not limit the execution subject and it is selected according to the needs of actual application.

[0204] The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted via computer-readable storage media. Computer-readable storage media can be any available medium that can be accessed by a computer, or a data storage device such as a server or data center that integrates one or more available media. Available media can include magnetic media or semiconductor media, etc.

[0205] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it can be used to execute any method of the above embodiments of the present invention, or to run any system of the above embodiments of the present invention.

[0206] Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transfer of computer programs from one location to another. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer. An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and storage medium can be located in an ASIC. Alternatively, the ASIC can be located in a user device. Of course, the processor and storage medium can also exist as discrete components in a communication device.

[0207] The computer-readable storage medium includes but is not limited to a flash memory, a hard disk, a solid-state drive, and the like.

[0208] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions according to the embodiments of the present invention are generated in whole or in part.

[0209] The above-mentioned embodiment of the present invention provides an intelligent motion target detection method and system based on weakly supervised dynamic optimization, which respectively extracts video and text features from the input unedited video and the corresponding natural language description, realizes multimodal feature interaction through a cross-modal Transformer codec, and constructs a cross-modal semantic representation; uses learnable object queries to generate multiple positive and negative Gaussian proposals, and introduces a pseudo-label-guided contrast loss function to improve the discrimination ability of positive and negative proposals; further combines the mask conditional Transformer module to perform text reconstruction on the positive and negative proposals, evaluates their semantic alignment effect, and introduces a progressive negative proposal constraint strategy to dynamically evolve the negative proposals from easy to difficult; finally, it integrates multiple supervisory signals to optimize the model's ability to accurately locate semantic segments in video clips. By constructing a learnable negative proposal and a dynamic pseudo-label constraint mechanism, the time domain positioning accuracy under weak supervision conditions is significantly improved.

[0210] Matters not mentioned in the above embodiments of the present invention are well known in the art.

[0211] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various variations or modifications within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A method for intelligent detection of moving targets based on weakly supervised dynamic optimization, characterized in that: include: Extracting video features and text features of the original video and text respectively, and fusing the video features and the text features to obtain cross-modal fusion features; Based on the fused features, generating a frame-level semantic similarity score as a pseudo label; Generate positive and negative proposal masks by using learnable object queries and fused features; Using the pseudo labels to guide feature contrast learning of positive proposals, the positive proposals are made to infinitely fit the text features in the semantic space, and the negative proposals are made to infinitely deviate from the relevant areas of the text features; Using the positive and negative proposal masks, text reconstruction is performed based on the proposal mask conditional Transformer, and semantic consistency training is performed on different proposals to obtain the video time domain positioning result; The video time domain positioning result is dynamically optimized to generate a final positioning result and complete the intelligent detection of moving targets.

2. The method for intelligent detection of moving targets based on weakly supervised dynamic optimization according to claim 1, characterized in that: The extracting of video features and text features from the original video and text respectively, and fusing the video features and text features to obtain cross-modal fusion features, includes: For a given raw video and text, a pre-trained 3D convolutional network is used to extract the temporal features of the video frames, and a pre-trained GloVe model is used to extract the word vector features of the text. The time series features and the word vector features are mapped to a unified embedding space through a fully connected layer, and the context information of the time series features and the word vector features are interactively encoded using the Transformer encoder and decoder to obtain a cross-modal fusion feature H that contains semantic and time series relationships, which is expressed as: H=Dec(V,Enc(Q)) Where Enc(·) represents the encoder; Dec(·) represents the decoder.

3. The method for intelligent detection of moving targets based on weakly supervised dynamic optimization according to claim 1, characterized in that: Generating a frame-level semantic similarity score as a pseudo label based on the fusion feature includes: The fusion feature H is input into the feedforward neural network to obtain the correlation score between each frame of video and text as a pseudo label, which is expressed as: S c =σ(FFN c (H)) Where S c is a pseudo label; σ is a Sigmoid activation function, which indicates the semantic relevance between each frame of video and text; FFN c It is a feed-forward neural network.

4. The method for intelligent detection of moving targets based on weakly supervised dynamic optimization according to claim 1, characterized in that: The method of using learnable object query and fusion features to generate positive and negative proposal masks includes: Introduce K positive query vectors Q p ∈R K×D and mK negative query vectors Q n ∈R K×D , and input it into a proposal generation model PFG (Proposal Feature Generator) together with the fusion feature H to obtain positive and negative proposal features, expressed as: F p =PFG(Q p ,H),F n =PFG(Q n ,H) Where, F p is a positive proposal feature; F n is the negative proposal feature; R represents the real number space; D is the dimension of each query vector; Based on the F p 、F n Predict the center position c of the corresponding time period k and width ω k , construct the Gaussian mask function on the time dimension, which is expressed as follows: Where, f (k) (t) is the value of the k-th time proposal at the t-th time position, which is used to construct the Gaussian mask corresponding to the k-th time proposal; N is the total number of frames in the video; t is the index of the current time frame; The outputs of the Gaussian mask function generated based on the positive and negative proposal features at each time position are stacked row by row to construct the positive proposal mask set M p ∈R K×N and the negative proposal mask set M n ∈R mK×N .

5. The method for intelligent detection of moving targets based on weakly supervised dynamic optimization according to claim 1, characterized in that: The method of using the pseudo labels to guide the feature contrast learning of the positive proposal so that the positive proposal infinitely fits the text feature in the semantic space and the negative proposal infinitely deviates from the relevant area of the text feature includes: Construct pseudo-label guided contrast loss L p As shown below: Where, is the i-th negative proposal mask in the negative proposal mask set; is the i-th positive proposal mask in the positive proposal mask set; K is the number of positive query vectors; S c is a pseudo label; α p is a hyperparameter that controls the contrast interval; T is the matrix transpose symbol; Use the pseudo-label to guide the contrast loss K p , guiding the feature contrast learning of the positive proposal, so that in the semantic space, the positive proposal infinitely fits the text feature, while the negative proposal infinitely deviates from the text feature.

6. The method for intelligent detection of moving targets based on weakly supervised dynamic optimization according to claim 1, characterized in that: The proposed mask-conditional Transformer performs text reconstruction and semantic consistency training on different proposals, including: Use mask-conditional Transformer to reconstruct text from positive proposal masks and weighted negative proposal masks; where: The weighted negative proposal mask is expressed as: Where, is the weighted negative proposal mask corresponding to the k-th positive proposal; is the Gaussian mask of the i-th negative proposal associated with the k-th positive proposal; m is the number of candidate negative proposals corresponding to each positive proposal; ω i is the importance weight of the i-th negative proposal mask, which is obtained as follows: Where FFN w It is a feedforward neural network; is the negative proposal mask of group i Multimodal features CLS characterization; The positive proposal mask, weighted negative proposal mask, and the mask corresponding to the entire original video are input into Transformer along with the masked text, the masked text is predicted, the original text is reconstructed, and the text reconstruction loss L is calculated. rec , perform semantic consistency evaluation on different proposals; wherein, the masked text refers to the masked input formed by randomly masking a set proportion of words from the original text Select the sample with the largest reconstruction error in each group of negative proposals as the training target, and perform semantic consistency training on different proposals; let the current training round be e, and the maximum number of rounds be e max , define the dynamic adjustment weight as: Where β and α are hyperparameters that control the training progress; Final progressive negative proposal cross entropy loss Expressed as: Where, is the cross entropy loss of weighted negative proposals; k * The number of the positive proposal with the smallest current loss.

7. The method for intelligent detection of moving targets based on weakly supervised dynamic optimization according to claim 1, characterized in that: The dynamically optimizing the video time domain positioning result to generate a final positioning result includes: The prediction output of the final time boundary is achieved by using a multi-loss joint optimization mechanism; among them: The multiple losses include: pseudo-label guided contrast loss L p , text reconstruction loss L rec And a video contrast loss L ivc and a proposal diversity loss L div ; The final optimization goal of the joint optimization is: L total =λ p L p +λ rec L rec +λ ivc L ivc +λ div L div Where λ p ,λ rec ,λ ivc and θ div The four coefficients are hyperparameters used to adjust the impact of each loss on model training.

8. An intelligent moving target detection system based on weakly supervised dynamic optimization, characterized in that: include: A feature extraction module is used to extract video features and text features of the original video and text respectively, and fuse the video features and the text features to obtain cross-modal fusion features; a pseudo-label construction module, which generates a frame-level semantic similarity score as a pseudo-label based on the fused features; A proposal generation module that uses learnable object queries and fused features to generate positive and negative proposal masks; A proposal optimization module, which uses the pseudo-labels to guide feature contrast learning of positive proposals, so that the positive proposals infinitely fit the text features in the semantic space and the negative proposals infinitely deviate from the relevant areas of the text features; A text reconstruction module, which uses the positive and negative proposal masks to reconstruct text based on the proposal mask conditional Transformer, and performs semantic consistency training on different proposals to obtain video temporal localization results; The positioning optimization module is used to dynamically optimize the video time domain positioning result, generate the final positioning result, and complete the intelligent detection of moving targets.

9. A computer terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When executing the computer program, the processor can be used to perform the method according to any one of claims 1 to 7, or run the system according to claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it can be used to perform the method according to any one of claims 1 to 7, or to run the system according to claim 8.

Images