Unsupervised video general time sequence boundary detection method and device, and storage medium
Through the unsupervised video universal timing boundary detection method, the convolutional neural network and boundary index calculations are used to solve the commonality problem of lens-level, event-level and action-level boundary detection in the prior art, and efficient and accurate boundary detection in online scenarios are achieved.
Patent Information
- Application Number
- CN202510593720.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-09
- Publication Date
- 2025-08-12
AI Technical Summary
The prior art lacks a general timing boundary detection model at the lens level, event level and action level, and the unsupervised model has poor detection effect in real-time online scenarios.
An unsupervised video universal timing boundary detection method is designed, and video frame features are extracted through convolutional neural networks, boundary index is calculated, and boundary detection is realized at the lens level, event level and action level through convolutional neural networks.
Accurate boundary detection at various semantic levels is realized, suitable for online scenarios, the boundary detection granularity is adjustable, without manual labeling, improving detection effect and efficiency.
Smart Images
Figure CN120472371A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer technology, relates to computer vision and video temporal boundary detection tasks, and is a method, device and storage medium for detecting video temporal boundary that is universal and unsupervised at all semantic level boundaries. Background Art
[0002] Due to the rapid development of computer technology, massive amounts of video data are generated daily on the internet. Intelligent video understanding models have become a hot topic in the field of computer vision. Mainstream video understanding tasks, including action recognition, action detection, video retrieval, and video localization, typically focus on understanding shorter videos and are difficult to apply to longer videos such as surveillance footage, sports videos, and movies. Therefore, understanding long videos requires semantic segmentation into a series of shorter segments. This requires detecting temporal boundaries within the video. Based on the semantic hierarchy, temporal boundaries can be categorized into shot, event, and action levels.
[0003] Existing video temporal boundary detection methods usually only target the problem of video temporal boundary detection at a single semantic level, and lack a universal model for temporal boundaries at the shot level, event level, and action level. On the other hand, current video temporal boundary detection methods are mostly supervised deep learning methods. However, supervised methods rely on manual data labeling, which is costly and prone to deviations due to different cognitions of labelers. Existing unsupervised temporal boundary detection methods can be classified into two types: clustering-based methods and boundary selection methods. Clustering methods are less efficient and cannot be applied to online scenarios; boundary selection methods lack good boundary selection criteria and usually only consider the feature differences across boundaries near the boundary frame. A single judgment criterion cannot take into account the semantic segmentation of different semantic levels at the shot level, event level, and action level, and is prone to false detection, resulting in inaccurate boundary detection results. Summary of the Invention
[0004] The technical problem addressed by this invention is that existing technologies lack a universal temporal boundary detection model for shot-, event-, and action-level temporal boundaries. Furthermore, existing unsupervised model training methods cannot meet the requirements of real-time online scenarios and accuracy. To address this problem, this invention designs a universal unsupervised video temporal boundary detection method for all semantic levels. This method is computationally efficient and can be applied to online temporal boundary detection scenarios.
[0005] The technical solution of the present invention is: an unsupervised universal temporal boundary detection method for video, which outputs the temporal boundaries of an input video having multiple segments at at least one semantic level. First, a convolutional neural network is used to extract features of each frame, and the probability of each frame being a temporal boundary is evaluated by calculating the boundary index of each frame. Then, Gaussian smoothing is performed on the boundary index in the temporal sequence, and the local maximum of the boundary index is taken as the candidate boundary. Finally, according to the desired semantic level segmentation granularity, a quantile threshold is used for screening to obtain the final temporal boundary.
[0006] The boundary index considers both the inter-frame feature differences across boundaries and the inter-frame feature similarities within the same segment. The boundary index is calculated as follows:
[0007] Take k frames before and after the t-th frame with a step size of d as a reference for calculating the boundary index of the t-th frame. The two parameters k and d jointly affect the receptive field of the sample. For the t-kd, t-(k-1)d, ..., td frames before the t-th frame, calculate the sum of the pairwise feature similarities:
[0008]
[0009] S(t-id,t-jd) represents the feature similarity between the t-id frame and the t-jd frame;
[0010] For the t+d, t+2d, ..., t+kd frames after the t-th frame, calculate the sum of the pairwise feature similarities:
[0011]
[0012] And calculate the sum of the pairwise similarities between the sampling frames before and after the t-th frame:
[0013]
[0014] The boundary index B(t) of the tth frame is:
[0015] B(t)=S left (t)+S right (t)-S cross (t)
[0016] The size of the boundary index corresponds to the possibility that the frame is a timing boundary.
[0017] Furthermore, after extracting the features of each frame, the similarities S(t,t+d),...,(t,t+kd) between the frames are first calculated and saved. Then, when calculating the boundary index, the saved similarity values of the corresponding frames are directly referenced.
[0018] Furthermore, the local maximum of the boundary index is specifically: if B(t * )≥B(t), then t * It is a local maximum point, that is, the local maximum is guaranteed to be the maximum value in a local window of length 2L+1 centered on itself, where L is set according to the requirements for segmentation of semantic segments and does not exceed the minimum number of frames of the semantic segment.
[0019] Furthermore, the method of quantile threshold screening is as follows:
[0020] The maximum x-quantile of the boundary index of all frames in the entire video is used as the threshold of the boundary index. Boundary indexes greater than the threshold are detected as temporal boundaries. The segmentation granularity of boundary detection is customized by adjusting x. Reducing x can detect more and finer-grained boundaries.
[0021] Furthermore, for online scenarios, when calculating the boundary index, the parameters k and d are set. The smaller k and d are, the shorter the detection delay is. When using the quantile threshold for screening, the maximum x-quantile of the boundary index of all frames before the current time point is used as the threshold to achieve quasi-real-time temporal boundary detection.
[0022] The semantic levels of the video of the present invention include three levels: shot level, event level and action level.
[0023] The present invention also provides an electronic device, which includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the above-mentioned unsupervised video general temporal boundary detection method, input an uncropped long video, and output the temporal boundaries of various semantic levels in the video, such as the shot level, event level, and action level.
[0024] The present invention also provides a computer-readable storage medium storing a computer-executable program. When the program is executed by a processor, the above-mentioned unsupervised universal time series boundary detection method is implemented.
[0025] This paper designs a boundary index as a selection criterion for temporal boundaries at each semantic level of a video, and designs a complete set of process steps based on the boundary index to achieve unsupervised universal temporal boundary detection in videos. Compared with the existing technology, this paper has the following advantages:
[0026] 1) Applicable to All Semantic Levels: Observations reveal consistent patterns around temporal boundaries at the shot, event, and action levels. This paper summarizes these patterns and proposes a boundary index as a universal criterion for selecting temporal boundaries across all semantic levels. By adjusting the receptive field when calculating the boundary index, temporal boundaries at different semantic levels can be detected.
[0027] 2) Improved boundary detection: The boundary index proposed by the present invention considers both the feature differences across boundary frames near the boundary frame and the feature similarities within the same segment, making it a more comprehensive boundary selection indicator in unsupervised scenarios. The present invention uses Gaussian filtering to smooth the noise on the boundary index curve and preliminarily screens the time series boundaries through local maxima. The maximum x-quantile is used as the threshold for screening the final time series boundary instead of a fixed value to avoid the influence of the absolute value of the boundary index on the detection results. These designs all play a role in improving the boundary detection effect.
[0028] 3) No training required: The present invention does not require supervised training or self-supervised learning of video features, which has advantages in the case of a single video in real scenarios.
[0029] 4) Adjustable boundary detection granularity: By adjusting x in the quantile threshold screening step, the present invention can customize the granularity of boundary detection. Reducing x can detect more and finer-grained boundaries.
[0030] 5) Easy to extend to online scenarios: The boundary index proposed in the present invention only cares about the local information of the video. It only needs to narrow the receptive field when calculating the boundary index and use the maximum x-quantile of the boundary index of all frames before the time point as the threshold when calculating the boundary index threshold. This method can be applied to online scenarios to achieve quasi-real-time temporal boundary detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 It is the overall framework flow chart of the method of the present invention.
[0032] Figure 2 Schematic diagram of the similarity characteristics between frames near the temporal boundary in the present invention. Based on this observation, the present invention proposes a boundary index as a criterion for boundary selection.
[0033] Figure 3 It is a visualization diagram of the reasoning process of a specific embodiment of the present invention.
[0034] Figure 4 This is the result of comparing the present invention with the prior art on the ClipShots dataset, where the ones with asterisks indicate methods that do not require any training, and the unsupervised methods without asterisks are methods that require self-supervised training.
[0035] Figure 5 This is the result of comparing the present invention with the prior art on the Kinetics-GEBD dataset, where the ones with asterisks indicate methods that do not require any training, and the unsupervised methods without asterisks require self-supervised training.
[0036] Figure 6This is the result of comparing the present invention with the prior art on the TAPOS dataset, where the ones with an asterisk indicate methods that do not require any training, and the unsupervised methods without an asterisk are methods that require self-supervised training. DETAILED DESCRIPTION
[0037] The present invention proposes an unsupervised universal temporal boundary detection method for videos, which inputs an uncropped long video and outputs the temporal boundaries of various semantic levels in the video, such as the shot level, event level, and action level. The long video here refers to a video with multiple segments at at least one semantic level. The present invention detects temporal boundaries to divide the semantic segments. The present invention is suitable for temporal boundary detection at various semantic levels, has good boundary detection effect, does not require any training, has adjustable boundary detection granularity, and can be easily extended to online scenarios. The present invention can serve as a basic preprocessing link for long video understanding and is widely used in long video understanding scenarios.
[0038] like Figure 1 As shown, the present invention first uses a convolutional neural network to extract the features of the video frames, then calculates the boundary index of each frame based on the feature similarity, and smoothes the boundary index in the temporal dimension. Then, the local maximum of the boundary index is taken as the candidate boundary, and finally the candidate boundary is further screened using the quantile threshold to obtain the final temporal boundary.
[0039] The specific implementation process of the present invention is as follows:
[0040] 1) Feature extraction: Given a video of length N frames, the frame-by-frame features extracted using a convolutional neural network are X = {x1, x2, ..., x N}.
[0041] 2) Boundary index calculation: Use cosine similarity to calculate the feature similarity between a pair of frames, and use S(i,j) to represent the feature similarity between the i-th frame and the j-th frame. Figure 2 The figure shows the boundary between the run-up and take-off phases of a long jump competition video. It can be observed that near the boundary frame, pairs of frames crossing the boundary exhibit feature differences, while on the other side of the boundary frame, frames within the same video segment exhibit feature similarities. Based on this observation, the present invention proposes a boundary index based on feature similarity as an indicator for determining whether a frame is a boundary frame.
[0042] In order to calculate the boundary index of the t-th frame to judge its probability of being a boundary, focus on some frames around it. Assume that the sampling step size is d and the sampling half-window size is k. Take k frames with a step size of d before and after the t-th frame as a reference for calculating the boundary index of the t-th frame. In other words, focus on the t-kd, t-(k-1)d, ..., td frames on the left of the t-th frame, and focus on the t+d, t+2d, ..., t+kd frames on the right. Here, the two parameters k and d jointly affect the receptive field of the sampling. For the t-kd, t-(k-1)d, ..., td frames on the left of the t-th frame, calculate the sum of the pairwise feature similarities:
[0043]
[0044] Similarly, for the t+d, t+2d, ..., t+kd frames to the right of the t frame, the sum of the pairwise feature similarities is also calculated:
[0045]
[0046] For the case of spanning the t-th frame, that is, between the t-kd, t-(k-1)d, ..., td frames and the t+d, t+2d, ..., t+kd frames, calculate the sum of the pairwise similarities:
[0047]
[0048] The calculation formula for defining the boundary index B(t) of the t-th frame is as follows:
[0049] B(t)=S left (t)+S right (t)-S cross (t)
[0050] When the boundary index is larger, it means that the possibility of the frame being a timing boundary is greater, otherwise the possibility of the frame being a timing boundary is smaller.
[0051] For some frames at the beginning and end of the video, it is impossible to obtain a sufficient number of frames before and after them as described above. For this special case, there are two specific implementation methods. One is to pad the beginning and end of the video with zeros. In this case, the boundary index will identify the first and last frames of the video as timing boundaries, which is reasonable. The other method is to directly determine that some frames at the beginning and end of the video are not timing boundaries. This is also reasonable because these frames do not have sufficient temporal context information to prove that they are timing boundaries. As long as the appropriate receptive field is adjusted through the two parameters k and d, this processing method will not miss the timing boundaries. Therefore, both processing methods are reasonable. The only difference lies in the judgment of whether the first and last frames of the video are timing boundaries. This judgment is not important because no matter how the first and last frames of the video are judged, the video will eventually be divided into the same segments by the timing boundaries.
[0052] 3) Gaussian smoothing: In order to reduce the noise on the boundary index curve and avoid interference with boundary detection, after calculating the boundary index of all frames of a video, a Gaussian filter is used to smooth the boundary index curve along the time dimension.
[0053] 4) Preliminary screening of time series boundaries through local maximum: The definition of local maximum is as follows: If B(t * )≥B(t), then the tth * The boundary index of the frame is a local maximum point. That is, the local maximum here is guaranteed to be at the tth * The maximum value in a local window of length 2L+1 centered on the frame itself. L is set according to the requirements for semantic segmentation and does not exceed the minimum number of frames of the semantic segment. For example, if the user hopes that the smallest segment to be segmented should not be less than a certain length, then set it to the number of frames of this length; if the user does not have this requirement, set it to a smaller value, such as the number of frames corresponding to 0.5 seconds. Most videos usually do not have multiple semantic switches in such a short time. If the video is a special scene with multiple switches in a short time, it can be set to a smaller value. If L is greater than the number of frames of a certain semantic segment, the boundary with the smaller boundary index in the front and back boundaries of this segment will be missed. At this time, the user does not have high requirements for the fineness of temporal segmentation, and such missed detection will not affect the detection results to meet user needs. Therefore, L can be set according to the requirements for semantic segmentation.
[0054] For the smoothed boundary index, the local maximum point is taken as the time series boundary for preliminary screening.
[0055] 5) Obtain the final temporal boundary through quantile threshold screening: The local maximum screened out in the previous step is not necessarily the real temporal boundary. There may be multiple local maxima in a real video segment. This is because there are also slight changes in semantic information within a video segment. For this reason, it is necessary to set a boundary index threshold to eliminate the influence of these pseudo-boundaries. Due to the diversity of video content, the absolute value of the boundary index will fluctuate. In addition, the setting of k when calculating the boundary index will also have a greater impact on the absolute value of the boundary index. Therefore, it is inappropriate to use a fixed threshold. The present invention uses the maximum x-quantile of the boundary index of all frames of the entire video as the threshold of the boundary index. For example, when x=5, the largest 1 / 5 of the boundary index of all frames of the video is selected as the threshold. By adjusting x, the granularity of boundary detection can be customized, and reducing x can detect more and finer-grained boundaries.
[0056] The following further proposes an efficient algorithm for calculating the boundary index of all frames in a single video. Through the implementation of the present method, it is noted that during the boundary index calculation process, the vast majority of the time is spent calculating frame feature similarity. This is because the computation time of feature similarity is linearly related to the feature dimension, which is typically large. For example, the features extracted by the commonly used ResNet-50 backbone network reach 2048 dimensions. Therefore, by reducing the number of feature similarity calculations, the computation time overhead of the boundary index can be reduced. A naive calculation method is to calculate the boundary index of a frame by performing the inter-frame similarity calculation when the similarity of a pair of frames is used. This involves redundant calculations. For example, when k ≥ 3, when calculating the boundary index B(t0) of frame t0, the feature similarities S(t0-d, t0-2d) of frames t0-d and t0-2d are calculated. When calculating the boundary index B(t0+d) of frame t0+d, S(t0-d, t0-2d) are still calculated. This is an example of redundant calculation. Therefore, we consider precalculating and storing the feature similarities between all possible frame pairs, and then using them directly when calculating the boundary index. We observe that for frame t in a video, only its feature similarities with frames t-kd, ..., t-2d, td, t+d, t+2d, ..., t+kd are used when calculating the boundary index for all frames. Therefore, for each frame t in the video, we only need to calculate and store S(t, t-kd), ..., S(t, td), S(t, t+d), ..., S(t, t+kd). Since S(i, j) = S(j, i), we only need to calculate and store S(t, t+d), ..., S(t, t+kd) for each frame. By processing the video in a time series, we can obtain the similarity of each frame. After extracting all frame features and before calculating the boundary indices for all frames, the present invention calculates and saves the similarities for all such frame pairs. Later, when calculating the boundary indices, the feature similarity between two frames can be directly referenced and used, avoiding the repeated calculation of similarities for many frame pairs. Mathematical derivation shows that the efficient algorithm requires almost 1 / 2k times as many feature similarities as the naive algorithm when calculating the boundary indices for all frames in a single video, significantly improving efficiency.
[0057] Furthermore, the present invention can be easily extended to online scenarios. As another example, if the present invention is intended for online use, simply set the parameters k and d to smaller values when calculating the boundary index. The values of k and d depend on the frame delay that can be tolerated in the online scenario. Since the boundary index calculation of a video frame uses the features of the following k*d frames, determining whether a frame is a boundary in an online scenario will involve a delay of at least k*d frames. Setting k and d smaller is intended to reduce this delay in online scenarios. Finally, when calculating the boundary index threshold, the maximum x-quantile of the boundary indexes of all frames before that time point is used as the threshold. In the aforementioned offline embodiment, the boundary indexes of all frames in the entire video are used to calculate the quantile threshold, and then the temporal boundaries of the entire video are output. In the online scenario, the video is detected in real time and the detection results of the temporal boundaries are output. In this case, the boundary index of the entire video cannot be used. Therefore, the quantile threshold is calculated based on the boundary indexes of each frame currently calculated to filter the temporal boundaries in the currently input video. In this way, the present invention can be applied to online scenarios, achieving near-real-time temporal boundary detection.
[0058] Figure 3 The reasoning process of the present invention on video is demonstrated through a specific embodiment. It can be seen that after the video is input, the model first calculates the boundary index of each frame, and then performs Gaussian smoothing on the boundary index. Among the smoothed boundary indexes, the local maximum value of the boundary index is taken as the candidate boundary, such as Figure 3 As shown in the red ellipse circled position in . Note that the local maximum in the present invention requires that there is no larger value within a certain length interval, which is different from the definition of local maximum in mathematics. Therefore, the green dotted ellipse circle is not a local maximum, because there is a point with a larger boundary index in a small range behind it. Then, the maximum 5-percentile threshold is used to screen and obtain the two boundaries that finally meet the conditions. It can be seen that in this embodiment, the model correctly detects the boundary and divides a long jump competition video into three segments of "run-up-take-off-landing" according to semantics.
[0059] The present invention is further implemented on the ClipShots dataset, Kinetics-GEBD dataset, and TAPOS dataset, and compared with other temporal boundary detection methods. There are a large number of shot-level temporal boundaries in the ClipShots dataset, including sudden shot switching and gradual shot switching; there are a large number of event-level temporal boundaries in the Kinetics-GEBD dataset; the TAPOS dataset crops action instances into input videos during use and detects the boundaries of sub-actions in the action instances, so it can be regarded as an action-level temporal boundary detection problem. When calculating the boundary index in the implementation, for sudden shot switching in the ClipShots dataset, the sampling step size is set to d=1 and the sampling half-window size is k=3; for gradual shot switching in the ClipShots dataset, the Kinetics-GEBD dataset, and the TAPOS dataset, the sampling step size is set to d=5 and the sampling half-window size is k=5. In the implementation, the ResNet-50 convolutional neural network is uniformly used to extract video frame features, the local maximum window size is uniformly set to L=15, and the boundary index threshold quantile x=5. The boundaries in the ClipShots dataset are annotated in the form of time intervals. The F1-score is used for evaluation on this dataset. If the predicted boundary overlaps with the true boundary in at least one frame, the prediction is considered correct. On the Kinetics-GEBD dataset and the TAPOS dataset, the F1-score under various relative distance thresholds is used to evaluate the performance of the model. The relative distance threshold represents the threshold of the relative distance between the predicted boundary and the true boundary. For example, when the relative distance is 0.05, it means that if the predicted boundary falls within the interval of [-0.05*video length, 0.05*video length] of the true boundary, the prediction is correct, otherwise the prediction is incorrect. Figure 4-6 This report compares the proposed method with other methods on three datasets. Methods marked with asterisks require no training, while unsupervised methods without asterisks require self-supervised training. The experimental results demonstrate that the proposed boundary index model outperforms other untrained methods on shot-level, event-level, and action-level boundary detection tasks, demonstrating that the proposed method can effectively detect temporal boundaries at various semantic levels.
[0060] The present invention can be implemented based on a computer program. Based on this, the present invention also provides an electronic device, the electronic device comprising a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or at least one program being loaded and executed by the processor to implement the above-mentioned unsupervised video universal temporal boundary detection method. A computer-readable storage medium is also provided, the computer-readable storage medium storing at least one instruction or at least one program, the at least one instruction or program being executed to implement the above-mentioned unsupervised video universal temporal boundary detection method. The instructions or program code for implementing the method of the present invention can be written in any combination of one or more programming languages. The instructions or program code can be executed entirely on a data processor, partially on a processor, as an independent software package, partially on a processor and partially on a remote device, or entirely on a remote device or server.
Claims
1. Unsupervised video general temporal boundary detection method, which is characterized by For an input video with multiple segments at at least one semantic level, the temporal boundaries of the video are output. First, a convolutional neural network is used to extract features from each frame. The probability of each frame being a temporal boundary is evaluated by calculating the boundary index of each frame. Then, Gaussian smoothing is performed on the boundary index in the temporal sequence, and the local maximum of the boundary index is taken as the candidate boundary. Finally, according to the desired semantic level segmentation granularity, a quantile threshold is used for screening to obtain the final temporal boundary. The boundary index considers both the inter-frame feature differences across boundaries and the inter-frame feature similarities within the same segment. The boundary index is calculated as follows: Take k frames before and after the t-th frame with a step size of d as a reference for calculating the boundary index of the t-th frame. The two parameters k and d jointly affect the receptive field of the sample. For the t-kd, t-(k-1)d, ..., td frames before the t-th frame, calculate the sum of the pairwise feature similarities: S(t-id,t-jd) represents the feature similarity between the t-id frame and the t-jd frame; For the t+d, t+2d, ..., t+kd frames after the t-th frame, calculate the sum of the pairwise feature similarities: And calculate the sum of the pairwise similarities between the sampling frames before and after the t-th frame: The boundary index B(t) of the tth frame is: B(t)=S left (t)+S right (t)-S cross (t) The size of the boundary index corresponds to the possibility that the frame is a timing boundary.
2. The unsupervised video universal temporal boundary detection method according to claim 1, characterized in that For the input video, after extracting the features of each frame, we first calculate and save the similarity S(t,t+d),...,(t,t+kd) between the frames. Then, when calculating the boundary index, we directly refer to the saved similarity values of the corresponding frames.
3. The unsupervised video universal temporal boundary detection method according to claim 1, wherein the boundary The local maximum of the index is specifically: if Then t * It is a local maximum point, that is, the local maximum is the maximum value in a local window of length 2L+1 centered on itself, where L is set according to the requirements for segmentation of semantic segments and does not exceed the minimum number of frames of the semantic segment.
4. The unsupervised video universal temporal boundary detection method according to claim 1, characterized in that The specific method of quantile threshold screening is: The maximum x-quantile of the boundary index of all frames in the entire video is used as the threshold of the boundary index. Boundary indexes greater than the threshold are detected as temporal boundaries. The segmentation granularity of boundary detection is customized by adjusting x. Reducing x can detect more and finer-grained boundaries.
5. The unsupervised video universal temporal boundary detection method according to claim 1, characterized in that For online scenarios, when calculating the boundary index, set the parameters k and d. The smaller k and d are, the shorter the detection delay is. When using the quantile threshold for screening, use the maximum x-quantile of the boundary index of all frames before the current time point as the threshold to achieve quasi-real-time temporal boundary detection.
6. The unsupervised video universal temporal boundary detection method according to claim 1, characterized in that The semantic levels include shot level, event level and action level.
7. An electronic device, characterized in that The device includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the unsupervised general video temporal boundary detection method described in any one of claims 1 to 6, inputting an uncropped long video and outputting temporal boundaries of various semantic levels at the shot level, event level, and action level in the video.
8. A computer-readable storage medium, characterized in that A computer executable program is stored, and when the program is executed by a processor, the unsupervised video universal temporal boundary detection method according to any one of claims 1 to 6 is implemented.