A method for intelligent serial and parallel video by target label

Abstract

The application relates to a method for intelligently stringing and parallelizing videos through target labels, which comprises the following steps: extracting a label string keyword and a structured feature attribute of a target in a main video; performing algorithm matching weight on the label string keyword and the structured feature attribute; and performing batch matching of the structured feature attribute and the label string keyword similarity in a target information table. As long as the two dimensions of the structured feature attribute similarity and the label string keyword matching degree of any target except a current target reach a certain percentage or more, the target is added to a similar target pool and waits for further inspection. The method can solve the problem that, in video work, because target features cannot be effectively utilized for intelligent comparison and stringing and parallelizing, omission occurs, thereby leading to incomplete stringing and parallelizing videos.

Description

Technical Field

[0001] This invention relates to video processing, specifically to a method for intelligent serialization and parallelization of videos using target tags. Background Technology

[0002] Currently, most video concatenation / parallelization processes involve identifying and processing a target image, extracting fixed features, and then performing a brute-force comparison. If the feature similarity percentage meets the criteria, the video is considered either concatenated or merged. This concatenation / parallel comparison method is one of the more mainstream methods, offering advantages such as convenience, speed, and efficiency, saving significant time spent reviewing videos and improving video processing efficiency. However, using target images for recognition also has some drawbacks, such as: 1. Human faces exhibit similarity; the differences between individuals are not significant, and most facial structures are quite similar. The use of makeup and the natural similarity between twins further increases the difficulty of recognition. 2. Human faces are variable; their shape is unstable, and facial expressions can vary significantly across different target images. 3. Facial recognition is also affected by factors such as lighting conditions, facial occlusions, and age, leading to differences in the final recognized features. 4. Facial recognition feature extraction is a program-defined algorithm, which cannot be effectively interfered with by humans. If more detailed features are required, alternative methods need to be devised. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for intelligent video concatenation using target tags. This method can solve the problem of incomplete video concatenation caused by missing clues due to the inability to effectively utilize target features for intelligent comparison and concatenation.

[0004] The technical solution adopted to achieve the purpose of this invention is a method for intelligent concatenation and parallelization of videos by combining target tags with structured feature attributes. The method includes the following steps:

[0005] S1: Extract the tag string keywords and structured feature attributes of the target in the main video;

[0006] S2: The algorithm matches the tag string keywords and structured feature attributes with weights. The system weights and adjusts the current structured feature attributes and tag string keywords.

[0007] S3: Perform batch matching of structured feature attributes and tag string keyword similarity in the target information table. As long as the similarity of structured feature attributes and the matching degree of tag string keywords of any other target besides the current target reach a certain percentage, then add the target to the similar target pool for further verification.

[0008] S4: For all targets in the similar target pool, match the video to which each target belongs in the relevant view library to detect all videos of the current target;

[0009] S5: Set all detected videos as videos to be strung together, and then sort the list of videos to be strung together in ascending order based on the structural feature similarity of the target, the keyword matching degree of the tag string, and the similarity of the video properties to which the target belongs.

[0010] In the above technical solution, the tag string keywords are features that the system cannot extract, and features that distinguish them from other targets are obtained manually.

[0011] In the above technical solution, in step S3, if the similarity of the structured feature attributes and the keyword matching degree of the tag string of any other target besides the current target both reach 30% or more, then the target is added to the similar target pool.

[0012] In the above technical solution, in step S4, a batch database search is performed on the relevant view library by matching all similar targets, and the video corresponding to each similar target is matched.

[0013] In the above technical solution, in step S5, the video category, video time, video location, attributes of the opposing target and attributes of the main video of each video are compared, filtered, sorted and deduplicated. The videos are then sorted in a positive order from large to small according to the similarity of the target structured attribute features, the matching degree of the same target label or similar target label, and the similarity of video attributes.

[0014] The difference between this invention and existing video conferencing methods lies in its use of a dual matching mechanism combining structured attributes and tag string keywords during video conferencing. Leveraging existing big data analysis experience, the program internally deduces and calculates the weight ratios of structured attributes and tags, forming a unique artificial intelligence algorithm that is more intelligent in its presentation and more efficient in video conferencing. Existing conferencing methods have regional limitations. For example, if a video of a certain nature occurs in one location at a certain time, and a similar video occurs in another region some time later, the two regions cannot effectively coordinate due to the cross-regional nature of the video. However, this invention, once deployed within the network, effectively confers all videos recorded into the system, solving the problem of cross-regional video conferencing.

[0015] The present invention has the following advantages:

[0016] 1. Multi-dimensional integrated video concatenation, resulting in more accurate and complete video concatenation.

[0017] 2. Intelligent serialization and parallelization in big data analysis greatly reduces the workload of frontline staff.

[0018] 3. Comparison across multiple dimensions reduces the time users spend repeatedly watching videos, improving efficiency.

[0019] 4. Accurate and efficient.

[0020] 5. Effectively combines target recognition technology, intelligent video analysis technology, and the practical experience and knowledge of viewers to provide more accurate parallel video information. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a method for intelligent serialization and parallelization of videos using target tags according to the present invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown, the method of intelligent video concatenation using target tags in this invention establishes a relationship between the video information table and the target information table by extracting tag string keywords from the target. Specifically, it includes the following steps:

[0024] S1: The program extracts the structured feature attributes and tag string keywords of the targets in the main video. The structured feature attributes come from the attributes of the current target after the recognition program. The tag string keywords are special attributes manually assigned to each target when it is entered into the target information table, in addition to the facial features extracted by the system for the current target. These attributes are generally relatively independent and obscure, with a low repetition rate between different targets.

[0025] S2: Based on extensive and effective big data analysis experience, the system uses internal artificial intelligence algorithms to internally weight and balance structured attributes and tag string keywords.

[0026] S3: Using weighted structured attributes and target tag string keywords, perform batch fuzzy matching of tag string keywords in the target information table. If any target tag string keyword other than the current target partially or completely matches, and the similarity of the structured feature attributes reaches 30% or more, the union of the two comparison dimensions is taken. Then, the target is added to the similar target pool for further verification. In this process, the structured feature attributes use intelligent algorithm similarity query and tag keyword matching (there are three methods: exact matching, fuzzy matching, and word segmentation matching). Therefore, the resulting set is a forward sorting of the weighted structured feature attributes and tag string keywords from high to low matching degree.

[0027] S4: Using all targets in the similar target pool, match the video to which each target belongs in the relevant view library. In this process, a batch brute-force search is performed on the relevant view library using all the matched similar targets to match the video corresponding to each similar target. Theoretically, these videos and the main video will have some similarity and coincidence.

[0028] S5: Set all detected videos as the videos to be strung together. Then, filter, sort, and deduplicate each video based on its category, time, location, opposing target, and other attributes, as well as the attributes of the main video. The similarity of the target's structured features and the keyword matching of its tag strings, along with the similarity of the properties of the videos to which the target belongs, are then used to forward sort the list of videos to be strung together in descending order.

[0029] S6: The system will repeat steps S2-S5 three times by default to ensure that there are no missed or false detections, thus preventing incomplete or missing results.

[0030] This invention proposes a method for manually adding target features during the target identification process. In this embodiment, the tag string keywords are the manually added features. These manually added features are those that the identification program cannot recognize. Features that the program can recognize include conventional attributes such as the target's clothing type, clothing color, bottom type, and bottom color. If there are more subtle or detailed features, such as: whether the target wears a necklace or pendant, earrings, one or more missing fingers on the target's left or right hand, a disability in the target's left or right leg, abnormal gait, tattoos on the target's arms or forearms, the tattoo design, knife wounds, or other defects, these are all human characteristics. Because some features are subtle and deeply hidden, the identification system often ignores them or cannot accurately identify them. These features can only be supplemented by observation of on-site surveillance video or descriptions from eyewitnesses. For example, facial features identified by a recognition program can lead to a large number of results when searching a view library. Suppose we also have a known, highly independent feature, such as the left eye of the target we need to search for being a fake eye. Using the keyword string "left eye is fake eye", we can first perform a filtering process during the search, eliminating a large number of results that do not meet the conditions. The keyword string "left eye is fake eye" can then be used as one of the conditions for associating the current target with other targets.

[0031] Based on this, this invention proposes a method for quickly and accurately analyzing all other videos that have similar or nearly identical features to the target person in the main video. The method involves intelligent tag-based video comparison. When a target is added to the database, its facial features are first extracted using a recognition program as its unique identifier. Then, the target's information is supplemented by reviewing on-site surveillance videos or obtaining descriptions from eyewitnesses. This information includes both general and specific details. General details include clothing color, type, and color and type of bottoms. Specific details include gender, age range, estimated height range, baldness, facial feature data, and whether the hair is long or short. Furthermore, this invention introduces a tag attribute as specific information. Tag attributes are typically detailed; for example, if the left index finger is missing, the tag "missing left index finger" is added to the database. This tag serves as a relatively independent and significant feature, marking the current target as a target tag string keyword. This tag string keyword then serves as one of the conditions for associating different or similar targets in different videos. When personnel perform operations, facial features are obtained through recognition programs. Because most faces are similar, the final result differs from the actual needs. After feature comparison, a large number of results are retrieved, but these results are only based on the similarity of facial features. If the similarity exceeds a certain value, the system will consider them to be the same target, resulting in redundant and inaccurate final serialization results. This invention introduces target tag string keywords, which are then used as a condition for retrieval during target comparison. For example, consider a scenario where personnel discover the time and image of a suspected target in a video extracted from a live feed. However, due to the small coverage area of ​​the on-site monitoring equipment or the equipment being old, the video image is blurry. In such cases, using target profiles for feature extraction may result in inaccurate or even unidentifiable identification. When features are incomplete or absent, using target feature tag keywords can filter out most targets that do not match the current suspected target. The filtered targets will then be roughly similar to the suspected targets in the on-site monitoring video. Based on this, comparing the target information database will yield a more accurate target set, resulting in a final video result set that is closer to the viewer's desired outcome.

[0032] This invention differs from the prior art in the following three aspects:

[0033] 1. A supplement to conventional video string concatenation, combining tag string keywords with structured video string concatenation.

[0034] Conventional video concatenation typically uses a single attribute of the video as a starting point for data matching. However, because there isn't a one-to-one correspondence between video information such as category, method, location, and object and the video itself, the accuracy of concatenating videos is low, wasting significant manpower and resources with unsatisfactory results. This invention first performs batch structuring of the surveillance videos captured on-site. Each target appearing in the video is identified and analyzed. The structured data serves as the unalterable base data, referred to as the source data. The tag string keywords supplement the structured data, serving several purposes: 1. Highlighting a highly significant feature of the current target; 2. Describing detailed, subtle features; 3. Supplementing certain shortcomings of the structuring process. During video concatenation, intelligent comparison is performed using the source data and tag string keywords, reducing the rate of matching unrelated videos, minimizing video omissions, and improving video recognition efficiency.

[0035] 2. Tag string keyword normalization

[0036] The source of tag string keywords and the determination of users are considered. Due to individual differences among users, different methods may be used when describing tag string keywords. For example, to describe a target person as left-handed, some tag string keywords are 'left-handed', while others are 'dominant left-handed'. Therefore, when users enter tag string keywords, they only need to enter the string 'left'. The system will then match and search for left-handed, dominant left-handed, and all other tags containing the string 'left'. This avoids tag mismatches due to differences in description, aiming to improve system convenience, reduce differences in description, and improve the accuracy of judgment.

[0037] 3. Assess weights

[0038] The system internally assigns weights to structured attributes and tag keywords. Through analysis and training with a large amount of existing data, the machine model automatically learns to assign higher weights to structured feature attributes and tag keywords with higher occurrence rates in comparison dimensions, and lower weights to attributes with lower occurrence rates. During intelligent video concatenation and comparison, videos are analyzed and sorted according to the degree of matching and weight, optimizing the display effect and bringing it closer to the results desired by users. The goal is to filter out unrelated video sets and improve the accuracy of video concatenation.

Claims

1. A method for intelligent serial and parallel video by target tagging, characterized in that, Comprise: S1: extract the target's label string keyword and structured feature attribute in the main video; S2: algorithm matching weight for label string keyword and structured feature attribute, the system weights the current structured feature attribute and label string keyword, and the weight; S3: batch matching structured feature attribute and label string keyword similarity in target information table, as long as the structured feature attribute similarity and label string keyword matching degree of any target other than the current target reach a certain percentage, the target is added to the similar target pool, waiting for further inspection; S4: for all targets in the similar target pool, match each target's video in the view library, and detect all videos of the current target; S5: set all detected videos as to-be-serialized videos, then according to the target's structured feature similarity and label string keyword matching degree, and the target's video property similarity, sort the to-be-serialized video list in ascending order.

2. The method for intelligently serializing and parallelizing videos through target labels according to claim 1, wherein the label string keyword is a feature that cannot be extracted by the system, and is obtained by distinguishing from other targets.

3. The method for intelligently serializing and parallelizing videos through target labels according to claim 1, wherein in step S3, if the structured feature attribute similarity and label string keyword matching degree of any target other than the current target reach 30% or more, the target is added to the similar target pool.

4. The method for intelligently serializing and parallelizing videos through target labels according to claim 1, wherein in step S4, all similar targets are matched to perform batch database searching on the view library, and the video corresponding to each similar target is matched.

5. The method for intelligently serializing and parallelizing videos through target labels according to claim 1, wherein in step S5, the category, video time, video location, attribute of opposite target and attribute of main video of each video are compared, filtered, sorted and de-duplicated, and then sorted in ascending order according to the target's structured attribute feature similarity, matching degree of the same target label or similar target label, and video attribute similarity.

Images