Video security and protection monitoring method and system based on artificial intelligence
By constructing an AI-based video security monitoring system, and utilizing skeleton and gait features to build a three-dimensional spatiotemporal identity fingerprint, the system solves the problem of target loss caused by facial occlusion and clothing changes, and achieves efficient continuous tracking and evidence chain generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LONGYAN TIANBO INFORMATION TECH CO LTD
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-19
AI Technical Summary
Traditional video surveillance systems cannot effectively track suspects when they wear masks, sunglasses, or hoods to obscure their facial features, and it is difficult to form a complete chain of evidence after suspects change their clothes, making it difficult to trace the source of cases and intercept them in real time.
Using an artificial intelligence-based approach, the system detects facial occlusion in real time, extracts steady-state physical features such as skeleton and gait features, constructs a three-dimensional spatiotemporally consistent identity fingerprint model, and calculates feature offset vectors by combining accessibility constraints between monitoring nodes, generating spatiotemporal trajectory records and executing audio-visual reminder actions.
It enables continuous tracking and early warning even when the face is covered or clothing is changed, reducing the cost of manual backtracking and improving the accuracy of target identification and the structure of the evidence chain.
Smart Images
Figure CN122067152A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of security monitoring technology, and in particular to a video security monitoring method and system based on artificial intelligence. Background Technology
[0002] With the development of artificial intelligence technology, video surveillance has been widely used in urban security and business management. However, traditional systems heavily rely on facial recognition. When suspects deliberately obscure their faces by wearing masks, sunglasses, or hoods, the system often fails to trigger effective tracking, leading to target loss. Furthermore, existing cross-camera tracking technologies are mostly based on transient visual features such as clothing color and style. In real-world scenarios, suspects often alter their visual characteristics by changing clothing or discarding items in blind spots. Traditional algorithms, due to significant feature shifts, cannot correlate the disguised target with the original target, making it difficult to form a complete chain of evidence and posing significant challenges to case tracing and real-time interception. Summary of the Invention
[0003] To overcome the above shortcomings, this invention provides a video security monitoring method and system based on artificial intelligence, which aims to solve the problem of continuous tracking and early warning of non-cooperative targets when their faces are obscured or they change clothes midway.
[0004] In a first aspect, the present invention provides the following technical solution: a video security monitoring method based on artificial intelligence, comprising:
[0005] S1. Real-time detection of the facial features of targets within the monitoring area; when the facial features of the target are detected to be occluded, the target is marked as a target to be tracked.
[0006] S2. Extract the steady-state physical features and transient apparent features of the target to be tracked, and construct a consistent identity fingerprint model of the target to be tracked in three-dimensional spatiotemporal based on the steady-state physical features.
[0007] S3. Obtain the target to be matched from adjacent monitoring nodes, and in combination with the reachability constraints between monitoring nodes, filter out the target to be matched whose matching degree with the consistent identity fingerprint model is greater than a preset threshold.
[0008] S4. Extract the transient appearance features of the target to be matched, compare them with the transient appearance features of the target to be tracked, and calculate the feature offset vector.
[0009] S5. If the feature offset vector exceeds the preset mutation threshold, the target to be tracked is determined to be an evasion suspect target, and the video frame sequence of the evasion suspect target before and after the feature mutation is associated to generate a spatiotemporal trajectory record.
[0010] S6. Execute an audio-visual alert on the security monitoring equipment in the area where the target to be tracked and the suspected target to be avoided are located.
[0011] Preferably, in step S1, the step of real-time detection of the facial features of the target within the monitoring area includes:
[0012] The facial feature point set of the target is located by the facial key point detection model, and the proportion of the visible points of the facial feature point set in the current image frame is calculated.
[0013] The target head region is scanned using an object detection model to identify whether there are any pre-set obstructions.
[0014] If the proportion of visible points is lower than a preset visibility threshold, or if there are obstructions in the head area, the facial representation is determined to be in an obstructed state.
[0015] Preferably, in step S2, the steady-state physical features include skeletal features and gait features; the transient appearance features include clothing features and carried items features.
[0016] Preferably, in step S2, the step of constructing a consistent identity fingerprint model of the target to be tracked in three-dimensional spatiotemporal space includes:
[0017] Based on skeletal features, the relative geometric proportions between various joints of the human body are calculated to generate static structural vectors.
[0018] Based on gait features, periodic motion parameters of key point displacements of the target within a preset time window are extracted to generate dynamic behavior vectors.
[0019] The static structure vector and the dynamic behavior vector are concatenated for features and mapped to a unified metric space to form a consistent identity fingerprint model of the target to be tracked.
[0020] Preferably, in step S3, the step of filtering out the targets to be matched whose matching degree with the consistent identity fingerprint model is greater than a preset threshold includes:
[0021] Based on the physical distance between adjacent monitoring nodes and the preset target movement speed range, determine the expected time window for the target to reach the adjacent monitoring nodes;
[0022] Extract the initial features of all targets to be compared that appear in adjacent monitoring nodes within the expected time window, and remove targets that do not conform to spatiotemporal logic;
[0023] The consistency fingerprint model of the target to be tracked is compared with the identity fingerprint of the initially selected target to be compared by calculating the similarity distance.
[0024] If the similarity distance is less than a preset measurement threshold, the corresponding target to be compared is determined to be a successful match and is used as the target to be matched.
[0025] Preferably, in step S4, the step of calculating the feature offset vector includes:
[0026] The transient appearance features of the target to be tracked and the target to be matched are divided into multiple local feature blocks, which include the head region, torso region, leg region and carrying object region;
[0027] Calculate the distances between the target to be tracked and the target to be matched in each corresponding local feature block;
[0028] The distance calculation results of each local feature block are weighted and combined to generate the feature offset vector that characterizes the degree of global appearance difference.
[0029] Preferably, in step S5, the step of generating the spatiotemporal trajectory record includes:
[0030] Automatically extract the first keyframe image of the target to be tracked before the feature mutation, and the second keyframe image of the suspected target to be avoided after the feature mutation;
[0031] In the first keyframe image and the second keyframe image, heatmaps are used to mark the regions of abrupt changes in apparent properties that cause the feature offset vector to exceed the threshold.
[0032] By associating the monitoring nodes corresponding to the first keyframe image and the second keyframe image, and integrating the target's movement path and corresponding timestamp information between different monitoring nodes, a structured spatiotemporal trajectory record is generated.
[0033] Preferably, in step S6, the step of performing the sound and light reminder action includes:
[0034] When the target to be tracked is identified, it is highlighted on the monitoring interface using a visual identifier.
[0035] When a suspected evasive target is identified, it will be highlighted on the monitoring interface using visual indicators, and alarms will be used to alert security personnel.
[0036] Secondly, the present invention provides the following technical solution: an artificial intelligence-based video security monitoring system for implementing any of the above-mentioned video security monitoring methods, the system comprising:
[0037] The facial occlusion recognition module is used to detect the facial features of targets within the monitoring area in real time. When the facial features of the target are detected to be occluded, the target is marked as a target to be tracked.
[0038] The identity fingerprint construction module is used to extract the steady-state physical features and transient apparent features of the target to be tracked, and to construct a consistent identity fingerprint model of the target to be tracked in three-dimensional spatiotemporal based on the steady-state physical features.
[0039] The identity fingerprint matching module is used to obtain the target to be matched from adjacent monitoring nodes, and, in combination with the reachability constraints between monitoring nodes, to filter out the target to be matched whose matching degree with the consistent identity fingerprint model is greater than a preset threshold.
[0040] The appearance feature comparison module is used to extract the transient appearance features of the target to be matched, compare them with the transient appearance features of the target to be tracked, and calculate the feature offset vector.
[0041] The trajectory recording generation module is used to determine that the target to be tracked is an evasion suspect if the feature offset vector exceeds a preset mutation threshold, and associate the video frame sequence of the evasion suspect before and after the feature mutation to generate a spatiotemporal trajectory record.
[0042] The audio-visual alert module is used to perform audio-visual alert actions on the security monitoring equipment in the area where the target to be tracked and the suspected target to be avoided are located.
[0043] The present invention has the following beneficial effects:
[0044] 1. This invention effectively solves the problem of monitoring non-cooperative targets by linking facial occlusion recognition with identity fingerprint modeling. The system does not rely on easily disguised facial information, but instead extracts steady-state features such as human skeleton and gait to construct identity fingerprints, ensuring that targets can still be accurately identified and continuously tracked even when they evade detection behaviors such as facial occlusion or changing clothes.
[0045] 2. This invention utilizes feature offset vectors and spatiotemporal constraints to achieve automated and intuitive evidence chain construction. By calculating the degree of abrupt changes in apparent features and combining this with the physical reachability of monitoring nodes, the system can automatically determine the target's evasion suspicion and generate a keyframe comparison map with heatmap annotations. This highly structured spatiotemporal trajectory recording transforms complex algorithmic judgments into intuitive visual evidence, significantly reducing the cost of manual backtracking. Attached Figure Description
[0046] Figure 1 A flowchart of an artificial intelligence-based video security monitoring method provided in an embodiment of the present invention;
[0047] Figure 2 This is a structural diagram of an artificial intelligence-based video security monitoring system provided in an embodiment of the present invention. Detailed Implementation
[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0049] Example 1
[0050] In the first embodiment of the present invention, the present invention provides a video security monitoring method and system based on artificial intelligence, such as... Figure 1 As shown, it includes:
[0051] S1. Real-time detection of the facial features of targets within the monitoring area; when the facial features of the target are detected to be occluded, the target is marked as a target to be tracked.
[0052] S2. Extract the steady-state physical features and transient apparent features of the target to be tracked, and construct a consistent identity fingerprint model of the target to be tracked in three-dimensional spatiotemporal based on the steady-state physical features.
[0053] S3. Obtain the target to be matched from adjacent monitoring nodes, and in combination with the reachability constraints between monitoring nodes, filter out the target to be matched whose matching degree with the consistent identity fingerprint model is greater than a preset threshold.
[0054] S4. Extract the transient appearance features of the target to be matched, compare them with the transient appearance features of the target to be tracked, and calculate the feature offset vector.
[0055] S5. If the feature offset vector exceeds the preset mutation threshold, the target to be tracked is determined to be an evasion suspect target, and the video frame sequence of the evasion suspect target before and after the feature mutation is associated to generate a spatiotemporal trajectory record.
[0056] S6. Execute an audio-visual alert on the security monitoring equipment in the area where the target to be tracked and the suspected target to be avoided are located.
[0057] Preferably, in step S1, the step of real-time detection of the facial features of the target within the monitoring area includes:
[0058] The facial feature point set of the target is located by the facial key point detection model, and the proportion of the visible points of the facial feature point set in the current image frame is calculated.
[0059] The target head region is scanned using an object detection model to identify whether there are any pre-set obstructions.
[0060] If the proportion of visible points is lower than a preset visibility threshold, or if there are obstructions in the head area, the facial representation is determined to be in an obstructed state.
[0061] In this embodiment, the system calls a pre-trained Dlib facial feature extraction operator to perform facial region locking on each candidate target in the video stream. Sixty-eight standard facial key points, including the outer contours of the eyes, the tip of the nose, the corners of the mouth, and the jawline, are extracted. A confidence threshold is set for each key point. If the detection confidence of a certain key point in the current frame is lower than... If the point's coordinates are outside the facial bounding box, then mark the point as a hidden point or an occluded point. Count the number of all visible points that meet the confidence threshold. Calculate its percentage of the total points. Distribution ratio .when When the preset visibility threshold is reached, the system logic determines that the key facial features of the target are missing.
[0062] Simultaneously launch an object detection model based on YOLOv8 or Faster R-CNN, which has been pre-trained on a dataset containing tens of thousands of labeled occluded object samples. Extract the upper 1 / 5 area of the human target on the monitoring interface as the detection window. The model identifies pixel features within the window and outputs category labels and their probabilities. The preset occlusion object library includes: masks, helmets, sunglasses, hats, etc. If the confidence score of any of the above labels exceeds 0.85, a determination result indicating the presence of a preset occluded object is output.
[0063] If the proportion of visible points is too low or any of the specific occlusions are detected, the system sets the target state to an occluded facial representation.
[0064] Through the above implementation methods, the present invention achieves accurate capture of the identity concealment behavior of non-cooperative targets. By using dual verification of missing geometric feature points and object attribute detection, the probability of identity misjudgment in the system is effectively reduced.
[0065] Preferably, in step S2, the steady-state physical features include skeletal features and gait features; the transient appearance features include clothing features and carried items features.
[0066] In this embodiment, the system uses the HRNet model to locate 17 key skeletal points of the target, including the shoulder, elbow, hip, knee, and ankle. By calculating the Euclidean distance between each joint point, the system extracts the limb length proportions that are unaffected by clothing thickness, forming a static geometric description of the target. The system tracks the displacement trajectory of the skeletal points in a continuous 30-60 frame video sequence and uses the ST-GCN model to extract the target's stride length, stride frequency, and joint swing curves. These features reflect the target's dynamic habits and have extremely high individual uniqueness. Subsequently, global average pooling technology is used to extract the main color and fabric texture features of the target's upper and lower garments. Then, semantic segmentation technology is used to identify the target's surrounding personal items, recording their category and relative position height as transient tags to assist in identity determination.
[0067] Through the above implementation methods, the present invention combines the skeletal gait at the biomechanical level with the clothing attributes at the visual level, solving the problem of easy target loss caused by traditional security relying on faces or clothing.
[0068] Preferably, in step S2, the step of constructing a consistent identity fingerprint model of the target to be tracked in three-dimensional spatiotemporal space includes:
[0069] Based on skeletal features, the relative geometric proportions between various joints of the human body are calculated to generate static structural vectors.
[0070] Based on gait features, periodic motion parameters of key point displacements of the target within a preset time window are extracted to generate dynamic behavior vectors.
[0071] The static structure vector and the dynamic behavior vector are concatenated for features and mapped to a unified metric space to form a consistent identity fingerprint model of the target to be tracked.
[0072] In this embodiment, the system extracts the coordinates of key points on the entire body of the target using a preset human pose estimation neural network, and calculates the Euclidean distance ratios between non-adjacent key points, such as: thigh length + calf length / torso length, and shoulder width / pelvic width. These relative geometric proportions do not change with the target's distance or scaling in the image, forming a static structural vector representing the anatomical structure of the human body.
[0073] The system records the displacement curves of key points such as the feet and knees within a preset sliding window time. The sliding window is preferably set to 2 seconds, corresponding to 1-2 complete gait cycles. Periodic motion parameters such as oscillation frequency, phase difference, and centroid undulation height in the motion trajectory are extracted using Fast Fourier Transform. These parameters reflect the target's unique walking posture, generating a dynamic behavior vector with biomechanical characteristics.
[0074] The system concatenates the static structure vector and dynamic behavior vector along the feature dimension. To eliminate observation biases caused by different sensors and perspectives, the system projects the concatenated features onto a unified, metrically meaningful three-dimensional vector space, ultimately forming a consistent identity fingerprint model for the target.
[0075] Through the above implementation methods, even when the target uses evasion methods such as masking their face or changing clothes, the present invention can still achieve accurate identification by relying on the unique fingerprint in three-dimensional spacetime, which greatly enhances the reliability of the monitoring system in real-world environments.
[0076] Preferably, in step S3, the step of filtering out the targets to be matched whose matching degree with the consistent identity fingerprint model is greater than a preset threshold includes:
[0077] Based on the physical distance between adjacent monitoring nodes and the preset target movement speed range, determine the expected time window for the target to reach the adjacent monitoring nodes;
[0078] Extract the initial features of all targets to be compared that appear in adjacent monitoring nodes within the expected time window, and remove targets that do not conform to spatiotemporal logic;
[0079] The consistency fingerprint model of the target to be tracked is compared with the identity fingerprint of the initially selected target to be compared by calculating the similarity distance.
[0080] If the similarity distance is less than a preset measurement threshold, the corresponding target to be compared is determined to be a successful match and is used as the target to be matched.
[0081] In this embodiment, the system pre-stores the topology map and physical distances between each camera node in the monitoring network. Based on the normal walking and running speeds of the human body, a movement speed range is set. These correspond to the minimum and maximum set speeds, respectively. After the target disappears from node A, the system automatically calculates the expected time window for it to reach the adjacent node B. ,in This indicates the start time of the time window. , indicates the end time of the time window.
[0082] The system only applies within the expected time window. Feature extraction is performed on targets appearing at node B. If a target appears much earlier than... or much later The system will automatically eliminate the target without performing subsequent complex fingerprint comparisons. This initial selection mechanism based on spatiotemporal logic can effectively filter out irrelevant pedestrians.
[0083] For candidate targets that pass the initial selection, the system extracts their identity fingerprints and calculates the similarity distance between them and the consistent identity fingerprint model of the target to be tracked. The cosine similarity is preferred when calculating the similarity distance. Set the measurement threshold to... If the calculated distance This proves that the skeletal proportions and gait patterns of the two are highly consistent. The system officially determines that the candidate target is a successful match and confirms it as a new target to be matched, maintaining the continuity of the identity chain.
[0084] Through the above implementation methods, the present invention uses reachability rules to establish a pre-filter, which not only greatly reduces the computing overhead of the back-end server in the context of massive traffic, but also eliminates a large number of mismatches caused by visual similarity from the root.
[0085] Preferably, in step S4, the step of calculating the feature offset vector includes:
[0086] The transient appearance features of the target to be tracked and the target to be matched are divided into multiple local feature blocks, which include the head region, torso region, leg region and carrying object region;
[0087] Calculate the distances between the target to be tracked and the target to be matched in each corresponding local feature block;
[0088] The distance calculation results of each local feature block are weighted and combined to generate the feature offset vector that characterizes the degree of global appearance difference.
[0089] In this embodiment, the system uses MaskR-CNN to divide the transient appearance feature map of the target into four independent semantic feature blocks according to the human anatomical structure, including the head region, torso region, leg region and object region.
[0090] The system extracts feature vectors of the target to be tracked and the target to be matched in the corresponding regions mentioned above, and calculates their similarity distance for each. For example, it calculates the distance in the torso region. If the target is wearing a white shirt at monitoring point A and changes into a black jacket at monitoring point B, then the local distance... Extreme numerical jumps may occur.
[0091] The system sets weighting coefficients based on the contribution of each part to identity recognition and clothing verification. Preferably, the torso and legs each account for 40% of the weight, and the head and carried items each account for 10%. The final generated feature offset vector... It consists of a weighted set of local distances, i.e.:
[0092] ;
[0093] in, , , , The distances correspond to the head area, torso area, leg area, and carried item area, respectively. , , , The corresponding weights for these regions. At this point, if... If the mutation threshold is exceeded, the system can determine that the target has engaged in apparent evasion behavior, provided that the identity fingerprint is successfully matched.
[0094] Through the above implementation methods, the present invention decomposes the overall features into local feature blocks to capture attribute mutations occurring in local areas, ensuring that the system can not only identify the same person before and after the change of clothes, but also accurately determine whether the person has the suspicion of subjectively evading monitoring.
[0095] Preferably, in step S5, the step of generating the spatiotemporal trajectory record includes:
[0096] Automatically extract the first keyframe image of the target to be tracked before the feature mutation, and the second keyframe image of the suspected target to be avoided after the feature mutation;
[0097] In the first keyframe image and the second keyframe image, heatmaps are used to mark the regions of abrupt changes in apparent properties that cause the feature offset vector to exceed the threshold.
[0098] By associating the monitoring nodes corresponding to the first keyframe image and the second keyframe image, and integrating the target's movement path and corresponding timestamp information between different monitoring nodes, a structured spatiotemporal trajectory record is generated.
[0099] In this embodiment, the last normal record of the target before the feature mutation is traced and extracted, which serves as the optimal visual image of the target at the previous monitoring node or the initial stage of the current node; the current image frame when the target is identified as an evasive suspect is captured in real time. These two keyframe images record the original appearance of the target before and after the feature mutation, providing a visual basis for subsequent comparison.
[0100] Based on the distances of each local feature block calculated in step S4, the system determines the specific location causing the vector to exceed the limit. The calculated sub-offset values are then mapped to heatmaps of different color depths, which are overlaid on corresponding positions in the first and second keyframes. For example, if a sudden increase in the offset vector is detected due to a change in the color of the upper garment from white to black, the system marks red heatmap patches on the torso position in both keyframes, visually representing the apparent attribute abrupt change that leads to the suspicion of evasion.
[0101] The system deeply correlates the aforementioned image evidence with physical spatiotemporal information. First, it automatically reads and records the monitoring node ID, geographical location information, and capture timestamp corresponding to the first and second keyframes. Then, it retrieves the historical trajectory of the target's movement among all monitoring nodes since it was marked as a target to be tracked, and combines it with the aforementioned timestamp information to generate a movement path map through chronological arrangement. Finally, it integrates the keyframe comparison map, attribute heatmap, and trajectory data containing path and time to generate a structured spatiotemporal trajectory record that can be traced back.
[0102] Through the above implementation methods, the present invention uses heat maps to intuitively mark abrupt changes in apparent attributes, and combines physical and spatiotemporal information in the monitoring network to strongly correlate the target's movement trajectory with clothing-changing behavior.
[0103] Preferably, in step S6, the step of performing the sound and light reminder action includes:
[0104] When the target to be tracked is identified, it is highlighted on the monitoring interface using a visual identifier.
[0105] When a suspected evasive target is identified, it will be highlighted on the monitoring interface using visual indicators, and alarms will be used to alert security personnel.
[0106] Specifically, when the system identifies a target with facial occlusion or missing identity information in step S1 and marks it as a target to be tracked, the monitoring backend triggers a basic alert mechanism. A semi-transparent, highlighted rectangle of a first preset color is automatically generated on the target's outline in the real-time monitoring screen, and real-time rendering and tracking are performed as the target moves.
[0107] When the system determines in steps S4 and S5 that the feature offset vector exceeds the limit and upgrades it to a suspected target for evasion, it triggers a high-level linkage alarm logic. The highlighted box on the monitoring interface immediately switches to the second preset color to distinguish it from regular targets. At the same time, the system links with the audio output device in the monitoring center to play an alarm sound to alert security personnel.
[0108] Through the above implementation methods, the present invention achieves differentiated processing of security levels.
[0109] Example 2
[0110] This invention also provides an artificial intelligence-based video security monitoring system, the structure of which is as follows: Figure 2 As shown, it includes:
[0111] The facial occlusion recognition module is used to detect the facial features of targets within the monitoring area in real time. When the facial features of the target are detected to be occluded, the target is marked as a target to be tracked.
[0112] The identity fingerprint construction module is used to extract the steady-state physical features and transient apparent features of the target to be tracked, and to construct a consistent identity fingerprint model of the target to be tracked in three-dimensional spatiotemporal based on the steady-state physical features.
[0113] The identity fingerprint matching module is used to obtain the target to be matched from adjacent monitoring nodes, and, in combination with the reachability constraints between monitoring nodes, to filter out the target to be matched whose matching degree with the consistent identity fingerprint model is greater than a preset threshold.
[0114] The appearance feature comparison module is used to extract the transient appearance features of the target to be matched, compare them with the transient appearance features of the target to be tracked, and calculate the feature offset vector.
[0115] The trajectory recording generation module is used to determine that the target to be tracked is an evasion suspect if the feature offset vector exceeds a preset mutation threshold, and associate the video frame sequence of the evasion suspect before and after the feature mutation to generate a spatiotemporal trajectory record.
[0116] The audio-visual alert module is used to perform audio-visual alert actions on the security monitoring equipment in the area where the target to be tracked and the suspected target to be avoided are located.
[0117] To verify the effectiveness of this system, a simulated surveillance test was conducted in a large luxury goods store and its surrounding public areas. The test environment deployed 20 interconnected monitoring nodes, covering entrances and exits, shelf areas, and fitting room aisles.
[0118] Testers simulated the suspect performing the following evasive maneuvers:
[0119] Initial entry: Wear a mask, sunglasses, and a baseball cap to enter the store.
[0120] During the journey, she changed her clothes: she took off her dark coat in a blind spot of the surveillance cameras, revealing a bright-colored shirt underneath, and changed the bag she was carrying.
[0121] Escape across points: Quickly use the escalator to reach another monitoring node on another floor.
[0122] Table 1 shows a comparison of the performance of this system with traditional facial recognition surveillance systems under the same conditions. Despite the suspect changing clothes, the gait proportions and skeletal structure features extracted by the identity fingerprint construction module remained stable. When matching adjacent nodes, the system, through the identity fingerprint matching module, still managed to lock onto the suspect with a 92.3% confidence level in a complex environment containing over 50 customers.
[0123] Table 1. Performance Comparison between the Invention and the Face Recognition Monitoring System
[0124]
[0125] The simulation test results above demonstrate that this system can effectively compensate for the blind spots in traditional surveillance when faced with behaviors such as deliberate obstruction and mid-journey clothing changes. The system does not rely on facial information; instead, it achieves continuous and accurate targeting of high-risk targets by constructing a deep identity fingerprint model and combining it with spatiotemporal logic constraints.
[0126] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A video security monitoring method based on artificial intelligence, characterized in that, include: S1. Real-time detection of the facial features of targets within the monitoring area; when the facial features of the target are detected to be occluded, the target is marked as a target to be tracked. S2. Extract the steady-state physical features and transient apparent features of the target to be tracked, and construct a consistent identity fingerprint model of the target to be tracked in three-dimensional spatiotemporal based on the steady-state physical features. S3. Obtain the target to be matched from adjacent monitoring nodes, and in combination with the reachability constraints between monitoring nodes, filter out the target to be matched whose matching degree with the consistent identity fingerprint model is greater than a preset threshold. S4. Extract the transient appearance features of the target to be matched, compare them with the transient appearance features of the target to be tracked, and calculate the feature offset vector. S5. If the feature offset vector exceeds the preset mutation threshold, the target to be tracked is determined to be an evasion suspect target, and the video frame sequence of the evasion suspect target before and after the feature mutation is associated to generate a spatiotemporal trajectory record. S6. Execute an audio-visual alert on the security monitoring equipment in the area where the target to be tracked and the suspected target to be avoided are located.
2. The video security monitoring method based on artificial intelligence according to claim 1, characterized in that, In step S1, the step of real-time detection of the facial representation of the target within the monitoring area includes: The facial feature point set of the target is located by the facial key point detection model, and the proportion of the visible points of the facial feature point set in the current image frame is calculated. The target head region is scanned using an object detection model to identify whether there are any pre-set obstructions. If the proportion of visible points is lower than a preset visibility threshold, or if there are obstructions in the head area, the facial representation is determined to be in an obstructed state.
3. The video security monitoring method based on artificial intelligence according to claim 1, characterized in that, In step S2, the steady-state physical features include skeleton features and gait features; The transient appearance features include clothing features and carried items features.
4. The video security monitoring method based on artificial intelligence according to claim 1, characterized in that, In step S2, the step of constructing a consistent identity fingerprint model of the target to be tracked in three-dimensional spatiotemporal space includes: Based on skeletal features, the relative geometric proportions between various joints of the human body are calculated to generate static structural vectors. Based on gait features, periodic motion parameters of key point displacements of the target within a preset time window are extracted to generate dynamic behavior vectors. The static structure vector and the dynamic behavior vector are concatenated for features and mapped to a unified metric space to form a consistent identity fingerprint model of the target to be tracked.
5. The video security monitoring method based on artificial intelligence according to claim 1, characterized in that, In step S3, the step of filtering out the targets to be matched whose matching degree with the consistent identity fingerprint model is greater than a preset threshold includes: Based on the physical distance between adjacent monitoring nodes and the preset target movement speed range, determine the expected time window for the target to reach the adjacent monitoring nodes; Extract the initial features of all targets to be compared that appear in adjacent monitoring nodes within the expected time window, and remove targets that do not conform to spatiotemporal logic; The consistency fingerprint model of the target to be tracked is compared with the identity fingerprint of the initially selected target to be compared by calculating the similarity distance. If the similarity distance is less than a preset measurement threshold, the corresponding target to be compared is determined to be a successful match and is used as the target to be matched.
6. The video security monitoring method based on artificial intelligence according to claim 1, characterized in that, In step S4, the step of calculating the feature offset vector includes: The transient appearance features of the target to be tracked and the target to be matched are divided into multiple local feature blocks, which include the head region, torso region, leg region and carrying object region; Calculate the distances between the target to be tracked and the target to be matched in each corresponding local feature block; The distance calculation results of each local feature block are weighted and combined to generate the feature offset vector that characterizes the degree of global appearance difference.
7. The video security monitoring method based on artificial intelligence according to claim 1, characterized in that, In step S5, the steps for generating spatiotemporal trajectory records include: Automatically extract the first keyframe image of the target to be tracked before the feature mutation, and the second keyframe image of the suspected target to be avoided after the feature mutation; In the first keyframe image and the second keyframe image, heatmaps are used to mark the regions of abrupt changes in apparent properties that cause the feature offset vector to exceed the threshold. By associating the monitoring nodes corresponding to the first keyframe image and the second keyframe image, and integrating the target's movement path and corresponding timestamp information between different monitoring nodes, a structured spatiotemporal trajectory record is generated.
8. The video security monitoring method based on artificial intelligence according to claim 1, characterized in that, In step S6, the steps for performing the audio-visual reminder action include: When the target to be tracked is identified, it is highlighted on the monitoring interface using a visual identifier. When a suspected evasive target is identified, it will be highlighted on the monitoring interface using visual indicators, and alarms will be used to alert security personnel.
9. A video security monitoring system based on artificial intelligence, characterized in that, The system for implementing the video security monitoring method according to any one of claims 1-8, the system comprising: The facial occlusion recognition module is used to detect the facial features of targets within the monitoring area in real time. When the facial features of the target are detected to be occluded, the target is marked as a target to be tracked. The identity fingerprint construction module is used to extract the steady-state physical features and transient apparent features of the target to be tracked, and to construct a consistent identity fingerprint model of the target to be tracked in three-dimensional spatiotemporal based on the steady-state physical features. The identity fingerprint matching module is used to obtain the target to be matched from adjacent monitoring nodes, and, in combination with the reachability constraints between monitoring nodes, to filter out the target to be matched whose matching degree with the consistent identity fingerprint model is greater than a preset threshold. The appearance feature comparison module is used to extract the transient appearance features of the target to be matched, compare them with the transient appearance features of the target to be tracked, and calculate the feature offset vector. The trajectory recording generation module is used to determine that the target to be tracked is an evasion suspect if the feature offset vector exceeds a preset mutation threshold, and associate the video frame sequence of the evasion suspect before and after the feature mutation to generate a spatiotemporal trajectory record. The audio-visual alert module is used to perform audio-visual alert actions on the security monitoring equipment in the area where the target to be tracked and the suspected target to be avoided are located.