Cultural relic safety protection Internet of Things system based on YOLOv8 behavior recognition algorithm
By using an IoT system based on the YOLOv8 behavior recognition algorithm, combined with multimodal data fusion and cloud management, the problem of behavioral intent recognition in complex scenarios of cultural relic security systems has been solved, achieving efficient and accurate behavior recognition and rapid response, thereby improving the security of cultural relics.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-03-10
AI Technical Summary
Existing cultural relic security systems are unable to effectively distinguish behavioral intentions in complex exhibition scenarios, resulting in a high false alarm rate, reduced response efficiency to high-risk behaviors, increased workload for security personnel, and potential security risks.
An IoT system based on the YOLOv8 behavior recognition algorithm combines visual data, distance data, and environmental data. Through multimodal fusion decision-making at the edge computing layer and coordinated control at the cloud management layer, it achieves accurate identification and differentiated response to behavioral intentions.
It achieves accurate identification of behavioral intent, reduces the false alarm rate to below 5%, improves the accuracy of target detection by 35% in low light conditions, and has a system response time of ≤3 seconds. It effectively distinguishes between unintentional approach and malicious behavior, thereby improving the safety of cultural relics.
Smart Images
Figure CN121644618A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the security technology field, and in particular to a cultural relic security protection Internet of Things system based on a YOLOv8 behavior recognition algorithm. BACKGROUND
[0002] Cultural relics are the material carriers of human civilization, and museums are the core places for the collection and display of cultural relics. The reliability of the security system of a museum is directly related to the safety of cultural relics. The current mainstream cultural relic security system still takes "sensor alarm + video monitoring" or "single visual recognition + simple linkage" as the core mode. Although it has certain improvement compared with the early manual inspection, there are still key technical bottlenecks in complex exhibition scenes. The most prominent problem is the lack of behavior intention recognition capability: the existing system can only realize basic event detection such as "whether personnel are close to the showcase", and cannot effectively distinguish the behavior intention - for example, the visual features of the audience bending over to tie shoelaces (unintentionally close) and squatting to pick the lock of the showcase (malicious behavior) are similar, and traditional algorithms are prone to misjudgment, resulting in a false alarm rate as high as 35% (according to industry statistics). This not only increases the invalid workload of security personnel, but also easily leads to the "boy who cried wolf" effect, reduces the response efficiency to real theft, damage and other high-risk behaviors, and brings hidden risks to cultural relic safety. SUMMARY
[0003] In view of the above deficiencies in the prior art, the purpose of the present application is to provide a cultural relic security protection Internet of Things system based on a YOLOv8 behavior recognition algorithm, which solves the problem of poor behavior intention recognition capability of the security protection system in the prior art.
[0004] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:
[0005] A cultural relic security protection Internet of Things system based on a YOLOv8 behavior recognition algorithm, comprising a perception layer, an edge computing layer, a cloud management layer and an execution control layer connected in sequence;
[0006] The perception layer is used to collect visual data, distance data and environmental data in the exhibition hall, and realize the spatio-temporal synchronization of multi-modal data;
[0007] The edge computing layer deploys a YOLOv8 target detection algorithm and a Transformer behavior recognition algorithm, and performs preprocessing, inference calculation and multi-modal fusion decision on the collected data, and outputs a risk level;
[0008] The cloud management layer receives risk level information, performs data storage, model iteration optimization and hierarchical decision, and generates a linkage control instruction;
[0009] The execution control layer executes a security linkage action according to the linkage control instruction.
[0010] Preferably, the perception layer comprises a visual perception unit, a distance perception unit, an environment perception unit and a data synchronization module;
[0011] The visual perception unit adopts a high-definition network camera and is arranged at four corners of an exhibition hall and above a showcase; the distance perception unit adopts a millimeter wave radar sensor and is installed at a front side of the showcase and divides a safety zone, a warning zone and a danger zone; the environment perception unit comprises a temperature and humidity sensor, an illumination sensor and a smoke sensor; and the data synchronization module realizes device time synchronization through an NTP protocol and associates multi-modal data based on a "time stamp + frame number".
[0012] Preferably, the edge computing layer comprises a data preprocessing module, a target detection module, a behavior recognition module and a multi-modal fusion decision module.
[0013] The data preprocessing module adopts a Retinex algorithm to enhance low-light details and an adaptive median filter to remove noise for visual data, and adopts Kalman filter denoising and a DBSCAN algorithm clustering for radar data.
[0014] The target detection module adopts a YOLOv8-L model, and a loss function thereof is as follows:
[0015] L total =λ cls ·L cls +λ box ·L box +λ obj ·L obj
[0016] Wherein, L cls is a binary cross-entropy classification loss, L box is a CIoU bounding box regression loss,
[0017] L obj is a confidence loss, λ cls =1.0, λ box =5.0 and λ obj =1.0.
[0018] Preferably, the behavior recognition module adopts an improved Transformer model, increases time position encoding and space position encoding, inputs a personnel behavior time sequence frame sequence, and outputs behavior categories such as normal visiting, close-range staying, squatting and approaching, carrying dangerous tools and prying a showcase and corresponding risk levels; and a multi-head self-attention calculation formula of the Transformer model is as follows:
[0019] MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O
[0020] in,
[0021]
[0022] h = 8, d k =32.
[0023] Preferably, the multimodal fusion decision module adopts a weighted voting fusion mechanism: when the radar detects a target entering the danger zone and the behavior is identified as a crouching approach, the risk level is increased by 1 level; when the vision detects a dangerous tool and the radar distance is <2m, it is directly judged as high risk; when the environmental sensor detects excessive smoke, the fire emergency mode is triggered.
[0024] Preferably, the cloud management layer includes a data storage module, a model iteration and optimization module, a decision engine module, and a remote monitoring module;
[0025] The model iteration and optimization module adopts the FedAvg algorithm, where local model parameters are uploaded from the edge to the cloud for aggregation to generate a global model. The aggregation formula is as follows:
[0026]
[0027] Where, n i w represents the number of samples at edge node i. i These are the model parameters for node i.
[0028] Preferably, the decision engine module outputs hierarchical linkage instructions based on rule engine and fuzzy logic: low risk only records data; medium risk pushes security warning information; high risk triggers audible and visual alarms and pre-locks the display case lock; extremely high risk triggers high-decibel audible and visual alarms, closes the exhibition hall access control, starts emergency protection of the display case and dispatches security robots; fire mode triggers the fireproof cover to fall and emergency evacuation broadcast.
[0029] Preferably, the execution control layer includes an audible and visual alarm unit, a display case emergency protection unit, an access control unit, and a security patrol robot unit; the display case emergency protection unit includes an electromagnetic lock, a transparent conductive film glass, and a high-temperature resistant silicone fireproof cover. In the event of an extremely high risk, the electromagnetic lock is locked and the conductive film glass becomes opaque when energized. In the event of a fire, the fireproof cover automatically falls.
[0030] Preferably, the edge computing layer uses TensorRT to accelerate the YOLOv8 model with INT8 quantization, achieving an inference speed of ≥30fps and an end-to-end inference time of ≤200ms; the overall system response time is ≤3 seconds, and the target detection mAP@0.5 accuracy is ≥95% under low light conditions, with an action recognition accuracy of ≥96%.
[0031] Preferably, the cultural relic security IoT system is compatible with museum security equipment, connects to existing cameras via the ONVIF protocol, and interfaces with the existing access control system via SDK; both the edge computing layer and the cloud management layer adopt a primary-backup redundant deployment to ensure system reliability.
[0032] Compared with the prior art, the present invention has the following beneficial effects:
[0033] 1. The improved Transformer model of this invention achieves accurate classification of typical museum behaviors through spatiotemporal position encoding and multi-head self-attention mechanism, with a behavior recognition accuracy of ≥96% and a false alarm rate of ≤5%, effectively distinguishing between "unintentional approach" and "malicious behavior", thus solving the core pain point of high false alarm rate in traditional systems;
[0034] 2. The Retinex low-light enhancement and adaptive median filtering de-reflection preprocessing techniques of this invention, combined with the YOLOv8-L model optimized by CIoU bounding box regression loss, achieve a target detection mAP@0.5≥95% in low-light (30lx) and glass reflection scenarios, which is more than 35% higher than traditional algorithms; the multimodal fusion of visual data and radar data further improves the system's stable operation rate (≥99.5%) in scenarios with personnel occlusion and environmental interference.
[0035] 3. The edge computing layer of this invention adopts TensorRT quantization acceleration, with YOLOv8 model inference speed ≥30fps, end-to-end inference time ≤200ms, and overall system response time (from the occurrence of a risk event to the initiation of linkage action) ≤3 seconds; the differentiated linkage mechanism based on risk level avoids over-response to minor anomalies and initiates comprehensive protection for extremely high-risk events. Attached Figure Description
[0036] Figure 1 This is a schematic diagram of the present invention. Detailed Implementation
[0037] To illustrate the technical content, structural features, objectives, and effects of the present invention in detail, the following description is provided in conjunction with the embodiments and accompanying drawings.
[0038] Please see Figure 1 A cultural relic security IoT system based on the YOLOv8 behavior recognition algorithm includes a sensing layer, an edge computing layer, a cloud management layer, and an execution control layer that are electrically connected in sequence.
[0039] The perception layer is used to collect visual data, distance data, and environmental data within the exhibition hall, and to achieve spatiotemporal synchronization of multimodal data. The perception layer includes a visual perception unit, a distance perception unit, an environmental perception unit, and a data synchronization module.
[0040] The visual perception unit employs high-definition network cameras, positioned at the four corners of the exhibition hall and above the display cases, enabling localized monitoring of the cases. The distance perception unit uses a 60-64GHz millimeter-wave radar sensor with a detection range of 0.1-10m and an accuracy of ±0.05m. It is installed in front of the display cases and divides the area into a safe zone (3-5m), a warning zone (1-3m), and a danger zone (0-1m). The environmental perception unit includes temperature and humidity sensors, a light sensor, and a smoke sensor, collecting environmental parameters in real time. The data synchronization module achieves device time synchronization via the NTP protocol and associates multimodal data based on "timestamp + frame number," ensuring spatiotemporal alignment.
[0041] The edge computing layer deploys the YOLOv8 object detection algorithm and the Transformer behavior recognition algorithm to preprocess the collected data, perform inference calculations, and conduct multimodal fusion decision-making, outputting a risk level. The edge computing layer includes a data preprocessing module, an object detection module, a behavior recognition module, and a multimodal fusion decision-making module.
[0042] The data preprocessing module uses the Retinex algorithm to enhance low-light details and adaptive median filtering to remove noise for visual data, and uses Kalman filtering for noise reduction and DBSCAN algorithm for clustering for radar data.
[0043] The target detection module uses the YOLOv8-L model, and its loss function is:
[0044] L total =λ cls ·L cls +λ box ·L box +λ obj ·L obj
[0045] Among them, L cls For binary cross-entropy classification loss, L box For CIoU bounding box regression loss,
[0046] L obj For confidence loss, λ cls =1.0, λ box =5.0, λ obj =1.0.
[0047] The behavior recognition module employs an improved Transformer model, incorporating temporal and spatial location encoding. The input is a sequence of temporal frames representing human behavior, and the output includes behavior categories and corresponding risk levels for normal visits, close-range stops, crouching approaches, carrying dangerous tools, and prying open display cases. The multi-head self-attention calculation formula for the Transformer model is as follows:
[0048] MultiHead(Q, K, V) = Concat(head1, head2,..., head h )W O
[0049] Among them,
[0050]
[0051] h = 8, d k = 32.
[0052] The multi-modal fusion decision-making module adopts a weighted voting fusion mechanism: when the radar detects that the target enters the danger zone and the behavior recognition is a crouching approach, the risk level is increased by 1 level; when the vision detects a dangerous tool and the radar distance < 2m, it is directly determined as a high risk; when the environmental sensor detects that the smoke exceeds the standard, the fire emergency mode is triggered.
[0053] The edge computing layer uses TensorRT to perform INT8 quantization acceleration on the YOLOv8 model, with an inference speed ≥ 30fps and an end-to-end inference time ≤ 200ms; the overall system response time ≤ 3 seconds, the target detection mAP@0.5 ≥ 95% in low-light environments, and the behavior recognition accuracy ≥ 96%.
[0054] The cloud management layer receives the risk level information, performs data storage, model iteration optimization, and hierarchical decision-making, and generates linkage control instructions. The cloud management layer includes a data storage module, a model iteration optimization module, a decision engine module, and a remote monitoring module;
[0055] The model iteration optimization module uses the FedAvg algorithm of federated learning. The edge device uploads the local model parameters to the cloud for aggregation to generate a global model. The aggregation formula is:
[0056]
[0057] Among them, n i is the number of samples of edge node i, and w i is the model parameter of node i.
[0058] The decision engine module outputs hierarchical linkage instructions based on the rule engine and fuzzy logic: low risk only records data; medium risk pushes security warning information; high risk triggers an audible and visual alarm and pre-locks the showcase lock; extremely high risk triggers a high-decibel audible and visual alarm, closes the exhibition hall access control, activates the showcase emergency protection, and deploys security robots; the fire mode triggers the fireproof cover to fall and an emergency evacuation broadcast.
[0059] The execution control layer executes security linkage actions according to the linkage control command. The execution control layer includes an audible and visual alarm unit, a display case emergency protection unit, an access control unit, and a security patrol robot unit; the display case emergency protection unit includes an electromagnetic lock, transparent conductive film glass, and a high-temperature resistant silicone fireproof cover. In the event of extremely high risk, the electromagnetic lock locks and the conductive film glass becomes opaque when energized; in fire mode, the fireproof cover automatically falls.
[0060] The cultural relic security IoT system is compatible with museum security equipment, connects to existing cameras via the ONVIF protocol, and interfaces with the existing access control system via SDK; both the edge computing layer and the cloud management layer adopt a primary and backup redundancy deployment to ensure system reliability.
[0061] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the present invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.
Claims
1. A cultural relic safety protection Internet of Things system based on a YOLOv8 behavior recognition algorithm, characterized in that, The perception layer, the edge computing layer, the cloud management layer and the execution control layer are sequentially and electrically connected; The perception layer is used for collecting visual data, distance data and environmental data in the exhibition hall, and realizing space-time synchronization of multi-modal data; The edge computing layer deploys a YOLOv8 target detection algorithm and a Transformer behavior recognition algorithm, and performs preprocessing, inference calculation and multi-modal fusion decision on the collected data, and outputs a risk level; The cloud management layer receives risk level information, performs data storage, model iteration optimization and hierarchical decision, and generates a linkage control instruction; The execution control layer executes a security linkage action according to the linkage control instruction.
2. The cultural relic security protection Internet of Things system based on the YOLOv8 behavior recognition algorithm of claim 1, wherein The perception layer includes a visual perception unit, a distance perception unit, an environmental perception unit and a data synchronization module; The visual perception unit adopts a high-definition network camera and is arranged at four corners of the exhibition hall and above the showcases; the distance perception unit adopts a millimeter wave radar sensor and is installed in front of the showcases and divides a safety zone, a warning zone and a danger zone; the environmental perception unit includes a temperature and humidity sensor, a light sensor and a smoke sensor; and the data synchronization module realizes device time synchronization through an NTP protocol and associates multi-modal data based on a "time stamp + frame number".
3. The cultural relic security protection Internet of Things system based on the YOLOv8 behavior recognition algorithm of claim 1, wherein The edge computing layer includes a data preprocessing module, a target detection module, a behavior recognition module and a multi-modal fusion decision module; The data preprocessing module adopts a Retinex algorithm to enhance low-light details and an adaptive median filter to remove noise for visual data, and adopts Kalman filtering for radar data denoising and a DBSCAN algorithm for clustering; The target detection module adopts a YOLOv8-L model, and a loss function thereof is: L total = λ cls · L cls + λ box · L box + λ obj · L obj where L cls is the binary cross-entropy classification loss, L box is the CIoU bounding box regression loss, L obj is the confidence loss, λ cls = 1.0, λ box = 5.0, and λ obj = 1.
0.
4. The cultural relic security protection Internet of Things system based on the YOLOv8 behavior recognition algorithm of claim 1, wherein The behavior recognition module adopts an improved Transformer model, increases time position encoding and space position encoding, and inputs a personnel behavior time sequence frame sequence and outputs behavior categories including normal visiting, close-range staying, squatting and approaching, carrying dangerous tools and prying a showcase and corresponding risk levels; a multi-head self-attention calculation formula of the Transformer model is: MultiHead(Q, K, V) = Concat(head1, head2,..., head h )W O wherein h = 8, d k = 32.
5. The cultural relic security protection Internet of Things system based on the YOLOv8 behavior recognition algorithm of claim 1, wherein The multi-modal fusion decision module adopts a weighted voting fusion mechanism: when a radar detected target enters a danger zone and behavior recognition is squatting and approaching, the risk level is increased by 1 level; when a dangerous tool is detected by vision and the radar distance is less than 2 m, it is directly determined as high risk; and when an environmental sensor detects that smoke exceeds a standard, a fire emergency mode is triggered.
6. The cultural relic security protection Internet of Things system based on the YOLOv8 behavior recognition algorithm of claim 1, wherein The cloud management layer includes a data storage module, a model iteration optimization module, a decision engine module and a remote monitoring module; The model iteration optimization module adopts a federated learning FedAvg algorithm, uploads local model parameters to the cloud to aggregate and generate a global model, and an aggregation formula is: where n i is the number of samples for edge node i, w i is the model parameter for node i.
7. The cultural relic security protection Internet of Things system based on the YOLOv8 behavior recognition algorithm according to claim 6, characterized in that, The decision engine module outputs hierarchical linkage instructions based on a rule engine and fuzzy logic: low risk only records data; medium risk pushes security warning information; high risk triggers sound and light alarms and pre-locks a showcase lock; extremely high risk triggers high-decibel sound and light alarms, closes an exhibition hall access control, starts a showcase emergency protection and dispatches a security robot; and the fire mode triggers a fire cover to fall and an emergency evacuation broadcast. 8. The cultural relic security protection Internet of Things system based on the YOLOv8 behavior recognition algorithm of claim 1, wherein The execution control layer comprises an acousto-optic alarm unit, a showcase emergency protection unit, an access control unit and a security patrol robot unit; the showcase emergency protection unit comprises an electromagnetic lock, a transparent conductive film glass and a high-temperature-resistant silica gel fireproof cover, the electromagnetic lock is locked and the conductive film glass is powered to become opaque in an extremely high risk situation, and the fireproof cover automatically falls in a fire mode.
9. The cultural relic security protection Internet of Things system based on the YOLOv8 behavior recognition algorithm of claim 1, wherein The edge computing layer adopts TensorRT to perform INT8 quantization acceleration on the YOLOv8 model, the inference speed is greater than or equal to 30 fps, the end-to-end inference time is less than or equal to 200 ms, the overall response time of the system is less than or equal to 3 seconds, the target detection mAP@0.5 in a low-light environment is greater than or equal to 95%, and the behavior recognition accuracy is greater than or equal to 96%.
10. The cultural relic security protection Internet of Things system based on the YOLOv8 behavior recognition algorithm of claim 1, wherein The cultural relic safety protection Internet of Things system is compatible with museum security equipment, accesses original cameras through an ONVIF protocol and interfaces with an original access control system through an SDK; the edge computing layer and the cloud management layer are both deployed in a primary and backup redundant mode, thereby guaranteeing the reliability of the system.