Abnormal behavior alarm device suitable for high-density crowd place

Abstract

The invention relates to the technical field of intelligent security and protection, in particular to an abnormal behavior warning device suitable for a high-density crowd place, which comprises a support structure and a functional module mounted on the support structure, and is characterized in that the functional module comprises a multi-mode sensing module, a double-strategy analysis module, an intelligent warning linkage module and a local data storage module; the multi-mode sensing module is electrically connected with the double-strategy analysis module, the double-strategy analysis module is electrically connected with the intelligent alarm linkage module, and the intelligent alarm linkage module and the multi-mode sensing module are electrically connected with the local data storage module; the multi-mode sensing module comprises an intelligent audio and video acquisition unit and a millimeter wave radar acquisition unit, and the intelligent audio and video acquisition unit is used for acquiring video pictures and audio signals of a high-density crowd place. The invention provides the abnormal behavior alarm device suitable for the high-density crowd place, and the device has the advantages of multi-scene adaptability, high recognition precision, low resource consumption and privacy security.

Description

Technical Field

[0001] This invention relates to the field of intelligent security technology, specifically to an abnormal behavior alarm device suitable for high-density crowd locations. Background Technology

[0002] As we all know, with the continuous increase in global urbanization rates, high-density crowd places such as subway stations, shopping malls, and transportation hubs are densely populated and have complex environments. Typical abnormal behaviors, such as escalator reversal, pedestrian falls, and abnormal gatherings, can easily lead to safety accidents or even public disorder if they cannot be identified and intervened in a timely manner. In addition, high-density crowd gatherings can also easily cause stampedes, physical conflicts, and other serious incidents. As the density of urban population continues to increase, the safety hazards in public spaces are becoming increasingly prominent, and the demand for abnormal behavior identification is showing an exponential growth trend, becoming a key area of ​​focus in the field of smart security.

[0003] Traditional abnormal behavior monitoring relies on manual video patrols, which has three major drawbacks: First, it is costly in terms of manpower, requiring dozens of security personnel to work in shifts in large venues, yet still struggling to cover all areas; second, it has a high false alarm and false negative rate, with a false alarm rate of 15%-20% and a false negative rate of over 15% in manual monitoring, and it is also slow to respond to brief abnormal behaviors (such as sudden falls) (with an average response time of over 5 minutes); and third, it has poor scene adaptability, failing to cope with complex environments such as dim lighting and crowd obstruction.

[0004] In recent years, with the rapid advancement of computer vision and artificial intelligence technologies, automatic identification of abnormal behavior based on deep learning models has become a research hotspot in monitoring scenarios, providing a new technical path for intelligent security. Although intelligent detection systems based on single models can partially replace manual labor, they still have significant shortcomings: First, there is a contradiction in scene adaptation. Complex locations (such as subways) require high-precision identification but are limited by the computing power of edge devices, while simple locations (such as communities) require low-power operation but suffer from resource waste due to large model redundancy. Second, the perception dimension is singular, relying heavily on video visual data. In dimly lit scenes (such as underground passages) or extremely densely populated scenes, the recognition accuracy drops sharply. Third, alarm coordination is weak. It can only output basic audio and visual alarms and cannot be linked with existing facilities in the location (such as turnstiles and broadcasts). Moreover, the alarm information lacks structured description (only indicating "abnormal" without time, location, or behavior type), resulting in low efficiency for security personnel to handle the situation. Fourth, privacy protection is insufficient. Some systems need to upload video data to the cloud for processing, posing a risk of data leakage and making them unsuitable for high-density locations with sensitive privacy, such as hospitals and schools. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides an abnormal behavior alarm device suitable for high-density crowd locations, which has the advantages of multi-scenario adaptability, high recognition accuracy, low resource consumption, and privacy security.

[0007] (II) Technical Solution

[0008] The above-mentioned technical objective of the present invention is achieved through the following technical solution: an abnormal behavior alarm device suitable for high-density crowd locations, comprising a support structure and functional modules installed on the support structure, wherein the functional modules include a multimodal perception module, a dual-strategy analysis module, an intelligent alarm linkage module, and a local data storage module; the multimodal perception module is electrically connected to the dual-strategy analysis module, the dual-strategy analysis module is electrically connected to the intelligent alarm linkage module, and both the intelligent alarm linkage module and the multimodal perception module are electrically connected to the local data storage module;

[0009] The multimodal perception module includes an intelligent audio and video acquisition unit and a millimeter-wave radar acquisition unit. The intelligent audio and video acquisition unit is used to acquire video images and audio signals in high-density crowd locations, and the millimeter-wave radar acquisition unit is used to acquire data on people's movement speed, direction, and spacing.

[0010] The dual-strategy analysis module includes a strategy switching unit, a dedicated small model combination unit, and a small model + large model collaborative unit. The strategy switching unit is used to select a detection strategy based on scene features. The dedicated small model combination unit is used for rapid anomaly identification in simple scenes. The small model + large model collaborative unit is used for high-precision anomaly determination in complex scenes.

[0011] The present invention is further configured such that: the intelligent audio and video acquisition unit includes a fisheye camera, a bullet camera, a night vision camera, and an audio sensor; the fisheye camera is installed at a high position on the supporting structure; the bullet camera is deployed at the upper and lower entrances and exits of the escalator; the night vision camera integrates an infrared fill light component; the audio sensor is installed in a one-to-one correspondence with the camera to collect abnormal audio signals such as quarreling sounds and cries for help.

[0012] The present invention is further configured such that: the millimeter-wave radar acquisition unit adopts frequency-modulated continuous wave radar, which is installed on the support structure around the camera, and is used to output point cloud data of personnel movement to assist in the identification of occlusion, falls, and rapid collisions.

[0013] The present invention is further configured such that: the strategy switching unit has a built-in scene recognition algorithm, which automatically switches the detection strategy by analyzing the crowd density, ambient light intensity and scene facility type output by the intelligent audio and video acquisition unit; when it is determined to be a simple scene, a dedicated small model combination unit is activated; when it is determined to be a complex scene, a small model + large model collaborative unit is activated.

[0014] The present invention is further configured such that: the dedicated small model combination unit includes a fall detection sub-model, an escalator reverse movement detection sub-model, and an anomaly aggregation detection sub-model;

[0015] The fall detection sub-model is based on the YOLOv5-tiny+DeepSORT algorithm. It tracks the trajectory of the head and shoulders and triggers fall posture recognition when the trajectory is lost for 5 consecutive frames. The escalator reverse movement detection sub-model is based on the HRNet lightweight skeletal keypoint algorithm. It calculates the ratio of the left and right thigh lengths for 10 consecutive frames and performs discrete Fourier transform. It determines reverse movement through frequency domain features. The abnormal clustering detection sub-model is based on the CSRNet lightweight algorithm. It regresses the crowd density heatmap and triggers an alarm when the local density is >5 people / ㎡ and the stay exceeds 3 minutes.

[0016] The present invention is further configured such that: the small model + large model collaborative unit includes a pre-screening sub-unit and a depth determination sub-unit; the pre-screening sub-unit adopts a lightweight YOLOv5 model to analyze the video stream in real time and extract suspicious segments with a confidence level of <90%;

[0017] The depth determination subunit includes the Qwen-2.5-VL visual big model and the DeepSeek text big model. The Qwen-2.5-VL model generates a structured description of suspicious fragments, and the DeepSeek model combines a pre-set rule base to determine abnormal behavior and outputs anomaly results with a confidence level of ≥90%.

[0018] The present invention is further configured such that: the intelligent alarm linkage module includes a local audible and visual alarm subunit, a security terminal alarm subunit, and a facility linkage subunit;

[0019] The local audio-visual alarm subunit includes a high-decibel audio-visual alarm and a directional voice broadcast;

[0020] The security terminal alarm subunit includes a handheld tablet and a monitoring screen. The handheld tablet receives structured alarm information including location, screenshots, and handling suggestions.

[0021] The facility linkage subunit is linked to the entrance and exit gates, fire protection system, and public address system via RS485 / Ethernet protocol.

[0022] The present invention is further configured such that the linkage logic of the facility linkage subunit includes: closing the entrance gate of the corresponding area and opening the evacuation passage when abnormal gathering occurs; starting the fire smoke exhaust fan when fire association is abnormal; and pushing alarm information with video and location to the regional security command center when there is a major abnormality.

[0023] The present invention is further configured such that: the local data storage module includes an industrial-grade storage server and a privacy protection subunit; the storage server hard drive supports RAID5 backup and stores audio and video data and abnormal logs; the privacy protection subunit includes local data transmission, face blurring processing, and three-level access control, all data is not uploaded to the cloud, the face area is Gaussian blurred, and administrators, security personnel, and maintenance personnel access data according to their permissions.

[0024] The present invention is further configured such that: the support structure includes a fixed bracket, a height adjustment rod, and an angle adjustment seat; the fixed bracket is bolted to the ground or wall of the site, the height adjustment rod is slidably disposed inside the fixed bracket, a stud is welded to the bottom of the front side of the height adjustment rod, and the front side of the stud extends to the outside of the fixed bracket and is slidably connected to the fixed bracket, and a fastening sleeve is threaded onto the surface of the stud, and the rear side of the fastening sleeve is in close contact with the fixed bracket;

[0025] The angle adjustment seat is installed on the top of the height adjustment rod and is connected to the intelligent audio and video acquisition unit and the millimeter-wave radar acquisition unit to adjust the acquisition angle.

[0026] (III) Beneficial Effects

[0027] Compared with the prior art, the present invention provides an abnormal behavior alarm device suitable for high-density crowd locations, which has the following beneficial effects:

[0028] This abnormal behavior alarm device, applicable to high-density crowd locations, innovatively adopts a dual-strategy analysis architecture. The strategy switching unit can automatically match detection strategies based on crowd density, light intensity, and scene facility type: In complex scenes (such as subways and shopping malls, with crowd density ≥ 5 people / ㎡ and facilities including escalators), a small model + large model collaborative unit is activated. The small model reduces invalid data through pre-screening, while the large model ensures anomaly recognition accuracy of ≥ 95% through deep semantic reasoning; In simple scenes (such as communities and kindergartens, with crowd density < 5 people / ㎡ and no complex facilities), a dedicated small model combination unit is activated. The lightweight model achieves millisecond-level response. By flexibly switching between the two strategies, it avoids the contradiction of traditional single-model missed detections in complex scenes and wasted resources in simple scenes, while adapting to the differentiated needs of different high-density locations. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0030] Figure 2 This is a schematic diagram of the support structure in this invention;

[0031] Figure 3 This is a schematic diagram of the abnormal behavior alarm device applicable to high-density crowd locations in this invention.

[0032] In the diagram: 1. Support structure; 2. Fixed bracket; 3. Height adjustment rod; 4. Angle adjustment seat; 5. Stud; 6. Fastening sleeve. Detailed Implementation

[0033] The technical solutions of 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.

[0034] Please see Figure 1-3 An abnormal behavior alarm device suitable for high-density crowd locations is proposed. It employs a four-layer modular architecture—sensing layer, analysis layer, alarm linkage layer, and data layer—to achieve a closed-loop process for abnormal behavior monitoring, identification, judgment, alarming, and handling. The modules work collaboratively, and the specific structure is as follows:

[0035] Overall structural composition

[0036] The device includes a support structure 1 and functional modules mounted on the support structure 1. The functional modules include a multimodal sensing module, a dual-strategy analysis module, an intelligent alarm linkage module, and a local data storage module. The multimodal sensing module and the dual-strategy analysis module are electrically connected via an industrial Ethernet, the dual-strategy analysis module is electrically connected to the intelligent alarm linkage module, and both the intelligent alarm linkage module and the multimodal sensing module are electrically connected to the local data storage module. The entire process does not rely on the cloud, ensuring real-time performance and privacy security.

[0037] Multimodal sensing module

[0038] As the sensing center of the device, it integrates audio, video, and radar data to overcome the environmental adaptability limitations of relying solely on visual recognition. Specifically, it includes an intelligent audio and video acquisition unit and a millimeter-wave radar acquisition unit.

[0039] Intelligent audio and video acquisition unit: includes fisheye camera, bullet camera, night vision camera and audio sensor; fisheye camera is installed at a high position of support structure 1 (such as the top of the column), with a monitoring radius of 10-30 meters, resolution ≥1080P, covering a large area of ​​crowd activity; bullet camera is deployed in key areas such as escalator entrances and exits, with a frame rate ≥25fps, focusing on local abnormal movements; night vision camera integrates infrared supplementary lighting components (supplementary lighting distance ≥15 meters), adapting to dim scenes with light <200 lux; audio sensor is installed one-to-one with camera, with a sampling rate of 44.1kHz, sensitivity -38dB±2dB, capturing abnormal audio above 85dB (argumentation, cries for help, glass breaking sound), and associating it with corresponding video footage to achieve synchronous audio and video analysis.

[0040] Millimeter-wave radar acquisition unit: It adopts a 77GHz frequency-modulated continuous wave radar with a detection range of 0.5-50 meters and an angular resolution of ≤1°. It is installed on the support structure 1 around the camera, avoiding metal obstructions. The radar outputs point cloud data by detecting the movement speed of people (0.1-10m / s), the direction of movement (angle with the axis of the facility), and the distance between people (0.1-5 meters). It helps to identify abnormal behaviors that are difficult to be captured visually, such as "obstruction and fall", "rapid collision" and "dense gathering". It is especially suitable for scenes with extremely dense crowds and dim lighting.

[0041] Dual-strategy analysis module

[0042] As the decision-making center of the device, it dynamically selects detection strategies based on scene characteristics, balancing accuracy and resource consumption. Specifically, it includes a strategy switching unit, a dedicated small model combination unit, and a small model + large model collaborative unit.

[0043] Strategy Switching Unit: Built-in scene recognition algorithm, which analyzes three major features output by the intelligent audio and video acquisition unit: crowd density, ambient light intensity, and scene facility type, and automatically switches detection strategies; when the crowd density is ≥5 people / ㎡, the light intensity is <200 lux, and any one of the following is present: escalator / entrance facilities, it is judged as a complex scene, and the small model + large model collaborative unit is activated; when the crowd density is <5 people / ㎡, the light intensity is ≥200 lux, and there are no complex facilities, it is judged as a simple scene, and the dedicated small model combination unit is activated, with a switching response time of <100ms.

[0044] Dedicated small model assembly unit: Designed for simple scenarios, it consists of a fall detection sub-model, an escalator reverse movement detection sub-model, and anomaly aggregation detection sub-model, deployed on an edge computing gateway to achieve millisecond-level inference.

[0045] Fall detection sub-model: Based on the YOLOv5-tiny+DeepSORT algorithm, it first detects the human head and shoulders target and tracks the trajectory. When the trajectory is lost for 5 consecutive frames (the target is occluded or falls to the ground), combined with the feature that the torso tilt angle is >45°, it is determined to be a fall.

[0046] Escalator reverse movement detection sub-model: Based on the HRNet lightweight skeleton key point algorithm, the ratio of the length of the left and right thigh bones of the human body is calculated in 10 consecutive frames. The frequency domain features are analyzed by Discrete Fourier Transform (DFT). If a specific frequency component (corresponding to the periodicity of reverse movement) is significantly enhanced, it is determined to be reverse movement.

[0047] Anomaly clustering detection sub-model: Based on the CSRNet lightweight algorithm, it regresses the population density heat map. When the density of a local area is greater than 5 people / ㎡ and the stay exceeds 3 minutes, it is judged as an abnormal clustering.

[0048] Small model + large model collaborative unit: Designed for complex scenarios, deployed on a local server, including a pre-screening subunit and a depth-based decision-making subunit:

[0049] Pre-screening subunit: Employs a lightweight YOLOv5 model to analyze the video stream in real time, extracting only suspicious segments with a confidence level of <90% (such as blurred motion or occluded targets), compressing the resolution to 720P, and reducing the frequency of large model calls;

[0050] The deep judgment subunit includes the Qwen-2.5-VL visual big model and the DeepSeek text big model. The Qwen-2.5-VL model extracts the spatiotemporal features of suspicious segments and generates structured natural language descriptions (e.g., 2024-10-01 14:30, subway platform A, 3 people in physical conflict, 1 person falls to the ground). The DeepSeek model combines a pre-set rule base (e.g., physical conflict definition: punching, pushing lasting ≥2 seconds), matches the semantic descriptions and outputs confidence scores. ≥90% of these are judged as valid anomalies.

[0051] Intelligent alarm linkage module

[0052] As the execution center of the device, it realizes multi-dimensional alarm and site facility linkage, specifically including local audible and visual alarm subunit, security terminal alarm subunit, and facility linkage subunit:

[0053] Local audio-visual alarm subunit: includes a high-decibel audio-visual alarm (80-120 decibels, red LED flashing light, flashing frequency 2Hz) and directional voice broadcast (coverage radius 5-10 meters); when an alarm is triggered, the audio-visual alarm is activated around the abnormal area, and the voice broadcast plays customized prompts according to the type of abnormality (such as "The escalator is going in reverse, please remain calm") to avoid panic among the crowd.

[0054] Security terminal alarm sub-unit: includes handheld tablet and security post monitoring screen; handheld tablet receives structured alarm work orders containing "abnormal time, location, screenshot, and handling suggestions", supports security personnel to report handling progress (such as arrival at the scene, risk elimination); monitoring screen displays the abnormality level of each area in the form of heat map (red = high risk, yellow = medium risk, green = normal), and clicking on the abnormal area can view real-time video and details.

[0055] Facility linkage subunit: Links with existing facilities (entrance and exit gates, fire protection system, public address system) in high-density locations via RS485 / Ethernet protocol; in case of abnormal gathering, closes the entrance gates of the corresponding area and opens adjacent evacuation routes; in case of fire-related anomalies (such as people running in panic), the linkage fire protection system starts the smoke exhaust fan; in case of major anomalies (such as knife fights), automatically pushes alarm information to the regional security command center, along with real-time on-site video and GPS location coordinates.

[0056] Local data storage module

[0057] As the data hub of the device, it balances data traceability and privacy protection, specifically including an industrial-grade storage server and a privacy protection sub-unit:

[0058] Industrial-grade storage server: Hard drive capacity ≥16TB, supports RAID5 redundancy backup, storage content includes: raw audio and video data (stored in separate directories by region, retained for 3-7 days), abnormal behavior logs (including alarm time, location, behavior type, confidence level, and handling records), system configuration parameters (policy switching threshold, model parameters); supports fast retrieval by time, region, and behavior type (response time <3 seconds), providing a basis for post-event review.

[0059] Privacy protection sub-unit: includes data localization processing, image desensitization, and three-level access control; all data is transmitted only between local devices (edge ​​gateway, server) and is not uploaded to the cloud; Gaussian blurring is applied to the facial areas of non-abnormally associated persons in the video (preserving the human body outline without affecting behavior recognition); three-level access is set for administrators (who can view complete data), security personnel (who can only view alarm-related segments), and maintenance personnel (who have no data access access), which complies with the requirements of the Personal Information Protection Law.

[0060] Support structure 1: Used to fix various functional modules and adapt to different installation requirements, including fixed bracket 2, height adjustment rod 3, and angle adjustment seat 4; fixed bracket 2 is bolted to the ground or wall of the location and is made of 304 stainless steel to ensure stability; height adjustment rod 3 is slidably installed inside fixed bracket 2, with an adjustment range of 2-5 meters. A stud 5 is welded to the bottom front side, and the stud 5 extends to the outside of fixed bracket 2 and is slidably connected. A fastening sleeve 6 is threaded on the surface of stud 5, and the rear side of the fastening sleeve 6 is in close contact with fixed bracket 2 to lock the height; angle adjustment seat 4 is installed on the top of height adjustment rod 3 and is connected to intelligent audio and video acquisition unit and millimeter-wave radar acquisition unit to ensure that the acquisition module is accurately aligned with the monitoring area.

[0061] The working principle of this embodiment is as follows: Fisheye cameras in the platform area capture 1080P video streams in real time, observing the overall distribution and movement trends of people on the platform; bullet cameras in the escalator area focus on people's movements as they go up and down the escalators, outputting continuous frame images; night vision cameras in the passageway area, due to sufficient lighting, temporarily disable infrared illumination and only capture regular video; audio sensors simultaneously collect on-site sounds, filtering out normal sounds such as subway arrival announcements (60-70 decibels). When an argument (90 decibels) is detected in the middle of the platform, it is automatically associated with the corresponding area's fisheye camera video and marked as... "Audio segment to be analyzed"; The platform area radar outputs point cloud data at a frequency of 5Hz, detecting that the distance between people in the middle of the platform is <0.8m (clustering characteristic), and that two people in the escalator area are moving in the opposite direction to the escalator's running direction (suspected reverse movement); All data are named according to "area-timestamp" (e.g., "Platform A-202410010745.mp4" "Escalator 1-202410010745.pcd"), and transmitted to the dual-strategy analysis module via industrial Ethernet, while being backed up to the local data storage module's temporary cache area;

[0062] The strategy switching unit receives audio and video data (10-second clip) and radar data. Through a lightweight density estimation algorithm, it calculates the crowd density in the platform area to be 8.5 people / m². The light sensor reads the ambient light level as 350 lux. The facility recognition model detects complex facilities such as escalators and turnstiles. Because it meets the criteria of "crowd density ≥ 5 people / m² + containing complex facilities", it is determined to be a complex scene. The strategy switching unit sends a start command to the "small model + large model collaborative unit" and shuts down the dedicated small model combination unit. The entire switching process takes less than 100ms and there is no data interruption.

[0063] A lightweight YOLOv5 model is used for real-time analysis of video streams from the platform and escalator areas: In the platform video, three people in the middle are identified with rapid limb movements (movement change rate > 60% / s), with a human confidence level of 85% (blurred due to crowd occlusion). This 3-second segment (07:45:02-07:45:05) is extracted and compressed to 720P resolution. In the escalator video, two people are identified moving in the opposite direction to the escalator, with a human confidence level of 88%. The corresponding 2-second segment is extracted. The two suspicious segments, along with audio and radar data, are transmitted to the depth determination subunit.

[0064] Qwen-2.5-VL visual large model processing: For suspicious segments in the platform area, extract the trajectory of key points of human skeletons and generate a structured description: "2024-10-01 07:45:03, in the middle of platform A of the subway station, 3 people are seen punching and pushing, 5 people are watching, and there is no obvious tendency to leave"; For suspicious segments in the escalator area, generate the description: "2024-10-01 07:45:04, escalator No. 1 of the subway station, 2 people are moving from bottom to top, opposite to the direction of the escalator (up and down), for 2 seconds";

[0065] DeepSeek text large model processing: Call the pre-set rule library ("Physical conflict definition: punching and pushing actions lasting ≥2 seconds, accompanied by onlookers; Escalator reverse movement definition: personnel moving in the opposite direction to the escalator running direction, lasting ≥2 frames"), and match the semantic description with the rules; the platform area segment matches the "physical conflict rule" with a confidence level of 96%; the escalator area segment matches the "escalator reverse movement rule" with a confidence level of 98%. Both are ≥90%, and are judged as "valid anomalies". The judgment result containing behavior type, time and location is output and transmitted to the intelligent alarm linkage module.

[0066] The alarm information is processed in a structured manner to generate a standardized alarm work order, which includes: basic information: anomaly 1 (physical conflict, 07:45:03, middle of platform A), anomaly 2 (escalator going in reverse, 07:45:04, escalator No. 1); auxiliary information: key frame screenshots of the two anomalies (the moment of punching during the physical conflict, and a side view of the person going in reverse on the escalator), real-time video links, and handling suggestions ("physical conflict: prioritize isolating the conflicting persons to avoid gathering; escalator going in reverse: immediately stop the escalator and guide the personnel to evacuate"); the work order is pushed to each alarm sub-unit simultaneously;

[0067] Multi-dimensional alarm triggering: Local audio-visual alarms: The audio-visual alarms in the middle of Platform A and next to Escalator No. 1 were activated, with red LED lights flashing at 2Hz and a simultaneous "beep" alarm sound; Directional voice broadcasts announced: "A conflict has occurred in the middle of Platform A. Please keep your distance. Security personnel are arriving soon." and "Escalator No. 1 is going in the wrong direction. Please stay away from the escalator. The escalator will stop immediately." The coverage radius was 10 meters, and no overall panic was caused on the platform; Security terminal alarms: Three nearby security personnel received work orders on their handheld tablets, clicked on the video link to view the real-time scene, one security personnel replied "I have gone to the middle of the platform," another replied "I have gone to Escalator No. 1," and one remained at the guard booth monitoring screen; The guard booth monitoring screen marked the two abnormal areas with red heat maps, displaying the platform and escalator images in a split screen, allowing the administrator to view the handling progress in real time;

[0068] Facility linkage: When the escalator control system receives the "emergency stop" command, escalator No. 1 stops operating within 3 seconds, and the guide lights at the top and bottom of the escalator turn red; the gate at the middle entrance of platform A automatically closes, and the gates of the adjacent evacuation passage open to prevent more people from rushing in; since the physical conflict has not escalated to a major anomaly, it will not be pushed to the command center for the time being; if it is not dealt with within 5 minutes, the system will automatically escalate the alarm.

[0069] Local data archiving and privacy protection (continuous archiving)

[0070] Data classification and storage:

[0071] Raw data: Audio and video data are stored in separate directories, "20241001-Platform A" and "20241001-Escalator 1", and retained for 7 days; radar point cloud data is retained for 1 day; Business data: Anomaly judgment logs (including confidence level and analysis process), alarm work orders (including security handling records "07:48:00 the conflict has been quelled, 07:47:30 escalator personnel have been evacuated"), and facility linkage status ("07:45:07 the escalator has stopped, 07:45:08 the turnstile has been closed") are stored on an industrial-grade server; Privacy protection processing:

[0072] The faces of onlookers in the video are Gaussian blurred, retaining only the clear outlines of those involved in the conflict and those going the wrong way (for post-incident tracing); security personnel can only view the 3-minute segment associated with the alarm and cannot access the complete platform video; maintenance personnel can only view the server operation logs and have no data access permissions; cache clearing: after 24 hours, the system automatically clears the temporary cache data, freeing up 10% of storage space to ensure the device continues to operate stably.

[0073] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. Those skilled in the art can make modifications to this embodiment without contributing any inventive step after reading this specification. Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An abnormal behavior alarm device suitable for high-density crowd locations, comprising a support structure (1) and functional modules installed on the support structure (1), characterized in that: The functional modules include a multimodal perception module, a dual-strategy analysis module, an intelligent alarm linkage module, and a local data storage module; the multimodal perception module is electrically connected to the dual-strategy analysis module, the dual-strategy analysis module is electrically connected to the intelligent alarm linkage module, and both the intelligent alarm linkage module and the multimodal perception module are electrically connected to the local data storage module. The multimodal perception module includes an intelligent audio and video acquisition unit and a millimeter-wave radar acquisition unit. The intelligent audio and video acquisition unit is used to acquire video images and audio signals in high-density crowd locations, and the millimeter-wave radar acquisition unit is used to acquire data on people's movement speed, direction, and spacing. The dual-strategy analysis module includes a strategy switching unit, a dedicated small model combination unit, and a small model + large model collaborative unit. The strategy switching unit is used to select a detection strategy based on scene features. The dedicated small model combination unit is used for rapid anomaly identification in simple scenes. The small model + large model collaborative unit is used for high-precision anomaly determination in complex scenes.

2. The abnormal behavior alarm device suitable for high-density crowd locations according to claim 1, characterized in that: The intelligent audio and video acquisition unit includes a fisheye camera, a bullet camera, a night vision camera, and an audio sensor; the fisheye camera is installed at a high position on the support structure (1); the bullet camera is deployed at the top and bottom of the escalator and in the entrance and exit areas; the night vision camera integrates an infrared fill light component; the audio sensor is installed one-to-one with the camera to collect abnormal audio signals such as quarrels and cries for help.

3. The abnormal behavior alarm device suitable for high-density crowd locations according to claim 1, characterized in that: The millimeter-wave radar acquisition unit adopts frequency-modulated continuous wave radar and is installed on the support structure (1) around the camera. It is used to output point cloud data of personnel movement to assist in the identification of occlusion, falls, and rapid collision behaviors.

4. The abnormal behavior alarm device suitable for high-density crowd locations according to claim 1, characterized in that: The strategy switching unit has a built-in scene recognition algorithm. By analyzing the crowd density, ambient light intensity, and scene facility type output by the intelligent audio and video acquisition unit, it automatically switches the detection strategy. When the scene is determined to be simple, a dedicated small model combination unit is activated. When the scene is determined to be complex, a small model + large model collaborative unit is activated.

5. The abnormal behavior alarm device suitable for high-density crowd locations according to claim 1, characterized in that: The dedicated small model combination unit includes a fall detection sub-model, an escalator reverse movement detection sub-model, and an anomaly clustering detection sub-model; The fall detection sub-model is based on the YOLOv5-tiny+DeepSORT algorithm. It tracks the trajectory of the head and shoulders and triggers fall posture recognition when the trajectory is lost for 5 consecutive frames. The escalator reverse movement detection sub-model is based on the HRNet lightweight skeletal keypoint algorithm. It calculates the ratio of the left and right thigh lengths for 10 consecutive frames and performs discrete Fourier transform. It determines reverse movement through frequency domain features. The abnormal clustering detection sub-model is based on the CSRNet lightweight algorithm. It regresses the crowd density heatmap and triggers an alarm when the local density is >5 people / ㎡ and the stay exceeds 3 minutes.

6. The abnormal behavior alarm device suitable for high-density crowd locations according to claim 1, characterized in that: The small model + large model collaborative unit includes a pre-screening subunit and a depth judgment subunit; the pre-screening subunit adopts a lightweight YOLOv5 model to analyze the video stream in real time and extract suspicious segments with a confidence level of <90%; The depth determination subunit includes the Qwen-2.5-VL visual big model and the DeepSeek text big model. The Qwen-2.5-VL model generates a structured description of suspicious fragments, and the DeepSeek model combines a pre-set rule base to determine abnormal behavior and outputs anomaly results with a confidence level of ≥90%.

7. The abnormal behavior alarm device suitable for high-density crowd locations according to claim 1, characterized in that: The intelligent alarm linkage module includes a local audible and visual alarm subunit, a security terminal alarm subunit, and a facility linkage subunit; The local audio-visual alarm subunit includes a high-decibel audio-visual alarm and a directional voice broadcast; The security terminal alarm subunit includes a handheld tablet and a monitoring screen. The handheld tablet receives structured alarm information including location, screenshots, and handling suggestions. The facility linkage subunit is linked to the entrance and exit gates, fire protection system, and public address system via RS485 / Ethernet protocol.

8. An abnormal behavior alarm device suitable for high-density crowd locations according to claim 7, characterized in that: The linkage logic of the facility linkage subunit includes: closing the entrance gate of the corresponding area and opening the evacuation passage when there is an abnormal gathering; starting the fire exhaust fan when there is an abnormal fire association; and pushing alarm information with video and location to the regional security command center when there is a major abnormality.

9. An abnormal behavior alarm device suitable for high-density crowd locations according to claim 1, characterized in that: The local data storage module includes an industrial-grade storage server and a privacy protection subunit; the storage server hard drive supports RAID5 backup and stores audio and video data and abnormal logs; the privacy protection subunit includes local data transmission, face blurring processing, and three-level access control, ensuring that all data is not uploaded to the cloud, face areas are Gaussian blurred, and administrators, security personnel, and maintenance personnel access data according to their permissions.

10. An abnormal behavior alarm device suitable for high-density crowd locations according to claim 1, characterized in that: The support structure (1) includes a fixed bracket (2), a height adjustment rod (3), and an angle adjustment seat (4); the fixed bracket (2) is bolted to the ground or wall of the site, the height adjustment rod (3) is slidably disposed inside the fixed bracket (2), a stud (5) is welded to the bottom of the front side of the height adjustment rod (3), and the front side of the stud (5) extends to the outside of the fixed bracket (2) and is slidably connected to the fixed bracket (2), and a fastening sleeve (6) is threaded onto the surface of the stud (5), and the rear side of the fastening sleeve (6) is in close contact with the fixed bracket (2); The angle adjustment seat (4) is installed on the top of the height adjustment rod (3) and connected to the intelligent audio and video acquisition unit and the millimeter-wave radar acquisition unit for adjusting the acquisition angle.

Images