People flow density monitoring device installed on barrier and control method thereof

Abstract

The invention relates to a people flow density monitoring device installed on a barrier and a control method thereof.The device comprises a shell (1) installed on the side face of the upper portion of the barrier (2), a ToF depth sensor (3), an inertial measurement unit (4) and a micro-control unit (5) are arranged in the shell (1), and the micro-control unit (5) is connected with the ToF depth sensor (3) and the inertial measurement unit (4); the ToF depth sensor (3) is used for measuring the real-time distance from a crowd to the edge of the barrier (2); the inertial measurement unit (4) is used for detecting the posture state change of the barrier (2) in real time; and the micro-control unit (5) is used for fusing and processing the real-time distance from the crowd to the edge of the barrier (2) and the posture state of the barrier (2) to obtain a crowd density estimated value. Compared with the prior art, the method has outstanding social safety benefits, the accuracy of people flow density monitoring is remarkably improved, and the false alarm rate and the missing report rate are effectively reduced.

Description

Technical Field

[0001] This invention relates to the fields of public safety management and crowd density monitoring technology, and in particular to a crowd density monitoring device installed on a barrier and its control method. Background Technology

[0002] With the increasing density of people at large public events, exhibitions, and rail transit stations, timely and accurate monitoring of crowds and prevention of stampedes has become a crucial issue for urban public safety management. Currently, widely used crowd monitoring technologies mainly include video image analysis, infrared counters, ultrasonic sensors, and top-mounted Time-of-Flight (ToF) sensors.

[0003] Video surveillance technology uses cameras installed at top or high locations to acquire real-time images and perform image analysis to monitor crowd density and identify abnormal behavior. However, this method is sensitive to ambient light, poses significant privacy risks, and has a high false recognition rate in high-density, frequently obstructed scenes. Furthermore, video surveillance solutions typically require high computing power and bandwidth, resulting in high costs and maintenance expenses, making widespread deployment difficult for large-scale or temporary events.

[0004] In recent years, top-mounted Time-of-Flight (ToF) sensors have been widely used due to their high ranging accuracy and strong privacy protection. For example, patent application CN110717408A discloses a people counting method based on a ToF camera. This method uses a combination of detection and tracking to accurately monitor and analyze people flow in a space in real time by installing a ToF camera on the top. However, top-mounting has stringent requirements on installation height and structural conditions. In temporary deployment scenarios, special brackets are often required, resulting in high costs and complex installation. More importantly, once the top sensor is obstructed, the monitoring blind zone increases significantly, and the accuracy of people flow estimation drops rapidly. At the same time, top sensors cannot monitor the status of physical obstacles (such as barriers used to guide crowds), and it is difficult to provide timely warnings and responses if barriers are accidentally moved or tilted.

[0005] To address the aforementioned shortcomings, some existing technologies propose using an inertial measurement unit (IMU) mounted separately on the barrier to sense its displacement or tilting status. However, the IMU itself can only monitor changes in the barrier's own attitude or acceleration, lacking effective perception of crowd density and failing to achieve comprehensive, coordinated monitoring of the crowd's status around the barrier. Therefore, the IMU solution cannot fully meet the needs of crowd management and synchronized barrier status monitoring in public safety scenarios.

[0006] Therefore, there is an urgent need for a sensing solution that is more flexible in installation, has a more comprehensive sensing range, is lower in cost, and can simultaneously monitor the status of barriers and the density of people, in order to solve the shortcomings and defects of the existing technologies. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art by providing a crowd density monitoring device and its control method that can be installed on a barrier, which has outstanding social safety benefits, significantly improves the accuracy of crowd density monitoring, and effectively reduces the false alarm and missed alarm rates.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] A crowd density monitoring device installed on a barrier includes a housing mounted on the upper side of the barrier. The housing houses a Time-of-Flight (ToF) depth sensor, an inertial measurement unit (IMU), and a microcontroller unit (MCU). The MCU is connected to the ToF depth sensor and the IMU. The ToF depth sensor measures the real-time distance from the crowd to the edge of the barrier. The IMU detects changes in the barrier's attitude in real time. The MCU fuses the real-time distance from the crowd to the barrier's edge and the barrier's attitude to obtain an estimated crowd density.

[0010] Furthermore, the outer shell is symmetrically provided with annular buckles on both sides, and the outer shell is fixed to the barrier by the annular buckles.

[0011] Furthermore, the housing includes a shell and a side cover, which are connected by snap-fit. A ToF window is provided on the side cover, and the ToF window is coaxially arranged with the ToF depth sensor.

[0012] Furthermore, the attitude state of the barrier includes various attitude states such as tilting, moving, and vibrating.

[0013] Furthermore, the microcontroller unit includes a processor and a memory. The processor is used to fuse the real-time distance between the crowd and the edge of the barrier and the posture state of the barrier to obtain an estimated value of the crowd density. The memory is used to store the raw data of the real-time distance between the crowd and the edge of the barrier and the posture state of the barrier, as well as the estimated value of the crowd density, in an independent storage unit.

[0014] Furthermore, the processor is equipped with a pre-trained LSTM model, which fuses the real-time distance between the crowd and the edge of the barrier and the posture state of the barrier to obtain an estimated value of the crowd density.

[0015] Furthermore, the specific steps for fusing the real-time distance from the crowd to the edge of the barrier and the pose state of the barrier using a pre-trained LSTM model are as follows:

[0016] The Time-of-Flight (ToF) feature sequence of the real-time distance from the crowd to the edge of the barrier and the IMU (Inertial Measurement Unit) feature sequence of the barrier's posture state are respectively input into independent feature precoding networks to obtain the ToF feature embedding vector and the IMU feature embedding vector.

[0017] Based on the ToF feature embedding vector and the IMU feature embedding vector, the feature information of the IMU mode is incorporated into the ToF mode representation through a multi-head attention mechanism to obtain cross-modal features;

[0018] The cross-modal features are added element-wise to the ToF feature embedding vector, and then layer normalization is performed to obtain the fused feature representation.

[0019] Furthermore, the housing is also equipped with a wireless communication module connected to the microcontroller unit. When the barrier's posture is in an abnormal state or the estimated pedestrian density exceeds a preset threshold, the wireless communication module immediately sends an alarm message to the background management system.

[0020] Furthermore, the abnormal posture states include various posture states such as tipping over and violent shaking.

[0021] According to another aspect of the present invention, a control method for a pedestrian density monitoring device installed on a barrier is provided, which monitors pedestrian density by using the pedestrian density monitoring device installed on the barrier as described above, and includes the following steps:

[0022] The inertial measurement unit monitors the changes in the attitude of the barrier in real time and performs attitude detection. If an abnormal attitude is detected, an alarm message is sent to the back-end management system via the wireless communication module.

[0023] The distance data from the crowd to the edge of the barrier is collected in real time using a ToF depth sensor;

[0024] The posture of the barrier and the distance between the crowd and the edge of the barrier are compressed and sent back to the microcontroller unit for fusion processing to obtain the estimated value of the crowd density.

[0025] The estimated population density is determined based on a preset threshold. If the estimated population density exceeds the preset threshold, an alarm message is sent to the backend management system via a wireless communication module.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. This invention couples an inertial measurement unit and a side-mounted ToF depth sensor within a crowd density monitoring device. A ToF window is provided on the side cover of the outer casing, and the ToF window and the ToF depth sensor are coaxially arranged. By sharing a common plate structure for the inertial measurement unit and the side-mounted ToF depth sensor, and designing the ToF depth sensor and the ToF window to be coaxial, the two can complete complementary sensing within a small volume, significantly improving the accuracy of crowd density monitoring.

[0028] 2. This invention monitors the attitude changes of the barrier in real time through an inertial measurement unit and performs attitude state detection. It also collects the distance data from the crowd to the edge of the barrier in real time through a ToF depth sensor and fuses the barrier attitude changes with the distance data from the crowd to the edge of the barrier through an LSTM model. This allows it to maintain stable output in high-density crowd impact scenarios and effectively reduce false alarm and false alarm rates.

[0029] 3. By simultaneously reporting the abnormal posture of barriers and the threshold event of pedestrian flow in the alarm information, this invention reduces the risks of network congestion and interface compatibility caused by the parallel operation of multiple systems. It can achieve "plug and play" with existing security platforms, significantly shortening the integration cycle. In scenarios such as public activities, exhibition passages, and emergency evacuation passages, it can achieve early detection and rapid response to the coupling risks of "pedestrian flow density-barrier posture", which has outstanding social security benefits. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of a pedestrian density monitoring device installed on a barrier according to the present invention.

[0031] Figure 2 This is a schematic diagram of the outer casing of a pedestrian density monitoring device installed on a barrier, as proposed in this invention.

[0032] Figure 3 This is a schematic diagram of the module connection of a pedestrian density monitoring device installed on a barrier according to the present invention;

[0033] Figure 4 This is a schematic diagram illustrating an application scenario of a pedestrian density monitoring device installed on a barrier, as proposed in this invention.

[0034] Figure 5 This is a schematic diagram illustrating the steps of a control method for a pedestrian density monitoring device installed on a barrier, as proposed in this invention.

[0035] Figure 6 This is a schematic flowchart illustrating the control method for a pedestrian density monitoring device installed on a barrier, as proposed in this invention.

[0036] Legend: 1. Outer shell; 101. Ring buckle; 102. Shell; 103. Side cover; 2. Barrier; 3. ToF depth sensor; 301. ToF window; 4. Inertial measurement unit; 5. Microcontroller unit; 501. Processor; 502. Memory; 6. Wireless communication module; 7. Independent storage unit; 8. Background management system. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0038] The following English abbreviations are involved:

[0039] Inertial Measurement Unit (IMU)

[0040] Time of Flight (ToF)

[0041] Microcontroller Unit (MCU)

[0042] Example 1

[0043] This embodiment provides a pedestrian density monitoring device installed on a barrier, such as... Figure 1 As shown, it includes a housing 1 mounted on the upper side of the barrier 2. Figure 2 As shown, the outer shell 1 has symmetrically arranged annular buckles 101 on both sides, and the outer shell 1 is fixed to the barrier 2 by the annular buckles 101. The outer shell 1 includes a housing 102 and a side cover 103, which are connected by buckles. A ToF window 301 is provided on the side cover 103, and the ToF window 301 is coaxially arranged with the ToF depth sensor 3.

[0044] The housing 1 houses a ToF depth sensor 3, an inertial measurement unit 4, a microcontroller unit 5, a wireless communication module 6, and an independent storage unit 7. The connections between the modules within the housing 1 are as follows: Figure 3 As shown, the microcontroller unit 5 is connected to the ToF depth sensor 3 and the inertial measurement unit 4. The ToF depth sensor 3 is used to measure the real-time distance from the crowd to the edge of the barrier 2; the inertial measurement unit 4 is used to detect the changes in the attitude state of the barrier 2 in real time, including various attitude states such as tilting, moving and vibrating; the microcontroller unit 5 is used to fuse and process the real-time distance from the crowd to the edge of the barrier 2 and the attitude state of the barrier 2 to obtain an estimated value of the crowd density.

[0045] The inertial measurement unit 4 uses a CNN model to detect the attitude changes of the barrier 2 in real time. This CNN model is trained on a large amount of barrier attitude data and can accurately identify key features such as the tilt angle, movement speed and vibration frequency of the barrier, thereby providing high-precision attitude information for the microcontroller unit 5.

[0046] The microcontroller unit 5 includes a processor 501 and a memory 502. The processor 501 is used to fuse the real-time distance between the crowd and the edge of the barrier 2 and the attitude state of the barrier 2 to obtain an estimated value of the crowd density. The memory 502 is used to store the raw data of the real-time distance between the crowd and the edge of the barrier 2 and the attitude state of the barrier 2, along with the estimated value of the crowd density, in an independent storage unit 7. The processor 501 is equipped with a pre-trained LSTM model. The pre-trained LSTM model fuses the real-time distance between the crowd and the edge of the barrier 2 and the attitude state of the barrier 2 to obtain an estimated value of the crowd density. The LSTM model can effectively process time-series data. Combined with the dynamic information provided by the ToF depth sensor 3 and the inertial measurement unit 4, it can more accurately reflect the changing trend of crowd density, and maintain high monitoring accuracy even when the crowd is moving rapidly or the barrier attitude changes frequently.

[0047] The specific steps for obtaining the crowd density estimate by fusing the real-time distance from the crowd to the edge of barrier 2 and the pose state of barrier 2 using a pre-trained LSTM model are as follows:

[0048] The real-time distance from the crowd to the edge of barrier 2 and the posture state of barrier 2 are preprocessed. Data preprocessing includes data cleaning, normalization, and time series construction. Data cleaning is used to remove outliers and noisy data; normalization scales the data to the same range to facilitate model processing; time series construction arranges the data in chronological order to form a time-series data format suitable for LSTM model processing.

[0049] The real-time distance between the crowd and the edge of barrier 2, after data preprocessing, and the posture state of barrier 2 are input into a pre-trained LSTM model in processor 501 for temporal analysis. The LSTM model consists of an input layer, an LSTM layer, a fully connected layer, and an output layer connected in sequence. The input layer receives the preprocessed real-time distance data and the temporal data of the posture state. The LSTM layer performs temporal analysis on the temporal data, capturing the time dependencies and dynamic changes in the data. The LSTM layer can effectively handle long-short-term dependencies. The fully connected layer extracts key features related to crowd density from the temporal data, further integrating and refining the features extracted by the LSTM layer to provide a more accurate feature representation for the output layer. The output layer outputs the estimated crowd density value. Based on the features extracted by the fully connected layer, the estimated crowd density value is finally generated through activation functions and other processing.

[0050] The specific implementation of obtaining the crowd density estimate by fusing the real-time distance from the crowd to the edge of barrier 2 and the pose state of barrier 2 using a pre-trained LSTM model is as follows:

[0051] Initialization Phase: Load the pre-trained LSTM model that has been switched to eval() mode. Initialize a fixed-length circular buffer (queue) to store features from the latest seq_len (sequence length) time steps. The length of the circular buffer should be consistent with the length of the input sequence of the LSTM model. Each element of the buffer can store feature data from the ToF sensor and inertial measurement unit. Prepare the output pipeline, including: logging, visualization, and alarm interfaces. Configure the logging system to record key information, such as data acquisition time, model inference results, and anomalies. Prepare the visualization interface to display sensor data and model output results in real time. Configure the alarm interface to trigger alarms promptly when the model detects anomalies.

[0052] At each time step (e.g., every second or when the sensor triggers), the latest frame of data is read from the ToF sensor and the inertial measurement unit, and the data is cleaned.

[0053] The latest processed ToF feature sequence and IMU feature sequence are appended to the buffer, and the oldest record is automatically discarded.

[0054] The acquired ToF feature sequences and IMU feature sequences are input into independent feature precoding networks, mapping features from different modalities to embedding spaces of the same dimension. For the raw features of the ToF modality at each time step, the input is fed into a ToF feature precoder (such as a multilayer perceptron (MLP) or a 1D convolutional network), outputting a low-dimensional ToF feature embedding vector. The features of the IMU modality are input into an IMU feature precoder to obtain a low-dimensional IMU feature embedding vector.

[0055] The Time-of-Flight (ToF) feature embedding vectors at all time points are used as queries, and the IMU feature embedding vectors are used as keys and values, respectively, and are input into the multi-head attention module. Attention weights are calculated based on the queries and keys, and the feature information of the IMU modality is integrated into the ToF modality representation according to the attention weights, realizing cross-modal feature interaction at the temporal level and obtaining cross-modal features.

[0056] The cross-modal features output by multi-head attention are added element-wise to the original ToF feature embedding vector, preserving the information of the original ToF features, while adding IMU feature information extracted through the multi-head attention mechanism. After layer normalization, the fused feature representation is obtained.

[0057] The fused feature representation is input into an LSTM classifier to obtain the predicted value of the density interval at the current time. The maximum value is then taken to obtain the class code. The class code is converted into a "sliding interval string" (such as "3-5") using a LabelEncoder or mapping table. The data in the "sliding interval string" corresponds to the number of people in the path at the current time (such as "3-5" representing 3-5 people in the path at the current time), thus obtaining the estimated value of the pedestrian density.

[0058] In another preferred embodiment, a prediction smoothing and alarm mechanism is also included.

[0059] The moving average is calculated by summing the midpoint values ​​of predicted categories over a period of time (e.g., the most recent n times). This effectively smooths the data curve, reducing drastic fluctuations caused by random noise or short-term outliers, and making the data trend clearer and more stable. During continuous monitoring of the moving average, a specific threshold range is set, determined based on the actual application scenario and target requirements. If the moving average crosses the boundary of this threshold range—that is, rising from within the range to outside or falling from outside to within—it indicates a significant change in the overall data trend. At this point, an alarm is triggered based on the direction of the moving average change. If the moving average rises and exceeds the upper limit of the threshold range, an "increased density" alarm is triggered; conversely, if the moving average falls and falls below the lower limit of the threshold range, a "decreased density" alarm is triggered. This alarm mechanism, based on a combination of the moving average and the threshold range, can promptly detect and alert users to significant changes in data density while maintaining data smoothness, providing a strong basis for subsequent decision-making and intervention.

[0060] The wireless communication module 6 is connected to the microcontroller unit 5. When the barrier 2 is in an abnormal posture or the estimated pedestrian density exceeds a preset threshold, the wireless communication module 6 immediately sends an alarm message to the background management system 8. The abnormal posture of the barrier 2 includes various postures such as tipping over and violent vibration.

[0061] The microcontroller unit 5 compares the estimated crowd density output from the LSTM model with a preset crowd density threshold. When the estimated crowd density exceeds the preset threshold, it is determined to be an abnormal crowd density. The preset crowd density threshold can be set according to the actual application scenario and safety requirements. For example, in densely populated places, the crowd density threshold can be set to 5 people per square meter; in critical areas such as evacuation routes, the crowd density threshold can be set to 3 people per square meter.

[0062] Alarm information includes the type of abnormal event (such as "barrier overturned", "excessive crowd density", etc.), the time of occurrence, the specific location, and related data (such as real-time distance data, posture status data, and crowd density estimates).

[0063] The wireless communication module 6 transmits alarm information to the back-end management system 8 in real time via a wireless network (such as Wi-Fi, 4G / 5G, etc.). After receiving the alarm information, the back-end management system 8 immediately issues an alarm prompt and notifies the management personnel to take corresponding emergency measures, such as evacuating people and checking the status of barriers, to ensure the safety of personnel and the normal operation of equipment.

[0064] The application scenario diagram of the crowd density monitoring device installed on the barrier provided in this embodiment is shown below. Figure 4 As shown, since the walkway is composed of multiple barriers 2 arranged continuously on both sides, pedestrian density monitoring devices can be installed on each barrier 2 to monitor the pedestrian density in real time, forming a monitoring network covering the entire walkway. Each monitoring device is responsible for monitoring the pedestrian density and barrier posture in the vicinity of its assigned barrier 2. By reasonably setting the spacing between the monitoring devices, it is ensured that the monitoring ranges of adjacent devices can be connected to each other, avoiding monitoring blind spots. For example, the spacing between the monitoring devices can be set to 5-10 meters depending on the width of the walkway and changes in pedestrian density.

[0065] Example 2

[0066] This embodiment provides a control method for a pedestrian density monitoring device installed on a barrier, such as... Figure 5 As shown, the method for monitoring pedestrian density using a pedestrian density monitoring device installed on a barrier as described in Example 1 includes the following steps:

[0067] S1. The inertial measurement unit 4 monitors the attitude changes of the barrier 2 in real time and performs attitude state detection. If there is an abnormal attitude state, an alarm message is sent to the background management system 8 through the wireless communication module 6.

[0068] like Figure 6 As shown, the inertial measurement unit 4 is activated to begin collecting IMU data, analyzing the attitude changes of the barrier 2, and performing a preliminary threshold check on the collected IMU data to determine if there are any anomalies. If an anomaly is found, an alarm is triggered, and an alarm message is sent to the backend management system 8 via the wireless communication module 6. The backend management system 8 receives the alarm data, records it, and processes it. If no anomaly is found, step S3 is performed.

[0069] S2. The distance data from the crowd to the edge of the barrier 2 is collected in real time by the ToF depth sensor 3.

[0070] Activate ToF depth sensor 3 to collect real-time distance data from the crowd to the edge of barrier 2.

[0071] S3. Compress the posture state of barrier 2 and the distance data from the crowd to the edge of barrier 2, and send them back to the microcontroller unit 5 for fusion processing to obtain the estimated value of the crowd density.

[0072] The attitude state of barrier 2 and the distance data from the crowd to the edge of barrier 2 are compressed. The compressed raw data is then sent back to the microcontroller unit 5 and stored in the independent storage unit 7 via memory 502. The processor 501 places the raw data into a buffer queue to await fusion processing. The raw data collected by the inertial measurement unit 4 and the ToF depth sensor 3 are synchronized in time. After time synchronization, the data is fused using a pre-trained LSTM model to obtain an estimate of the crowd density.

[0073] S4. The estimated value of the crowd density is judged according to the preset threshold. If the estimated value of the crowd density exceeds the preset threshold, an alarm message is sent to the background management system 8 through the wireless communication module 6.

[0074] The estimated population density is determined based on a preset threshold. If the estimated population density exceeds the preset threshold, an alarm message is sent to the back-end management system 8 via the wireless communication module 6, and the estimated population density is stored in the independent storage unit 7 for data archiving. If the estimated population density does not exceed the preset threshold, the estimated population density is directly stored in the independent storage unit 7 for data archiving.

[0075] The rest is the same as in Example 1.

[0076] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A pedestrian density monitoring device installed on a barrier, characterized in that, The device includes a housing (1) installed on the upper side of the barrier (2). Inside the housing (1) are a ToF depth sensor (3), an inertial measurement unit (4), and a microcontroller unit (5). The microcontroller unit (5) is connected to the ToF depth sensor (3) and the inertial measurement unit (4). The ToF depth sensor (3) is used to measure the real-time distance from the crowd to the edge of the barrier (2). The inertial measurement unit (4) is used to detect the changes in the posture of the barrier (2) in real time. The microcontroller unit (5) is used to fuse the real-time distance from the crowd to the edge of the barrier (2) and the posture of the barrier (2) to obtain an estimated value of the crowd density.

2. The pedestrian density monitoring device installed on a barrier according to claim 1, characterized in that, The outer shell (1) is symmetrically provided with ring buckles (101) on both sides, and the outer shell (1) is fixed to the barrier (2) by the ring buckles (101).

3. The pedestrian density monitoring device installed on a barrier according to claim 1, characterized in that, The outer casing (1) includes a housing (102) and a side cover (103), which are connected by a snap fastener. A ToF window (301) is provided on the side cover (103), and the ToF window (301) is coaxially arranged with the ToF depth sensor (3).

4. The pedestrian density monitoring device installed on a barrier according to claim 1, characterized in that, The attitude states of the barrier (2) include various attitude states such as tilting, moving and vibrating.

5. The pedestrian density monitoring device installed on a barrier according to claim 1, characterized in that, The microcontroller unit (5) includes a processor (501) and a memory (502). The processor (501) is used to fuse the real-time distance between the crowd and the edge of the barrier (2) and the posture state of the barrier (2) to obtain an estimated value of the crowd density. The memory (502) is used to store the raw data of the real-time distance between the crowd and the edge of the barrier (2) and the posture state of the barrier (2) and the estimated value of the crowd density in an independent storage unit (7).

6. The pedestrian density monitoring device installed on a barrier according to claim 5, characterized in that, The processor (501) is equipped with a pre-trained LSTM model. The pre-trained LSTM model is used to fuse the real-time distance between the crowd and the edge of the barrier (2) and the posture state of the barrier (2) to obtain the estimated value of the crowd density.

7. The pedestrian density monitoring device installed on a barrier according to claim 6, characterized in that, The specific steps for fusing the real-time distance from the crowd to the edge of the barrier (2) and the pose state of the barrier (2) using a pre-trained LSTM model are as follows: The ToF feature sequence of the real-time distance from the crowd to the edge of the barrier (2) and the IMU feature sequence of the posture state of the barrier (2) are respectively input into independent feature precoding networks to obtain the ToF feature embedding vector and the IMU feature embedding vector. Based on the ToF feature embedding vector and the IMU feature embedding vector, the feature information of the IMU mode is incorporated into the ToF mode representation through a multi-head attention mechanism to obtain cross-modal features; The cross-modal features are added element-wise to the ToF feature embedding vector, and then layer normalization is performed to obtain the fused feature representation.

8. The pedestrian density monitoring device installed on a barrier according to claim 1, characterized in that, The outer casing (1) is also equipped with a wireless communication module (6) connected to the microcontroller unit (5). When the posture of the barrier (2) is in an abnormal posture state or the estimated value of the crowd density exceeds the preset threshold, the wireless communication module (6) immediately sends an alarm message to the background management system (8).

9. The pedestrian density monitoring device installed on a barrier according to claim 8, characterized in that, The abnormal posture states include various posture states such as tipping over and violent shaking.

10. A control method for a pedestrian density monitoring device installed on a barrier, characterized in that, Monitoring pedestrian density using a pedestrian density monitoring device installed on a barrier as described in any one of claims 1-9 includes the following steps: The inertial measurement unit (4) monitors the attitude state changes of the barrier (2) in real time and performs attitude state detection. If there is an abnormal attitude state, an alarm message is sent to the background management system (8) through the wireless communication module (6). The distance data from the crowd to the edge of the barrier (2) is collected in real time by the ToF depth sensor (3); The posture state of the barrier (2) and the distance data from the crowd to the edge of the barrier (2) are compressed and sent back to the microcontroller (5) for fusion processing to obtain the estimated value of the crowd density. The estimated value of the crowd density is judged according to the preset threshold. If the estimated value of the crowd density exceeds the preset threshold, an alarm message is sent to the background management system (8) through the wireless communication module (6).

Images