Visitor flow dynamic identification method and system based on multi-sensor fusion

Abstract

The invention discloses a multi-sensor fusion-based people flow dynamic identification method and system, relates to the technical field of intelligent traffic, and is used for solving the problems of people flow perception distortion and shielding tracking interruption in a complex environment. The method comprises the following steps: firstly, constructing an environment interference matrix, and dynamically distributing sensor weights according to a radar electromagnetic attenuation rate, thermal imaging illumination abrupt change and a piezoelectric mechanical noise factor; when the thermal imaging density weight is too low and the radar speed weight is greater than a threshold value, triggering piezoelectric gait verification and outputting calibration fusion data; mapping the fusion data to a dynamic grid, calculating three indexes of density entropy, velocity mutation and flow direction consistency, generating a congestion vector, and marking a risk boundary; for a shielding target, generating a virtual track by combining the final known position, speed and density distribution, and optimizing model parameters when a re-identification error exceeds a limit; and finally, a hierarchical dredging instruction is judged and triggered through the virtual track and the risk area space overlapping, and the gate is dynamically adjusted based on the congestion vector to release.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and specifically to a method and system for dynamic identification of pedestrian flow based on multi-sensor fusion. Background Art

[0002] With the acceleration of urbanization and the widespread construction of large public spaces, the need for safety management in crowded areas such as subway stations, shopping malls, and stadiums is increasing. Especially during peak hours or emergencies, the risk of crowds, congestion, and even stampedes increases significantly, placing higher demands on real-time perception of crowd flow, risk prediction, and emergency response. To address this, there is an urgent need to establish a highly timely and accurate dynamic crowd flow identification mechanism that can operate stably in complex, multi-source environments and promptly output key crowd characteristic information to provide decision-making support for safe guidance and flow control.

[0003] Most current crowd identification methods rely on single sensors or static models. These methods are prone to target recognition failure, trajectory interruption, or density estimation bias when faced with complex application scenarios such as complex occlusion, illumination changes, or signal interference. Single visual or infrared technologies struggle to stably identify individuals in low-light, reflective, or heat-interference environments. Radar, while penetrating, struggles to accurately reflect density structures. Piezoelectric sensors are significantly affected by background vibrations and struggle to independently locate targets. Furthermore, traditional algorithms often employ fixed thresholds or static model parameters, lacking the ability to adapt to dynamic environments. This makes it impossible to accurately track individual trajectories when obstructed by people or when traffic flows fluctuate dramatically. It also makes it difficult to effectively predict congestion trends and adjust traffic strategies, limiting the system's practicality and responsiveness in large-scale, highly dynamic scenarios. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a method and system for dynamic identification of pedestrian flow based on multi-sensor fusion, which solves the problems of the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a dynamic identification method for human flow based on multi-sensor fusion, comprising the following steps: S1. According to the electromagnetic attenuation rate of the radar, the illumination mutation factor of the thermal imaging and the mechanical noise factor of the piezoelectric sensor, an environmental interference matrix is constructed, and the weight value of each sensor is dynamically calculated according to the inverse relationship of each factor in the matrix; S2. When the density weight value of the thermal imaging in the blocked area is lower than the first threshold and the speed weight value of the radar corresponding area is higher than the second threshold, the gait spectrum characteristics of the piezoelectric sensor are called to verify the existence of the target, and the fused density, speed and flow direction data after weight calibration are output; S3. The fused data is mapped to the dynamic grid unit, and the weight of the calculated Calculate the density entropy value, speed mutation index and flow consistency index of each grid. When the three indicators exceed the adaptive threshold at the same time, fuse the density entropy and speed mutation gradient to generate the congestion propagation direction vector, and mark the boundary of the risk area; S4. For the obscured target, according to the last known position, speed vector and grid density distribution data of the target, generate a virtual trajectory through the density-constrained trajectory prediction model. When the target re-identification position error exceeds the tolerance value, adjust the model parameters; S5. Make a spatial overlap judgment on the predicted virtual trajectory and the risk area boundary. If there is an intersection, trigger the directional diversion instruction. At the same time, according to the spatial gradient direction of the congestion propagation vector, dynamically adjust the gate release frequency according to the negative correlation principle.

[0006] Furthermore, an environmental interference matrix is constructed based on the electromagnetic attenuation rate of the radar, the illumination mutation factor of the thermal imaging, and the mechanical noise factor of the piezoelectric sensor. The matrix includes the following steps: calculating the attenuation slope of the radar spectrum curve as the electromagnetic attenuation rate factor, analyzing the overall amplitude of the grayscale change of pixels in adjacent frames of the thermal imaging as the illumination mutation factor, and accumulating the energy of the piezoelectric sensor vibration signal in the characteristic frequency band of the human gait to generate the mechanical noise factor; integrating the three factors into a two-dimensional matrix, and correlating the instantaneous interference states of different sensors through matrix operations.

[0007] Furthermore, dynamically calculating the weight value of each sensor according to the inverse relationship of each factor in the matrix includes the following steps: establishing a mapping relationship between the interference factor and the sensor weight, allocating the weight coefficient according to the exponential inverse relationship based on the numerical value of each factor in the environmental interference matrix, and dynamically allocating the weights of the radar, thermal imager and piezoelectric sensor through preset adjustment rules; the adjustment rule logic is as follows: when the electromagnetic attenuation rate factor is detected to be increased, the radar weight is increased, when the illumination mutation factor is enhanced, the thermal imaging weight is reduced, and when the mechanical noise factor exceeds the limit, the piezoelectric sensor weight is reduced.

[0008] Furthermore, step S2 includes the following steps: pre-dividing the monitoring area into several sensing units; when the weighted density value collected by the thermal imaging device in a unit is lower than a set threshold and the weighted speed value collected by the radar in the unit is higher than another set threshold, the system triggers the piezoelectric sensor array deployed in the unit to collect gait vibration signals; performing spectral analysis on the collected signals to screen out the frequency band energy that is consistent with the characteristics of human steps, and then multiplying the three types of observation results of thermal imaging, radar, and piezoelectricity by corresponding weights and performing weighted fusion to output the fused crowd density, movement speed and main flow direction of the unit.

[0009] Furthermore, the fused data is mapped to dynamic grid cells, and the density entropy, velocity mutation index and flow consistency index of each grid are calculated, including the following steps: the monitoring area is divided into dynamically scalable hexagonal grids, and the density entropy is calculated by the information entropy of the temperature distribution in the grid to quantify the disorder of crowd gathering; the velocity mutation index is generated by normalizing the absolute value of the difference of the velocity vectors of adjacent frames to capture the motion instability characteristics; the flow consistency index is calculated based on the inverse normalization of the directional angle variance to identify the stagnation trend of the crowd.

[0010] Furthermore, when the three indicators exceed the adaptive threshold at the same time, the gradient of density entropy and velocity mutation is fused to generate a congestion propagation direction vector, and the boundary of the risk area is marked, including the following steps: only when the density entropy value, velocity mutation index and flow consistency index exceed the adaptive threshold simultaneously, the spatial gradient of the product of density entropy and velocity mutation is calculated; along the gradient direction, a congestion propagation vector with a length proportional to the gradient modulus is generated to indicate the direction and intensity of congestion diffusion, and the continuous connected grid with the largest gradient modulus is marked as the boundary of the risk area.

[0011] Furthermore, for the obscured target, a virtual trajectory is generated through a density-constrained trajectory prediction model based on the target's last known position, velocity vector and grid density distribution data, including the following steps: obtaining the target's last position coordinates and corresponding velocity direction before being obscured, and then constructing a density field constraint based on the density distribution of the target's grid and surrounding grids; based on the target's current position and velocity direction, guided by the density gradient, calculating the target's possible acceleration or deceleration trend, and on this basis, generating continuous predicted position points in a time series. During the prediction process, the intersection risk of the target and the surrounding crowd's trajectory is monitored in real time. If the predicted path is found to intersect with the paths of others, the predicted speed or direction is adjusted through the obstacle avoidance algorithm to keep the trajectory consistent with the actual crowd behavior.

[0012] Furthermore, when the target re-identification position error exceeds the tolerance value, the model parameters are adjusted, including the following steps: when the re-identification position deviation exceeds the allowable range, a loss function for the trajectory error is established, the grid density and velocity information when the error occurs are statistically analyzed, and the main factors of the prediction deviation are analyzed; based on the analysis results, the density-sensitive parameters, velocity response coefficients, and obstacle avoidance weights in the trajectory prediction model are increased or decreased.

[0013] Furthermore, step S5 includes the following steps: spatially comparing the generated virtual predicted trajectory with the marked risk area boundary grid, mapping the predicted path to the grid coordinate system and determining whether it overlaps or contacts the risk area outline; if it is determined that there is a spatial intersection, determining the coordinates of the intersection of the risk boundary and the trajectory, and using this as a reference to select the nearest safe guidance position and generate guidance instructions; obtaining the congestion propagation vector field calculated in all grids in the current time slot, and identifying the directionality of the vector in the grid where the gate is located; if the direction is opposite to the normal output passage direction of the gate, automatically issuing an instruction to reduce the release frequency of the gate; if the vector direction is consistent with the release direction, then increasing the release frequency.

[0014] The dynamic identification system of human flow based on multi-sensor fusion includes the following modules: interference weight calculation module, occlusion target verification module, grid feature extraction module, trajectory prediction module, and guidance control module; the interference weight calculation module is used to construct an environmental interference matrix according to the electromagnetic attenuation rate of the radar, the illumination mutation factor of the thermal imaging and the mechanical noise factor of the piezoelectric sensor, and dynamically calculate the weight value of each sensor according to the inverse relationship of each factor in the matrix; the occlusion target verification module is used to call the gait spectrum characteristics of the piezoelectric sensor to verify the existence of the target when the density weight value of the thermal imaging in the occlusion area is lower than the first threshold and the speed weight value of the radar corresponding area is higher than the second threshold, and output the weight-calibrated fusion density, speed and flow direction data; the grid feature extraction module is used to convert The fused data is mapped to dynamic grid cells, and the density entropy value, speed mutation index and flow consistency index of each grid are calculated. When the three indicators exceed the adaptive threshold at the same time, the gradient of the density entropy and speed mutation is fused to generate the congestion propagation direction vector, and the boundary of the risk area is marked; the trajectory prediction module is used to generate a virtual trajectory for the obscured target based on the last known position, speed vector and grid density distribution data of the target through a density-constrained trajectory prediction model, and adjust the model parameters when the target re-identification position error exceeds the tolerance value; the diversion control module is used to perform spatial overlap judgment on the predicted virtual trajectory and the risk area boundary. If there is an intersection, a directional diversion instruction is triggered, and at the same time, the gate release frequency is dynamically adjusted according to the negative correlation principle based on the spatial gradient direction of the congestion propagation vector.

[0015] The present invention has the following beneficial effects: (1) A dynamic identification method for pedestrian flow based on multi-sensor fusion. This method realizes adaptive perception of complex environmental interference and joint processing of multi-source data. On the one hand, the environmental interference matrix can uniformly quantify the electromagnetic attenuation information of the radar, the illumination mutation information of the thermal imaging, and the mechanical noise information of the piezoelectric sensor, and assign dynamic weights to each sensor, so that the system can still maintain high sensor availability under various interference conditions such as rain, fog, strong light, and vibration. On the other hand, the density, velocity, and flow direction data after weighted fusion are mapped to the dynamic grid, and indicators such as density entropy, velocity mutation, and flow direction consistency are calculated in real time. When the three indicators exceed the threshold together, the congestion propagation direction vector is automatically generated and the risk area boundary is marked, realizing the rapid detection and spatial positioning of the congestion situation. The synergistic effect of the above functions effectively improves the recognition accuracy and real-time performance of pedestrian flow changes, and at the same time enhances the robustness and reliability of the system under multiple interferences such as occlusion, illumination, and noise. For targets temporarily out of visual or radar observation, a density-constrained trajectory prediction model generates virtual trajectories, providing continuous trajectory clues when the target reappears. When the predicted position deviates from the actual re-identified position beyond the tolerance, a parameter adjustment mechanism is automatically triggered, updating the trajectory prediction model in real time to minimize the impact of occlusion on recognition. Furthermore, the virtual trajectories are spatially overlapped with the risk area boundaries, and congestion propagation vectors are combined to make traffic flow decisions and adjust gate frequency. This allows for the timely guidance and diversion of people who may enter high-risk areas, effectively alleviating localized congestion and improving overall traffic efficiency.

[0016] (2) A dynamic crowd flow identification system based on multi-sensor fusion integrates five modules: interference weight calculation, occlusion verification, grid analysis, trajectory prediction, and diversion control, forming a closed-loop dynamic crowd flow identification and control platform. In the interference weight calculation module, the weight of each sensor can be adjusted in real time according to the environmental state, making the data fusion result more accurate; in the occlusion verification module, multi-modal sensing compensates each other, achieving high-reliability identification of targets in the occluded area; in the grid feature extraction module, real-time grid density entropy, velocity mutation, and flow consistency indicators can accurately depict the dynamic behavior of the crowd; in the trajectory prediction module, the virtual trajectory generated for the occluded target and the dynamic parameter update mechanism enable the system to respond quickly to the occlusion state; in the diversion control module, the predicted trajectory is combined with the congestion propagation vector to perform spatial overlap judgment, and the gate release frequency is negatively adjusted, achieving accurate early warning and active intervention of potential congestion risks. Through the collaborative work of the above modules, the system significantly improves the overall accuracy and efficiency of crowd flow monitoring in complex urban scenarios, providing reliable technical support for public safety, emergency management, and smart transportation.

[0017] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the method for dynamic identification of pedestrian flow based on multi-sensor fusion of the present invention.

[0019] Figure 2 This is a flow chart of the dynamic pedestrian flow recognition system based on multi-sensor fusion of the present invention. DETAILED DESCRIPTION

[0020] The embodiments of the present application solve the key problems of inaccurate dynamic perception of pedestrian flow, delayed response to congestion risks, and interruption of target occlusion tracking in complex environments through a method and system for dynamic identification of pedestrian flow based on multi-sensor fusion.

[0021] The overall idea of the solution in the embodiments of this application is as follows: First, a multi-source perception system is built using radar, thermal imaging and piezoelectric sensors. Based on the interference characteristics of each sensor in different environments, the electromagnetic attenuation rate, light mutation factor and mechanical noise factor are extracted, an environmental interference matrix is constructed and perception weights are dynamically allocated to improve the reliability of sensor data.

[0022] Secondly, when thermal imaging detection fails, the radar speed information and piezoelectric gait spectrum are combined to verify the presence of the target, thereby achieving accurate identification of people in the obscured area.

[0023] Subsequently, the fused density, velocity and flow direction data are mapped to a dynamic grid, and indicators such as density entropy, velocity mutation index and flow direction consistency are extracted to identify local aggregation and motion abnormality areas.

[0024] Furthermore, for targets that are obscured or temporarily out of contact, a density-constrained trajectory prediction model is used to generate their virtual paths, and the model is dynamically corrected when the error exceeds the tolerance.

[0025] Finally, the predicted trajectory is spatially matched with the boundary of the risk area. If a potential collision trend is found, the gate is linked through evacuation instructions, and the gate release rhythm is dynamically adjusted according to the trend of human flow spread, realizing early risk warning and local evacuation control.

[0026] See also Figure 1, an embodiment of the present invention provides a technical solution: a dynamic identification method for human flow based on multi-sensor fusion, comprising the following steps: S1. constructing an environmental interference matrix based on the electromagnetic attenuation rate of the radar, the illumination mutation factor of the thermal imaging, and the mechanical noise factor of the piezoelectric sensor, and dynamically calculating the weight value of each sensor according to the inverse relationship of each factor in the matrix; S2. when the density weight value of the thermal imaging in the blocked area is lower than the first threshold and the speed weight value of the radar corresponding area is higher than the second threshold, calling the gait spectrum characteristics of the piezoelectric sensor to verify the existence of the target, and outputting the fused density, speed and flow direction data after weight calibration; S3. mapping the fused data to the dynamic grid unit, calculating each grid The density entropy value, speed mutation index and flow consistency index of the target are used. When the three indicators exceed the adaptive threshold at the same time, the gradient of density entropy and speed mutation is integrated to generate the congestion propagation direction vector, and the boundary of the risk area is marked; S4. For the obscured target, according to the last known position, speed vector and grid density distribution data of the target, a virtual trajectory is generated through the density-constrained trajectory prediction model. When the target re-identification position error exceeds the tolerance value, the model parameters are adjusted; S5. The predicted virtual trajectory is judged to overlap with the boundary of the risk area in space. If there is an intersection, the directional diversion instruction is triggered. At the same time, the gate release frequency is dynamically adjusted according to the negative correlation principle based on the spatial gradient direction of the congestion propagation vector.

[0027] In this implementation, Step S1 primarily assesses and integrates the reliability of multi-source sensor data. The electromagnetic attenuation rate reflects the degree of energy attenuation caused by obstruction or interference in the radar signal; greater attenuation indicates less reliable radar data. The illumination mutation factor measures the stability of thermal imaging equipment images under complex lighting conditions (such as direct sunlight or changing shadows). The mechanical noise factor measures signal fluctuations in piezoelectric sensors caused by ground vibrations, frequent footsteps, or other mechanical disturbances. By constructing an environmental interference matrix, these factors are standardized and reflected in the weight distribution, employing the inverse proportional principle: greater interference results in a smaller sensor weight, thereby enhancing overall perception accuracy. This addresses the issue of varying reliability of multi-sensor data in dynamic scenarios, enabling adaptive weight adjustment and laying the foundation for subsequent fusion. Step S2 improves the ability to detect people in obscured areas. Occlusion areas refer to locations where the field of view is limited or where thermal imaging equipment cannot directly detect (such as under awnings or in corners). Density weights refer to the thermal imaging sensor's contribution to the crowd density within the current area. Speed weights refer to the radar's response to speed changes within the area. Gait spectrum features refer to signals such as gait rhythm, frequency, and pressure changes detected by the piezoelectric sensor, which serve as important evidence of an individual's presence. When thermal imaging information is unreliable but radar detects significant speed changes, the piezoelectric signature is used for confirmation, ultimately generating fused density, speed, and flow direction data weighted and calibrated across the three sensor types. This avoids false positives and missed detections due to thermal imaging unavailability, and enhances the accuracy of identifying human presence in complex scenarios through redundant verification. Step S3: This step implements local dynamic analysis of human flow within a spatial area. Dynamic grid cells: Divide the monitoring area into multiple two-dimensional small areas (grids) to facilitate local data analysis. Density entropy: Describes the degree of disorder in crowd distribution. High entropy indicates dramatic density fluctuations and uneven clustering. Speed mutation index: Measures dramatic speed fluctuations within a grid, such as sudden stops or turns. Flow consistency: Indicates whether people within an area are moving in the same direction. Low consistency may indicate chaos or stagnation. When all three indicators are abnormal, indicating a potential risk (such as clustering or conflict) in the area, a congestion propagation direction vector is generated by superimposing density and velocity gradients, and the risk boundary is marked accordingly. This enables early identification of abnormal crowd dynamics, supporting proactive demarcation of risk areas and subsequent guidance. Step S4: This step addresses the issue of individuals losing contact for short periods within obscured areas. Density-constrained trajectory prediction model: Building on conventional physical motion prediction, grid density data is introduced as a constraint to simulate the motion deviation characteristics of individuals in high-density areas. Re-identification position error: The spatial difference between the model's predicted position and the actual re-identification result. Tolerance: The maximum acceptable error value, exceeding which triggers model self-learning or correction. If the error is too large, the key model parameters such as density sensitivity coefficient and velocity response factor are reversely optimized through the trajectory error function to make the prediction gradually converge.Enhance the ability to continuously track targets that are temporarily occluded, and improve the stability and robustness of crowd tracking. Step S5: This step realizes the linkage response of the crowd management system and the traffic control system. Spatial overlap judgment: By calculating whether the predicted path intersects with the boundary of the risk area, it is judged whether the individual is approaching the risk area; Directional guidance instructions: prompt the monitoring system or intelligent indicators to guide the crowd to move in a safe direction; Gate release frequency: The number of people allowed to pass per unit time is dynamically adjusted according to the degree of match between the vector direction and the traffic direction. If the congestion vector is opposite to the traffic direction, the release rate is reduced to ease the entrance pressure; if it is consistent, the traffic efficiency is appropriately improved to balance the flow of people. Open up the detection-prediction-control closed loop to achieve active intervention and adaptive adjustment of peak traffic flow.

[0028] Specifically, an environmental interference matrix is constructed based on the electromagnetic attenuation rate of the radar, the illumination mutation factor of the thermal imaging, and the mechanical noise factor of the piezoelectric sensor. The matrix includes the following steps: calculating the attenuation slope of the radar spectrum curve as the electromagnetic attenuation rate factor, analyzing the overall amplitude of the grayscale change of pixels in adjacent frames of the thermal imaging as the illumination mutation factor, and accumulating the energy of the piezoelectric sensor vibration signal in the characteristic frequency band of the human gait to generate the mechanical noise factor; integrating the three factors into a two-dimensional matrix, and correlating the instantaneous interference states of different sensors through matrix operations.

[0029] In this implementation, the electromagnetic attenuation factor is calculated by analyzing the radar's echo power changes at different distances or environmental conditions to quantify the degree of electromagnetic signal attenuation, which serves as the basis for subsequent weight allocation. The following steps describe the extraction of the radar electromagnetic attenuation factor: By analyzing the attenuation trend of the radar return signal strength with distance, the electromagnetic attenuation factor reflecting the propagation interference is extracted. The expression is: ; Parameter explanation: :Electromagnetic attenuation factor; :Receive signal at distance The intensity value at : radial distance between target and radar; : Expressed as a common logarithm or natural logarithm, it is used to linearize the power-distance relationship. Technical meaning: This factor represents the dynamic attenuation intensity change of the signal during propagation, reflecting the degree of instantaneous interference to the radar in complex environments (such as wall obstruction and metal reflection). Calculation steps for the thermal imaging illumination mutation factor: Calculate the global grayscale change of the thermal imaging image between adjacent frames to quantify the impact of illumination fluctuations on infrared recognition performance. Mathematical expression: ; Parameter explanation: :Light mutation factor; :Timeframe Lower Pixel Gray value of : The width and height of the thermal imaging image; : Image time frame index. Steps for generating the piezoelectric sensor mechanical noise factor: Perform spectrum analysis on the piezoelectric sensor vibration signal and calculate its total energy within the human gait frequency band to generate an indicator reflecting the mechanical interference intensity. Mathematical expression: ; Parameter explanation: : Mechanical noise factor; : spectrum function of the piezoelectric vibration signal after Fourier transform; : Lower limit of the main frequency band of human gait (e.g. starting value of gait frequency); : Upper limit of the main frequency band of human gait (e.g., cadence termination value); : Frequency energy density function. Technical meaning: This factor quantitatively measures the total energy of the piezoelectric signal in a specific frequency band within a certain time window, representing the degree of potential mechanical vibration interference (such as resonance and ground impact) in the scene. Steps for constructing a two-dimensional environmental interference matrix: The three interference factors mentioned above are normalized and integrated to form a multi-sensor interference state matrix for dynamic weight adjustment. Expression form: For the single-moment interference state of an area: ; Multiple regions or multiple moments are combined into a matrix: ; Parameter explanation: : No. Normalized electromagnetic attenuation factor of the region; : No. Normalized illumination mutation factor of the region; : No. Normalized mechanical noise factor of the region; : the total number of regions or sampling points; : Environmental interference matrix, serving as the input basis for subsequent weight adjustment and fusion decision-making. Technical meaning: Matrix It is used to reflect the differences in interference status of various sensors in the current environment, and serves as the basis for weight allocation and information confidence adjustment in sensor fusion.

[0030] Specifically, dynamically calculating the weight value of each sensor according to the inverse relationship of each factor in the matrix includes the following steps: establishing a mapping relationship between interference factors and sensor weights, allocating weight coefficients according to an exponential inverse relationship based on the numerical values of each factor in the environmental interference matrix, and dynamically allocating the weights of radar, thermal imager and piezoelectric sensor through preset adjustment rules; the adjustment rule logic is as follows: when the electromagnetic attenuation rate factor is detected to be increased, the thermal imaging weight is reduced when the illumination mutation factor is enhanced, and the piezoelectric sensor weight is reduced when the mechanical noise factor exceeds the limit.

[0031] In this implementation, the interference factor and sensor weight are established: to avoid the problem of insensitivity to gradient changes caused by direct linear inverse proportion, an exponential mapping function is introduced to make the sensor have a nonlinear weight response to abnormal interference. Mathematical expression: ; ; Parameter explanation: : Weight coefficient of radar sensor; : Weight coefficient of thermal imaging sensor; : Weight coefficient of piezoelectric sensor; : The interference factor value of the radar at the current moment (electromagnetic attenuation rate); : The interference factor value of thermal imaging at the current moment (light mutation factor); : The interference factor value (mechanical noise factor) of the piezoelectric sensor at the current moment; : Adjust the sensitivity coefficient, corresponding to the mapping response slope of radar, thermal imaging and piezoelectric channels respectively; : Normalization factor, ensuring that the sum of the weights of the three sensors is 1. Technical meaning: This formula ensures that when environmental interference increases, the contribution value of the corresponding sensor decreases rapidly through exponential decay, rather than simple mean distribution, so that the fusion result is closer to the actual signal-to-noise ratio. Dynamic adjustment logic based on adjustment rules, step description: By setting the change trend trigger conditions of each type of factor, dynamically adjust the corresponding sensor weight priority to achieve adaptive response. Adjustment rule logic explanation: Radar weight increase rule: If Increases (i.e., the radar is interfered with less), then the downward trend of the exponential term in the weight function slows down, resulting in Increase; Explanation logic: Radar signal penetration is enhanced, suitable for taking the lead in flow velocity monitoring. Thermal imaging weight reduction rules: If Increase (i.e., drastic changes in illumination), the exponential term changes drastically, resulting in Rapid decline; Explanation logic: Infrared images are affected by high-contrast lighting, and the risk of distortion increases, so it is advisable to reduce participation. Piezoelectric weight suppression rule: If > (Threshold judgment), then set γ→γ′>γ to enhance the sensitive response of the piezoelectric signal to noise, so that Rapid compression; Explanation: Ground mechanical impact interference exceeds the human gait threshold, limiting the risk of misjudgment. Technical Mechanism Summary: Through an exponential inverse mapping model, each sensor weight dynamically responds to changes in interference factors. This offers the following advantages: Strong robustness: Avoids the blunt response of linear functions to abnormal fluctuations; High fusion accuracy: Weight changes reflect the dynamic characteristics of information confidence; Good adjustability: Parameters α, β, and γ allow for customizing weight sensitivity in different scenarios; Deployment-friendly: Does not rely on large sample training and can be embedded in lightweight edge systems.

[0032] Specifically, step S2 includes the following steps: pre-dividing the monitoring area into several sensing units; when the weighted density value collected by the thermal imaging device in a unit is lower than a set threshold and the weighted speed value collected by the radar in the unit is higher than another set threshold, the system triggers the piezoelectric sensor array deployed in the unit to collect gait vibration signals; performing spectral analysis on the collected signals to screen out the frequency band energy that is consistent with the characteristics of human steps; then multiplying the three types of observation results of thermal imaging, radar, and piezoelectricity by the corresponding weights and performing weighted fusion to output the fused crowd density, movement speed and main flow direction of the unit.

[0033] In this implementation, the sensing unit division and trigger condition check divide the monitoring area into several sensing units, each of which is indexed by express; Unit, obtains the current crowd density estimate from the thermal imaging sensor ; Unit, obtains the current average speed of the crowd from the radar sensor ; System preset thermal imaging density threshold and radar speed threshold ; Trigger condition: If and , then go to the next step, otherwise continue monitoring. Parameter explanation: : Perception unit index; : No. The thermal imaging weighted density estimation of the unit is obtained by multiplying the original thermal imaging density value by the infrared weight of the unit; : No. The radar weighted average speed of the unit is obtained by multiplying the radar raw speed by the radar weight of the unit; : Thermal imaging density judgment threshold, used to determine whether the density is too low; : Radar speed judgment threshold, used to determine whether the speed is too high. Piezoelectric array gait vibration signal acquisition and spectrum analysis are performed in the first In the unit, the system starts the piezoelectric sensor matrix, and each node outputs a time domain vibration signal, which is recorded as ,in Represents the piezoelectric node index; for each Perform Fourier transform to get the spectrum ; In the human gait characteristic frequency band The accumulated energy is calculated Piezoelectric gait energy of the unit: like , then it is confirmed that the human gait exists; otherwise, the piezoelectric is determined to be invalid and further processing of this unit is terminated. Parameter explanation: : No. Unit No. Time domain vibration signal of the piezoelectric node; : The first obtained after Fourier transform Node spectrum; : lower and upper limits of gait characteristic frequency band; : The accumulated energy in the gait frequency band represents the gait intensity detected by the piezoelectric; : Gait energy judgment threshold, used to distinguish real human gait from environmental vibration. Multi-source observation weighted fusion, after confirming the existence of the target in the a-th unit, obtains three types of observations: thermal imaging density value ; Radar speed value ; The crowd density or speed value calculated by piezoelectricity is recorded as ; The system calculates the dynamic weights of the three sensors at that moment ,satisfy ; Final weighted fusion output: ; ; ; Parameter explanation: : Dynamic weights of infrared, radar, and piezoelectric in unit a; : Crowd density estimation extracted by three types of sensors; : Crowd velocity estimation extracted by three types of sensors; : Flow direction angles measured by three types of sensors; : population density after fusion; : The speed of the crowd after fusion; : mainstream flow direction after fusion; : The unit phase vector represents the direction; : Complex phase extraction operator, used to obtain the flow angle after weighted synthesis.

[0034] Specifically, the fused data is mapped to dynamic grid cells, and the density entropy, velocity mutation index and flow consistency index of each grid are calculated, including the following steps: the monitoring area is divided into dynamically scalable hexagonal grids, and the density entropy is calculated by the information entropy of the temperature distribution in the grid to quantify the disorder of crowd gathering; the velocity mutation index is generated by normalizing the absolute value of the difference of the velocity vectors of adjacent frames to capture the motion instability characteristics; the flow consistency index is calculated based on the inverse normalization of the directional angle variance to identify the stagnation trend of the crowd.

[0035] In this implementation, hexagonal grid division and dynamic scaling are described as follows: the entire monitoring area is divided into several regular hexagonal units that can be dynamically scaled, and each grid is indexed According to the current crowd density or the preset scaling rules, the hexagon side length is automatically adjusted to ensure that the grid is finer in high-density areas and coarser in low-density areas, thus taking into account both spatial resolution and computational efficiency. Dynamic scaling basis: If the crowd density of a grid at the previous moment exceeds the high-density threshold , the grid side length is reduced to a certain proportion of the original; if the density is lower than the low density threshold , then the side length is appropriately enlarged to reduce the amount of calculation. Hexagonal grid index: The hexagonal grid is , the grid center coordinates are The purpose of calculating the grid density entropy value is to quantify the disorder of the distribution of people or temperature data in the grid through information entropy, reflecting the state of crowd gathering and sparseness. The temperature value or normalized population density is divided into subintervals, corresponding to the probability distribution , then the density entropy value of the grid is Calculated by the following formula: ; Parameter explanation, : No. The density entropy value of each grid is used to quantify the disorder of the crowd or heat distribution; : The total number of sub-intervals in each grid can be set dynamically according to temperature or density range; : No. Grid The probability of the population (or temperature) in each subinterval; : natural logarithm, used to calculate entropy; calculation process: the fused crowd density or thermal imaging temperature value is calculated in the grid The internal structure is divided into two types according to equal distance or equal frequency. small intervals; count the proportion of samples in each interval to the total number of grid samples, and get ; Substitute the above probability into the formula to obtain the grid disorder The purpose of calculating the grid speed mutation index is to capture the violent fluctuation or unstable characteristics of the motion state by measuring the speed change of the crowd in the grid at adjacent moments. At the moment and The average velocity vectors detected are and , then the velocity mutation index of the grid is It can be calculated and normalized by the following formula: ; Parameter explanation :No. The speed mutation index of each grid is used to indicate the degree of speed change; :time No. The average velocity vector of the crowd at the grid; :time No. The average velocity vector of the crowd at the grid; : Speed difference vector, which represents the speed change between two adjacent moments; : The modulus of the velocity difference vector, calculated as ; :time Or a globally set velocity normalization factor (which can be the maximum normal velocity value of the grid), used to map the mutation index to the [0,1] interval; : A very small constant used to avoid zero division errors. Calculation process, statistical grid All individuals within at time and The average value of the velocity vector is obtained and ; Calculate the velocity difference vector , and find its model ; Use normalization factor Standardize and obtain The purpose of calculating the grid flow consistency index is to identify whether the crowd is stagnant, hesitant or confused by counting the discrete degree of the movement direction of all individuals in the grid. , first calculate its direction vector and Cp, and then normalize it to get the flow consistency index Qp: ; Parameter explanation, Qp: flow direction consistency index of the p-th grid, the closer the value is to 1, the direction is highly consistent, and the closer it is to 0, the direction is discrete; K: the total number of individuals tracked in the grid; φm: the movement direction angle of the m-th individual (radians), usually calculated from the velocity vector; : represents the projection of the direction angle φm onto the x-axis and y-axis, respectively; Cp: the modulus of the sum of the direction vectors, used to quantify the overall directional concentration within the grid. Calculation process: Calculate the instantaneous motion direction φm for each tracked object in the grid Gp; convert each direction into the unit vector form [cos(φm), sin(φm)]T; sum all K unit vectors and take the modulus Cp; divide Cp by K for normalization to obtain the consistency index Qp.

[0036] Specifically, when the three indicators exceed the adaptive threshold at the same time, the gradient of density entropy and velocity mutation is fused to generate a congestion propagation direction vector, and the boundary of the risk area is marked, including the following steps: only when the density entropy value, velocity mutation index and flow consistency index exceed the adaptive threshold simultaneously, the spatial gradient of the product of density entropy and velocity mutation is calculated; along the gradient direction, a congestion propagation vector with a length proportional to the gradient modulus is generated to indicate the direction and intensity of congestion diffusion, and the continuous connected grid with the largest gradient modulus is marked as the boundary of the risk area.

[0037] In this embodiment, when The density entropy value of the grid , speed mutation index and flow consistency index Satisfy at the same time ; That is, the synchronization exceeds the respective adaptive threshold 、 、 When , the following steps are performed: Calculate the coupling field strength and define the coupling strength function At each grid center point Chuwei ; : No. The coupling strength of the grid; : No. The density entropy value of the grid; : No. The velocity mutation index of the grid. Calculate the spatial gradient in the two-dimensional grid plane and calculate the coupling strength For horizontal (+ direction) and vertical (+ direction) to construct the gradient component: ;in: : No. The horizontal gradient component of a grid is determined by its right neighbor grid With the left neighboring grid Calculation of coupling strength difference; : No. The vertical gradient component of a grid is determined by the gradient of its upper neighboring grid. and the lower adjacent grid Calculation of coupling strength difference; :represent the The grid's right, left, top, and bottom neighboring grid indices in the hexagonal grid layout. Generate the congestion propagation direction vector by combining the above gradient components into a vector: ; Let the modulus of this vector be ; and set its length to match It is proportional to the diffusion intensity of congestion at the grid; the vector direction indicates the diffusion direction of congestion. : No. The congestion propagation direction vector of the grid; : Gradient modulus, indicating the direction strength of the fastest local coupling change. Mark the risk area boundary, for all , , The grid is filtered out to find its gradient modulus Greater than a certain adaptive gradient threshold A subset of ; : Gradient modulus threshold, used to identify local strong coupling change areas; in the collection In the process, connectivity clustering is performed according to the spatial adjacency relationship of the grid (adjacent hexagonal sides), and several continuous connected subsets are obtained. For each connected subset , marking its outer boundary grids (grids adjacent to grids in the subset but not belonging to the subset form the boundary outline). These outer boundary grids are defined as the risk area boundaries and are used for subsequent reminders and guidance. : connected subset index ; Risk area boundary: by The set of identified outer edges of the mesh.

[0038] Specifically, for the obscured target, a virtual trajectory is generated through a density-constrained trajectory prediction model based on the target's last known position, velocity vector and grid density distribution data, including the following steps: obtaining the target's last position coordinates and corresponding velocity direction before being obscured, and then constructing a density field constraint based on the density distribution of the target's grid and surrounding grids; based on the target's current position and velocity direction, guided by the density gradient, calculating the target's possible acceleration or deceleration trend, and on this basis, generating continuous predicted position points in a time series. During the prediction process, the intersection risk of the target and the surrounding crowd's trajectory is monitored in real time. If the predicted path is found to intersect with the paths of others, the predicted speed or direction is adjusted through the obstacle avoidance algorithm to keep the trajectory consistent with the actual crowd behavior.

[0039] In this implementation, the last state before the occlusion is obtained: before the target enters the occlusion area, the system knows its last observable position and velocity information, which is used as the initial condition for prediction. The last known position of is the vector ; Let the velocity direction unit vector at that time be ; The actual speed is recorded as These data are used to initialize the subsequent prediction process. Construct density field constraints: Using the gridding results, obtain the crowd density distribution of the target grid and its surrounding grids to form a continuous "density field" to guide the trajectory deviation. Let the grid number be , then The density of the grid is ; Expand each grid density into a continuous function by smooth interpolation or bilinear interpolation , represents the local density at any position; the current position is obtained through the density function The density value at The density field As a constraint, the target is guided to show a clear avoidance trend in the high-density area. Calculation of density gradient and acceleration trend: Using the spatial gradient of the density field, the avoidance or deceleration trend of the target in the occluded area is estimated, and the acceleration vector is constructed during prediction. Density gradient calculation: By taking the partial derivative of the interpolated density field, the acceleration vector at position The gradient vector at ; : density field gradient, used to indicate the direction in which the density rises fastest; :Respectively and Acceleration vector construction: The avoidance trend of the target within the grid can be calculated using the acceleration vector It is expressed in the form of an inverse relationship with the density gradient: ; : Density sensitivity coefficient, used to adjust the response intensity to density gradient; : Speed inertia coefficient, used to maintain the original motion trend of the target; :time The predicted speed of :time The unit velocity direction vector of Indicates that the target tends to avoid the direction of high density; the second Keep a certain amount of inertia to avoid sudden changes in trajectory. Time series predicted position update description: Based on the current velocity and acceleration information, the discrete kinematics formula is used to generate the predicted position at the next moment. ; ; :No. and Predicted position at the moment; :No. and The speed at the moment; :No. and The unit vector of velocity direction at the moment; : prediction time step; :No. Acceleration vector at the moment. Real-time collision and intersection risk monitoring instructions: After each position update, it is necessary to determine whether the current predicted position has intersections or path conflicts with other surrounding targets or known trajectories. If a significant overlap risk with other trajectories is detected, the current predicted vector is fine-tuned to avoid collision. Crossing risk determination: Compare the predicted positions Set of predicted or known waypoints to other targets ,like ; then it is considered that there is cross risk, among which Is the preset safety distance threshold. Obstacle avoidance adjustment: If a conflict is detected, modify the current acceleration direction or speed direction, for example: ; : adjusted acceleration; : A unit vector perpendicular to the current velocity direction, used for lateral avoidance; : Obstacle avoidance fine-tuning coefficient, controlling the lateral avoidance range; : Recalculate the velocity direction based on the adjusted acceleration.

[0040] Specifically, when the target re-identification position error exceeds the tolerance value, the model parameters are adjusted, including the following steps: when the re-identification position deviation exceeds the allowable range, a loss function of the trajectory error is established, the grid density and speed information when the error occurs are statistically analyzed, and the main factors of the prediction deviation are analyzed; based on the analysis results, the density-sensitive parameters, speed response coefficients and obstacle avoidance weights in the trajectory prediction model are increased or decreased.

[0041] In this implementation, error determination and loss function construction, error determination: at time , the system re-identifies the actual position of the target as This position is different from the previous predicted position The spatial deviation between is defined as: ;if ,in If the error tolerance value is set in advance, the parameter optimization process is started; otherwise, the current model parameters are kept unchanged. Loss function construction: To evaluate the deviation between the predicted trajectory and the actual trajectory, a trajectory error loss function is defined , consider the cumulative error between all predicted positions and the true position during the period from the occlusion point to the re-identification point. Time index ,but: ; Parameter explanation: : average trajectory error loss; : From the last observable moment To re-identification time The number of time steps; : At the moment The virtual position vector generated by the prediction model; : At the moment The true position or the approximate actual position obtained by re-identification interpolation; : Euclidean norm, used to calculate the distance of position vector difference. Loss function It reflects the fitting error between the predicted trajectory and the actual trajectory during the entire occlusion period. The grid density and velocity information statistics when the error occurs, the positioning error moment: find the moment Make ; That is, the time index with the largest positioning difference. Grid density and speed statistics: The grid number is recorded as , the density distribution and average velocity information of this grid and its adjacent grids are as follows: ; Parameter explanation: : current grid index; : With grid adjacent grid index; : Grid and the density values of its adjacent grids; : Grid The above information is used to analyze the environmental factors that lead to prediction deviation and is the basis for subsequent parameter adjustment. The density and velocity information of its adjacent grids and the prediction error Combined, determine the main reasons for deviation, including: High-density congestion deviation: If or its adjacent Significantly higher than the average density of the environment, indicating that the target may deviate from the original predicted path due to avoidance; Insufficient speed response: If the speed gradient of the grid If the error is large, it means that the prediction model has a delayed response to the speed mutation; the obstacle avoidance behavior is insufficient: if there are other targets occupying the path in the adjacent grid, it means that the model has not fully considered the obstacle avoidance weight. Model parameter adjustment Based on the deviation analysis results, the following three key parameters are adjusted to modify the prediction model: density sensitive parameters If the high-density deviation dominates, increase Otherwise, keep or reduce appropriately. Adjustment method example (gradient update): ; : density sensitive parameter; : corresponding learning rate; : Loss function pair The partial derivative of , when calculating, associates the weights of the acceleration term and the density gradient term in the predicted trajectory. Velocity response coefficient If the speed suddenly deviates significantly, increase ; On the contrary, it can be reduced or maintained. Gradient update form: ; : speed response coefficient; : speed adjustment learning rate; : Loss function pair The partial derivative of reflects the influence of this coefficient on the overall trajectory error. If the conflict deviates from the problem, then increase Otherwise, do nothing or reduce it moderately. Gradient update example: ; : obstacle avoidance weight coefficient; : Obstacle avoidance adjustment learning rate; : The partial derivative of the loss function with respect to the obstacle avoidance weight parameter.

[0042] Specifically, step S5 includes the following steps: spatially comparing the generated virtual predicted trajectory with the marked risk area boundary grid, mapping the predicted path to the grid coordinate system and determining whether it overlaps or contacts the risk area outline; if it is determined that there is a spatial intersection, determining the coordinates of the intersection of the risk boundary and the trajectory, and using this as a reference to select the nearest safe guidance position and generate guidance instructions; obtaining the congestion propagation vector field calculated in all grids in the current time slot, and identifying the directionality of the vector in the grid where the gate is located; if the direction is opposite to the normal output passage direction of the gate, automatically issuing an instruction to reduce the release frequency of the gate; if the vector direction is consistent with the release direction, then increasing the release frequency.

[0043] In this implementation, the trajectory is spatially compared with the risk boundary, the trajectory is mapped to grid coordinates, and the virtual predicted trajectory is represented as a series of time-series position points { , ,…, },in =( , ) is the predicted coordinate at the nth moment. For each position point , which is converted into the corresponding grid index according to the grid division rules ,For example: ; : represents the hexagonal grid index where the prediction point at time n falls; GridIndex(x,y): a function that maps continuous coordinates to the grid they belong to. To determine the spatial intersection, let the risk boundary grid set be B={ , ,…}. If there exists a time n such that ∈B, the predicted trajectory is considered to have “spatially overlapped or touched” the risk area boundary; formally: If the above conditions are met, proceed to the next step; otherwise, do not perform guidance and gate adjustment. Determine the intersection and generate guidance instructions, determine the intersection coordinates of the risk boundary and the trajectory, and In the time set, select the first time index that appears , its corresponding position That is the "intersection coordinates": ; This point indicates the position where the predicted trajectory officially enters the risk area boundary. Select the nearest safe guidance position and define the safe guidance grid set , contains the indices of all grid cells that are adjacent to the risk boundary but not marked as risk. Grid index Perform the calculation: ; : The safe grid index closest to the intersection point; : Grid The center coordinates of : Euclidean distance. Generate guidance instructions based on the security grid The center coordinates of , issue "crowd guidance" or "route diversion" instructions in the system, instructing the target and surrounding personnel to follow The instruction may include text prompts, electronic screen displays or intelligent broadcasts to ensure that personnel leave the risk area in time. Dynamic adjustment of gate release frequency Obtain the congestion propagation vector field of the current time slot for all grid indexes , the system has calculated the corresponding congestion propagation vector . Determine the grid index where a gate is located. , get its corresponding vector The correlation between the judgment vector and the gate release direction is set as the unit vector corresponding to the normal gate release direction Calculate the congestion vector Release direction The dot product of: if , indicating that the congestion propagation direction is opposite to the release direction; if , then the two directions are consistent; if , indicating that the two directions are approximately perpendicular and no significant adjustment is required. :The system automatically reduces the release frequency of the gate, such as extending the gate opening interval or reducing the number of people released at the same time; when :The system appropriately increases the release frequency to speed up the evacuation of personnel; if , you can keep the current frequency or make slight adjustments.

[0044] See also Figure 2The dynamic identification system of human flow based on multi-sensor fusion includes the following modules: interference weight calculation module, occlusion target verification module, grid feature extraction module, trajectory prediction module, and guidance control module; the interference weight calculation module is used to construct an environmental interference matrix according to the electromagnetic attenuation rate of the radar, the illumination mutation factor of the thermal imaging and the mechanical noise factor of the piezoelectric sensor, and dynamically calculate the weight value of each sensor according to the inverse relationship of each factor in the matrix; the occlusion target verification module is used to call the gait spectrum characteristics of the piezoelectric sensor to verify the existence of the target when the density weight value of the thermal imaging in the occlusion area is lower than the first threshold and the speed weight value of the radar corresponding area is higher than the second threshold, and output the weight-calibrated fusion density, speed and flow direction data; the grid feature extraction module is used to convert the fusion The combined data is mapped to dynamic grid cells, and the density entropy value, speed mutation index and flow consistency index of each grid are calculated. When the three indicators exceed the adaptive threshold at the same time, the gradient of density entropy and speed mutation is fused to generate the congestion propagation direction vector, and the boundary of the risk area is marked; the trajectory prediction module is used to generate a virtual trajectory for the obscured target based on the target's last known position, speed vector and grid density distribution data through a density-constrained trajectory prediction model, and adjust the model parameters when the target re-identification position error exceeds the tolerance value; the diversion control module is used to judge the spatial overlap between the predicted virtual trajectory and the risk area boundary. If there is an intersection, a directional diversion instruction is triggered, and the gate release frequency is dynamically adjusted according to the negative correlation principle based on the spatial gradient direction of the congestion propagation vector.

[0045] In this implementation, the interference weight calculation module collects key interference factors from three heterogeneous sensor types: radar, thermal imaging, and piezoelectric sensors—including electromagnetic attenuation, illumination mutation, and mechanical noise—and constructs them into an environmental interference matrix. Subsequently, based on the inverse relationship between these factors, dynamic weights for each sensor are calculated in real time to guide subsequent fusion processing. The environmental interference matrix construction integrates interference factors of different dimensions into a matrix structure. Matrix operations or matrix decompositions (such as covariance and SVD) are used to analyze the interference coupling patterns between sensors. This provides multi-dimensional correlation information for dynamic weight assignment, rather than relying solely on single-factor evaluation. The occluded target verification module deploys the piezoelectric sensor array deployed within the perception unit to extract gait spectrum features when thermal imaging recognition fails (e.g., due to occlusion) and the radar detects a change in speed. This module verifies the presence of a real moving target in the area and outputs weight-calibrated fused density, velocity, and flow direction data. The fused density, velocity, and flow direction parameters obtained by this module are assigned a higher confidence level, directly improving the accuracy of subsequent grid analysis. The innovation of this module lies in achieving highly robust recognition of targets in occluded areas through dynamic triggering and cross-validation of multiple sensors, breaking through the bottleneck of reduced accuracy of a single technology in occluded scenarios. Grid feature extraction module: This module maps the fused density, velocity, and flow direction data into a dynamically scalable hexagonal grid, calculates the density entropy value, velocity mutation index, and flow consistency index of each grid, and generates a congestion propagation direction vector through gradient calculation when the three indicators simultaneously exceed the adaptive threshold, marking the boundary of the risk area. When the three indicators simultaneously exceed the threshold, the module calculates the coupled field of "density entropy × velocity mutation" and calculates its spatial gradient. It innovatively regards the gradient direction as the congestion propagation direction, and the gradient modulus represents the diffusion intensity. Through hexagonal grid connectivity clustering, the connected area with the largest gradient modulus value is marked as the risk boundary, achieving high-precision extraction of the congestion boundary. The trajectory prediction module generates continuous virtual trajectories for obscured or short-term lost targets based on their last known position, velocity direction, and grid density distribution through a "density-constrained trajectory prediction model." If the position error exceeds the limit during identification and re-identification, the model parameters are automatically adjusted. During the prediction process, the module monitors the potential intersection risk between the virtual trajectory and the trajectories of other surrounding targets in real time, and uses a "lateral avoidance vector" to fine-tune the acceleration or direction. This intersection risk detection with other people's predicted trajectories, combined with an obstacle avoidance mechanism, makes the virtual trajectory closer to the real crowd movement pattern in multi-person mixed traffic scenarios. If the error exceeds the tolerance during target re-identification, the system uses an error loss function to automatically adjust the key model coefficients of "density sensitivity parameter," "velocity response coefficient," and "obstacle avoidance weight" based on the grid density and velocity information at the moment of maximum error.The diversion control module performs a spatial comparison between the virtual predicted trajectory and the marked risk area boundary grid. If an overlap is found, a "directional diversion instruction" is triggered, and the gate release frequency is dynamically adjusted in real time based on the negative correlation principle between the congestion propagation vector and the gate release direction. By using the grid index to accelerate the comparison, it is possible to determine whether the predicted path enters the risk area boundary through a single grid mapping, rather than a point-by-point and pixel-by-pixel comparison, which greatly improves real-time performance. Once an intersection is found, the module automatically calculates the coordinates of the first intersection of the risk boundary and the trajectory, and selects the "nearest safe grid" in the grid near that point as the guidance target, directly generating a concise and clear "offset direction" and "avoidance prompt"; this strategy avoids complex path planning and can quickly guide personnel to deviate based on the nearest safe point, which is helpful for emergency response in the event of sudden congestion.

[0046] In summary, this application has at least the following effects: A dynamic crowd flow identification method and system based on multi-sensor fusion uses an environmental interference matrix and dynamic weight assignment to adaptively fuse radar, thermal imaging, and piezoelectric sensors in complex scenarios, significantly improving the reliability and accuracy of multi-source data. When thermal imaging fails and radar can only detect speed changes, the gait spectrum characteristics of piezoelectric sensors are used for cross-validation, effectively reducing missed detections and false alarms in occluded areas. A dynamic hexagonal grid-based multi-metric joint analysis of density entropy, velocity mutation, and flow direction consistency accurately identifies local congestion patterns and risk boundaries, enabling rapid location and early warning of potential congestion areas. A density-constrained trajectory prediction model, driven by density gradients and real-time obstacle avoidance correction, effectively maintains continuous tracking of occluded targets, significantly improving the reliability and continuity of predicted trajectories. By determining the spatial overlap of virtual trajectories with risk areas and adjusting the negative correlation between the congestion propagation vector and gate release direction, this system achieves proactive crowd flow guidance and dynamic gate frequency control, enhancing the system's responsiveness and control effectiveness in sudden congestion scenarios.

[0047] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0048] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0049] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0050] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0051] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0052] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A dynamic identification method of human traffic based on multi-sensor fusion, characterized by: The following steps are involved: S1. Construct an environmental interference matrix based on the electromagnetic attenuation rate of the radar, the illumination mutation factor of the thermal imaging, and the mechanical noise factor of the piezoelectric sensor. Dynamically calculate the weight value of each sensor based on the inverse relationship between the factors in the matrix. S2. When the density weight of the thermal image in the obstructed area is lower than a first threshold and the velocity weight of the radar in the corresponding area is higher than a second threshold, the piezoelectric sensor's gait spectrum characteristics are used to verify the target's presence and output weight-calibrated fused density, velocity, and flow direction data. S3. Map the fused data to dynamic grid cells and calculate the density entropy value, velocity mutation index, and flow consistency index for each grid. When all three indicators exceed the adaptive threshold, the density entropy and velocity mutation gradient are fused to generate the congestion propagation direction vector and mark the risk area boundary. S4. For the occluded target, a virtual trajectory is generated by a density-constrained trajectory prediction model based on the target's last known position, velocity vector, and grid density distribution data. The model parameters are adjusted when the target re-identification position error exceeds the tolerance value. S5. The predicted virtual trajectory and the risk area boundary are spatially overlapped. If there is an intersection, a directional diversion instruction is triggered. At the same time, the gate release frequency is dynamically adjusted according to the negative correlation principle based on the spatial gradient direction of the congestion propagation vector.

2. The method for dynamic identification of pedestrian flow based on multi-sensor fusion according to claim 1 is characterized by: Based on the electromagnetic attenuation rate of the radar, the illumination mutation factor of the thermal imaging, and the mechanical noise factor of the piezoelectric sensor, an environmental interference matrix is constructed, which includes the following steps: The attenuation slope of the radar spectrum curve is calculated as the electromagnetic attenuation rate factor, the overall amplitude of the grayscale change of pixels in adjacent frames of thermal imaging is analyzed as the illumination mutation factor, and the energy of the piezoelectric sensor vibration signal in the characteristic frequency band of human gait is accumulated to generate the mechanical noise factor. The three factors are integrated into a two-dimensional matrix, and the instantaneous interference states of different sensors are correlated through matrix operations.

3. The method for dynamic identification of human flow based on multi-sensor fusion according to claim 2 is characterized by: The dynamic calculation of the weight value of each sensor according to the inverse relationship of each factor in the matrix includes the following steps: Establish a mapping relationship between interference factors and sensor weights. According to the numerical value of each factor in the environmental interference matrix, the weight coefficient is assigned according to the exponential inverse relationship. The weights of radar, thermal imager and piezoelectric sensor are dynamically allocated through preset adjustment rules. The logic of the adjustment rules is as follows: when the electromagnetic attenuation rate factor is detected to be increased, the radar weight is increased; when the illumination mutation factor is enhanced, the thermal imaging weight is reduced; when the mechanical noise factor exceeds the limit, the piezoelectric sensor weight is reduced.

4. The method for dynamic identification of pedestrian flow based on multi-sensor fusion according to claim 3 is characterized by: Step S2 includes the following steps: The monitoring area is pre-divided into several sensing units. When the weighted density value collected by the thermal imaging device in a unit is lower than a set threshold and the weighted velocity value collected by the radar in the unit is higher than another set threshold, the system triggers the piezoelectric sensor array deployed in the unit to collect gait vibration signals; The collected signals are subjected to spectrum analysis to filter out the frequency band energy that matches the characteristics of human steps. Then, the three types of observation results, thermal imaging, radar, and piezoelectricity, are multiplied by the corresponding weights and weighted fusion is performed to output the fused crowd density, movement speed, and main flow direction of the unit.

5. The method for dynamic identification of human flow based on multi-sensor fusion according to claim 4 is characterized in that: The fused data is mapped to dynamic grid cells, and the density entropy value, velocity mutation index, and flow consistency index of each grid are calculated, including the following steps: The monitoring area is divided into dynamically scalable hexagonal grids, and the density entropy value is calculated based on the information entropy of the temperature distribution within the grid to quantify the disorder of crowd gathering; The velocity mutation index is generated by normalizing the absolute value of the difference of velocity vectors of adjacent frames to capture the motion instability characteristics; The flow consistency index is calculated based on the inverse normalization of the directional angle variance to identify the stagnation trend of the crowd.

6. The method for dynamic identification of human flow based on multi-sensor fusion according to claim 5 is characterized by: When the three indicators exceed the adaptive threshold at the same time, the density entropy and the gradient of the speed mutation are integrated to generate the congestion propagation direction vector and mark the risk area boundary, including the following steps: Only when the density entropy value, velocity mutation index and flow consistency index simultaneously exceed the adaptive threshold, the spatial gradient of the product of density entropy and velocity mutation is calculated; A congestion propagation vector with a length proportional to the gradient modulus is generated along the gradient direction to indicate the congestion diffusion direction and intensity. At the same time, the continuous connected grid with the largest gradient modulus is marked as the boundary of the risk area.

7. The method for dynamic identification of human flow based on multi-sensor fusion according to claim 6 is characterized by: For the occluded target, a virtual trajectory is generated using a density-constrained trajectory prediction model based on the target's last known position, velocity vector, and grid density distribution data. The model includes the following steps: Obtain the target's last position coordinates and corresponding velocity direction before being blocked, and then construct a density field constraint based on the density distribution of the target's grid and surrounding grids. Based on the target's current position and speed direction, guided by the density gradient, the possible acceleration or deceleration trend of the target is calculated. On this basis, continuous predicted position points are generated in a time series. During the prediction process, the intersection risk of the target and the surrounding people's trajectory is monitored in real time. If the predicted path is found to intersect with the paths of others, the predicted speed or direction is adjusted through the obstacle avoidance algorithm to keep the trajectory consistent with the actual crowd behavior.

8. The method for dynamic identification of human flow based on multi-sensor fusion according to claim 7 is characterized by: When the target re-identification position error exceeds the tolerance value, the model parameters are adjusted, including the following steps: When the re-identification position deviation exceeds the allowable range, a loss function for the trajectory error is established, the grid density and velocity information when the error occurs are statistically analyzed, and the main factors of the prediction deviation are analyzed; According to the analysis results, the density-sensitive parameters, speed response coefficients and obstacle avoidance weights in the trajectory prediction model are increased or decreased.

9. The method for dynamic identification of human flow based on multi-sensor fusion according to claim 8, characterized in that: Step S5 includes the following steps: The generated virtual predicted trajectory is spatially compared with the marked risk area boundary grid, the predicted path is mapped to the grid coordinate system and it is determined whether it overlaps or contacts with the risk area outline; If it is determined that there is a spatial intersection, the coordinates of the intersection of the risk boundary and the trajectory are determined, and based on this, the nearest safe guidance position is selected to generate guidance instructions; Obtain the congestion propagation vector field calculated in all grids in the current time slot and identify the directionality of the vector in the grid where the gate is located; If the direction is opposite to the normal output direction of the gate, an instruction to reduce the gate release frequency will be automatically issued. If the vector direction is consistent with the release direction, the release frequency will be increased.

10. A dynamic identification system for human traffic based on multi-sensor fusion, applied to the dynamic identification method for human traffic based on multi-sensor fusion according to any one of claims 1 to 9, characterized in that: It includes the following modules: interference weight calculation module, occlusion target verification module, grid feature extraction module, trajectory prediction module, and grooming control module; The interference weight calculation module is used to construct an environmental interference matrix based on the electromagnetic attenuation rate of the radar, the illumination mutation factor of the thermal imaging, and the mechanical noise factor of the piezoelectric sensor, and dynamically calculate the weight value of each sensor according to the inverse relationship between the factors in the matrix; The occluded target verification module is used to call the gait spectrum characteristics of the piezoelectric sensor to verify the existence of the target when the density weight value of the thermal imaging in the occluded area is lower than the first threshold and the speed weight value of the radar corresponding area is higher than the second threshold, and output the weight-calibrated fused density, speed and flow direction data; The grid feature extraction module is used to map the fused data to dynamic grid cells, calculate the density entropy value, velocity mutation index and flow consistency index of each grid, and when the three indicators simultaneously exceed the adaptive threshold, fuse the density entropy and velocity mutation gradient to generate the congestion propagation direction vector and mark the risk area boundary; The trajectory prediction module is used to generate a virtual trajectory for the occluded target based on the target's last known position, velocity vector and grid density distribution data through a density-constrained trajectory prediction model, and adjust the model parameters when the target re-identification position error exceeds the tolerance value; The diversion control module is used to determine the spatial overlap between the predicted virtual trajectory and the risk area boundary. If there is an intersection, a directional diversion instruction is triggered. At the same time, the gate release frequency is dynamically adjusted according to the negative correlation principle based on the spatial gradient direction of the congestion propagation vector.

Images