An Adaptive Perimeter Monitoring Method Based on Cooperative Sensing of Microwave and LiDAR

Abstract

This invention provides an adaptive perimeter monitoring method based on microwave and lidar collaborative sensing, relating to the fields of security monitoring and perimeter protection technology. The method involves deploying microwave and lidar arrays in the monitored perimeter area, and constructing a static 3D baseline model of the environment by combining data from both types of radar. The raw microwave radar signals are filtered to extract target distance and velocity information, while the lidar point cloud is denoised and smoothed, outputting a synchronous multi-source data stream. An attention mechanism feature fusion network is constructed, mapping the preprocessed features of the two types of radar to channel and spatial attention weights respectively, and then weighted and fused. A lightweight target detection model is then used for inference, outputting target classification results, confidence levels, and location information in real time. Real-time environmental parameters are adjusted according to a preset mapping relationship, adjusting the weights of the two types of radar in the fusion network and the lidar scanning parameters, while dynamically updating the target detection confidence threshold. When the target confidence level exceeds the threshold and the positioning error meets the standard, a tiered alarm is triggered.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of security monitoring and perimeter protection, in particular to a self-adaptive perimeter monitoring method based on microwave and laser radar collaborative sensing. BACKGROUND

[0002] In the existing perimeter security field, single sensor monitoring or simple multi-sensor parallel use is the mainstream solution, but they have a series of inherent and difficult to overcome defects when facing complex and variable real environment.

[0003] Traditional visible light camera is the most widely used device, but it is severely dependent on environmental lighting and weather conditions. In low-visibility conditions such as night, rain, snow, fog, etc., the image quality decreases sharply and the target is difficult to identify. In addition, it can only provide two-dimensional texture and color information, lacks accurate distance and speed data, is easily disturbed by light changes, leaf shaking, etc., and has a very high false alarm rate.

[0004] Microwave radar has good ranging and speed measuring capabilities and strong environmental penetration, and is not affected by light and weather. However, its biggest shortcoming is that the angle resolution and target recognition capability are very poor. Radar can usually only detect that an object is moving, but it is difficult to accurately determine the specific contour, size and category of the object. This leads to the fact that when it is used alone, either the real threat is missed, or a large number of false alarms are generated due to the inability to identify.

[0005] Laser radar can generate accurate three-dimensional point clouds, provide rich target geometric shape information, and has high recognition accuracy. However, it is vulnerable to adverse weather. Rain, snow, heavy fog, and smoke can severely attenuate laser signals, resulting in a significant reduction in effective detection distance, and even complete failure, making it impossible to achieve stable monitoring all day long. In addition, early high-line laser radars are high in cost, which also limits their large-scale deployment. SUMMARY

[0006] In order to solve the above problems, the present application proposes a self-adaptive perimeter monitoring method based on microwave and laser radar collaborative sensing, which comprises:

[0007] Microwave radar arrays and laser radar arrays are arranged in the perimeter area to be monitored. After the system is started, a static environment three-dimensional baseline model is constructed by combining the data of the two types of radars through a self-learning stage;

[0008] The original signal of the microwave radar is filtered to extract target distance and speed information, and the laser radar point cloud is denoised and smoothed to output synchronized multi-source data streams;

[0009] A fusion network of attention mechanism is constructed to map the features of the two types of radars into channel and spatial attention weights respectively and then fuse them, and a lightweight target detection model is used for inference to output the target classification results, confidence and position information in real time.

[0010] Real-time environmental parameters are used to adjust the weights of the two types of radars in the fusion network and the scanning parameters of the laser radar according to a preset mapping relationship, and the confidence threshold of target detection is dynamically updated.

[0011] The hierarchical alarm mechanism is triggered when the target confidence exceeds the confidence threshold and the positioning error meets the standard.

[0012] Further, the step of constructing a multi-source decision layer fusion algorithm to calculate the secondary confidence before triggering the hierarchical alarm mechanism includes:

[0013] Synchronously acquire the initial confidence C of the current target output by the AI target recognition module initial , the target trajectory data and environmental parameters of multiple detection periods;

[0014] Calculate the trajectory continuity coefficient S trace : If the target displacement deviation of adjacent periods in multiple periods is ≤0.5m, determine the trajectory continuity, S trace =1, otherwise S trace =0;

[0015] Determine the environmental adaptation coefficient K env : If the environment is rain and fog environment, K env =0.8; if the environment is sunny environment, K env =1.0;

[0016] Calculate the secondary confidence C final :

[0017] ;

[0018] Only when C final ≥0.8, the hierarchical alarm mechanism is triggered.

[0019] Further, an adaptive threshold algorithm is used to dynamically update the confidence threshold of target detection, including:

[0020] (1) Collect rainfall, fog, and light intensity to determine rain and fog environment or sunny and strong light environment;

[0021] (2) Generate N initial individuals based on chaotic sequence, and divide them into male and female categories; calculate the fitness of the two categories, update the individual optimal and global optimal; calculate the temperature threshold and search energy Q, and update the position; after iteration meets the standard, output the global optimal;

[0022] (3) Global optimization through mmWave 6G backhaul, edge computing unit replaces parameters, and continuous multiple detection cycles verify that the indicators meet the standard for stable operation.

[0023] Further, the fitness calculation and optimal value recording steps are as follows: the fitness of the value of the jth parameter of the first type search unit f and the value of the jth parameter of the second type search unit m is calculated respectively f(X f,j ), f(X m,j ); the individual optimal value P best and the global optimal value G best are updated, and X food =G best ;

[0024] The fitness of the value X i,j of the jth parameter of the search unit i is as follows:

[0025] ;

[0026] Where: FPR is the false positive rate; mAP is the average precision of multi-target detection; is the system response time delay; is the reference delay threshold.

[0027] Further, the fitness function assigns weights according to =0.35, =0.45, =0.2, and sets the initial parameters according to the environmental quantization results: in a rainy and foggy environment, the initial value of the target detection confidence threshold is 0.85-0.9, and in a sunny and strong light environment, the initial value of the target detection confidence threshold is 0.8-0.85;

[0028] The iteration termination condition is: the current iteration number t≥the maximum iteration number t max or the fitness of the global optimal value G best is improved by ≤0.001 for consecutive multiple times, and the output optimal confidence threshold is adapted in conjunction with the microwave radar weight and the laser radar scanning frequency.

[0029] Further, the process of AI-driven feature fusion and target recognition is as follows: the feature fusion network uses a multi-channel interactive attention module, maps the microwave radar speed feature to a channel weight through 2 layers of multilayer perception and Sigmoid, maps the laser radar point cloud feature to a spatial weight through 3x3 two-dimensional convolution and global pooling, and the two are fused at the feature pyramid level; the target detection is embedded with a convolution block attention module after the feature extraction module, an SE attention block is embedded in the fusion layer, a 3x3 convolution branch is added to the second layer of the feature pyramid for small target detection, and ≥30 frames / second inference and ≥95% average precision mean are achieved.

[0030] ​Further, the microwave radar array and the laser radar array are arranged at a preset interval in the to-be-monitored perimeter area to form a sensor network with overlapping detection areas; after the system is started, a self-learning stage is performed to collect environmental background features, laser radar point cloud data and microwave radar echo statistical data are combined, a three-dimensional baseline model of the static environment is constructed through point cloud voxelization processing, multi-dimensional echo statistics and a coordinate registration method combining GPS coarse registration and ICP fine registration, and the spatial coordinates of the two types of radars are unified to a world coordinate system.

[0031] Further, environmental parameters including at least rainfall, fog degree and light intensity are monitored in real time, the contribution weights of the microwave radar and the laser radar in the feature fusion network are dynamically adjusted according to a preset environmental parameter and sensor weight mapping relationship, and the scanning parameters of the laser radar are adaptively adjusted; based on local statistical features and environmental context, an adaptive threshold algorithm is used to dynamically update the confidence threshold of target detection.

[0032] Compared with the prior art, the present application has the following beneficial effects:

[0033] The present method has all-weather survival capability of the microwave radar and high-precision hole capability of the laser radar through deep fusion of the microwave radar and the laser radar and system-level adaptive optimization. Whether in a sunny day or in a night with extremely low visibility or in severe weather, the system can work stably to realize reliable detection and accurate classification of the intrusion target, and the robustness and environmental adaptability of the monitoring are significantly improved.

[0034] The present method uses an attention mechanism feature fusion network, and a dynamic weighting fusion strategy can always extract the most reliable feature combination for target detection, so that the confidence of the final output target classification result is higher and the false alarm rate is greatly reduced. The lightweight detection model ensures the real-time performance of high-precision reasoning, meets the stringent requirements of perimeter security on immediate response. At the same time, the ability to dynamically update the confidence threshold makes the alarm triggering mechanism more intelligent and flexible, and further filters out accidental false alarms caused by environmental interference or small animals. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A schematic diagram for arranging the microwave radar array and the laser radar array;

[0036] Figure 2 A target identification schematic diagram;

[0037] Figure 3 A median filtering schematic diagram. DETAILED DESCRIPTION

[0038] The application will be described in detail below with specific examples. The following examples will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the application. These are within the scope of protection of the application.

[0039] Example 1

[0040] S1, sensor deployment and environment modeling

[0041] In the perimeter area to be monitored, microwave radar arrays and laser radar arrays are arranged at a preset interval to form a sensor network with overlapping detection areas, as shown in the schematic diagram of Figure 1 ; wherein the microwave radar is used to obtain the radial velocity information of the target and has the ability to penetrate rain and fog, and the laser radar is used to obtain the high-resolution three-dimensional contour information of the target; after the system is started, a self-learning phase is performed to collect and learn the environmental background characteristics, and based on the point cloud data of the laser radar and the echo statistical data of the microwave radar, a three-dimensional baseline model representing the static environment is constructed, and the spatial coordinates of the microwave radar and the laser radar are unified to a world coordinate system through a coordinate registration algorithm, providing a geometric reference for data fusion.

[0042] Specifically, the microwave radar and laser radar arrays are arranged at equal intervals on the perimeter terrain, and the detection radius of a single radar is ≥120m, and the array spacing is controlled to be not more than 100m, so as to ensure that the overlap rate of adjacent coverage areas is between 15%-20%, achieving blind area-free monitoring. The microwave radar uses a 24GHz high-frequency beam, which has strong penetration ability for rain, fog and vegetation, and can achieve accurate velocity measurement on a stationary target. As shown in Figure 1 , it is a schematic diagram of target recognition.

[0043] The laser radar selects a multi-line scanning model with a point cloud density ≥10 pts / m 2 , providing high-resolution contours. After the system is started, a 30-minute self-learning phase is used to continuously collect background noise such as wind, rain and vegetation, and a three-dimensional baseline model is constructed based on point cloud voxelization and radar echo statistics, and spatial coordinates are unified through ICP and timestamp alignment to form a unified world coordinate system, providing a reliable geometric reference for subsequent fusion.

[0044] In the preferred embodiment, when constructing the three-dimensional baseline model of the static environment in the system self-learning phase, the laser radar point cloud and microwave radar echo data need to be combined to achieve reproducible modeling through the following detailed steps:

[0045] (1) Laser radar point cloud voxelization processing

[0046] For the performance parameters in the specification, the laser radar point cloud density is ≥10 pts / m 2 The voxel size is set to 0.2m x 0.2m x 0.2m, which can ensure sufficient number of points in a single voxel, preserving environmental spatial details and avoiding information redundancy. During voxelization, two core features of each voxel are extracted and retained simultaneously: one is the average reflectivity, used to distinguish static objects of different materials such as walls, vegetation, and ground; the other is the number of points, used to filter invalid voxels with no points or less than 2 points, ensuring the effectiveness of the modeling data.

[0047] (2) Microwave radar echo statistical analysis

[0048] Based on the spatial unit divided by voxelization, multi-dimensional statistics are performed on the microwave radar echo data, and the statistical dimensions are defined as: echo intensity variance, Doppler shift distribution standard deviation, and static target echo proportion. When constructing the three-dimensional baseline model, voxels with a static target echo proportion of ≥90% are set as background voxels and included in the baseline model; voxels below this threshold are temporarily excluded to avoid the dynamic interference of temporary obstructions affecting the stability of the baseline.

[0049] (3) Multi-radar coordinate registration implementation

[0050] To unify the spatial coordinates of the microwave radar and the laser radar, a two-step method of GPS coarse registration and ICP fine registration is adopted: the first step is to perform initial alignment based on the GPS coordinates of the installation positions of the two types of radars, with the spatial coordinate deviation controlled within 1m to lay a foundation for fine registration; the second step is to perform ICP algorithm fine registration, with the iteration termination condition set as: iteration number ≥50 times or registration error ≤0.1m, which ensures the coordinate unification accuracy and avoids excessive iteration to waste self-learning efficiency.

[0051] Through the above steps, the spatial features of laser point clouds and the static determination of microwave echoes can be combined to form a precise and reproducible three-dimensional baseline model, providing a reliable geometric reference for subsequent multi-source data fusion.

[0052] The specific implementation parameters of the three-dimensional baseline model construction include: the laser radar point cloud voxelization size is set to 0.2m x 0.2m x 0.2m, and the average reflectivity and point cloud number features are retained during voxelization; the microwave radar echo statistical dimensions are echo intensity variance, Doppler shift distribution standard deviation, and static target echo proportion, and the three-dimensional baseline model construction threshold is set to voxels with a static target echo proportion of ≥90% as background; the ICP coordinate registration adopts a two-step method of GPS coarse registration and ICP fine registration, and the ICP iteration termination condition is iteration number ≥50 times or registration error ≤0.1m.

[0053] S2, multi-source data preprocessing and synchronization

[0054] The original signal of the microwave radar output is filtered to suppress noise, and the distance and speed information of the target is extracted; the point cloud data collected by the laser radar is denoised and smoothed to eliminate isolated noise points while retaining the edge details of the target; a high-precision time synchronization protocol is used to time stamp the data streams of the microwave radar and the laser radar uniformly, to realize the spatio-temporal alignment of the data of the two types of sensors within millisecond accuracy, and to eliminate abnormal data based on a statistical threshold, to output clean, time-synchronized multi-source data streams.

[0055] As shown in Figure 3 , the original IQ stream of the microwave radar output distance and speed is filtered by median filtering to remove isolated spikes, and then filtered by low-pass filtering to suppress high-frequency noise; the laser radar point cloud is smoothed by radius search median smoothing to remove isolated noise points and retain edge details.

[0056] To achieve millisecond-level synchronization, an IEEE1588 PTP hardware timestamp is deployed in the edge computing unit to align the time references of the two types of sensors with a synchronization accuracy of ±50 ns, the error is controlled within 1 ms, meeting the spatio-temporal registration requirements of high-speed targets. After synchronization, the abnormal detection module eliminates drift and electromagnetic interference based on a statistical threshold, and outputs clean data streams with uniform timestamps for the fusion layer.

[0057] In the preferred embodiment, the filtering and denoising processing is performed according to the following parameters:

[0058] 1. Microwave radar signal filtering

[0059] For the one-dimensional distance / speed signal output by the microwave radar, first, a median filter with a window length of 5 is used to remove isolated spike noise in the signal; then, a 4th order Butterworth low-pass filter is used to set the cutoff frequency to 5 Hz, to accurately suppress high-frequency interference caused by rain and fog scattering and electromagnetic radiation, and to retain the effective signal characteristics of the target.

[0060] 2. Laser radar point cloud denoising

[0061] Radius search denoising is performed on the laser radar point cloud, with a radius threshold of 0.3 m. If the number of point clouds in a single search area is ≤3, it is determined to be an isolated noise point and is removed, while retaining the edge details of the target.

[0062] 3. Abnormal data elimination

[0063] Based on the statistical threshold, clean data is selected: distance data exceeding the radar detection radius ±5%, speed data exceeding ±50 m / s, and point cloud reflectivity intensity below 0.1. If any of the conditions is met, the data is eliminated, providing a reliable data basis for subsequent millisecond-level spatio-temporal alignment and feature fusion.

[0064] The microwave radar signal is first filtered by a median filter with a window length of 5 to remove isolated spikes, and then filtered by a 4th order Butterworth low-pass filter to suppress high-frequency noise; the laser radar point cloud denoising adopts a radius threshold of 0.3m radius search, and when the number of point clouds in a single search area is ≤3, it is determined as an isolated noise point and removed; the abnormal data removal criteria are that the distance data exceeds the radar detection radius ±5%, the speed data exceeds ±50m / s, and the point cloud reflection intensity is lower than 0.1, and any one of the conditions is removed.

[0065] S3, AI-driven feature fusion and target recognition

[0066] A feature fusion network based on attention mechanism is constructed, which maps the target speed features extracted from the preprocessed microwave radar data to channel attention weights, and maps the target contour features extracted from the preprocessed laser radar point cloud data to spatial attention weights; on the feature pyramid network, the channel attention weights and the spatial attention weights are concatenated and weighted fused to realize adaptive enhancement of cross-modal features; a lightweight target detection neural network model is used to perform end-to-end inference on the fused features, and real-time output of target classification results, confidence and position information containing at least personnel, vehicles and animals is realized.

[0067] The fusion network adopts a multi-channel interactive attention module, which maps the speed features of the radar to channel weights, and maps the contour features of the laser point cloud to spatial attention, and concatenates and splices them on the feature pyramid to realize adaptive weighting of cross-scale features. On this basis, lightweight YOLOv5-S is introduced and CBAM / SE attention blocks are embedded, which significantly improves the mAP of small target detection to more than 95%, while maintaining a real-time inference speed of ≥30FPS, meeting the edge power limit. The model completes end-to-end inference on the edge GPU, realizing multi-class real-time classification and pose estimation of personnel, vehicles and animals.

[0068] In the preferred embodiment, the multi-channel interactive attention is specifically implemented as follows:

[0069] Velocity feature channel weight W c Mapping formula:

[0070] ;

[0071] wherein Norm(V) represents standardizing the target velocity vector V extracted by the microwave radar, the dimension = 32, corresponding to 32 tracking targets, MLP is a 2-layer fully connected network, input 32 dimensions, hidden 64 dimensions, output fusion network channel number, Sigmoid activation function.

[0072] Contour feature spatial attention W s Mapping formula:

[0073] ;

[0074] Wherein, P is the BEV feature map generated by the laser radar point cloud, Conv2d is a 3*3 convolution layer, the output channel number = 1, and Pool is a global average pooling. The general process of BEV laser radar perception mainly converts point cloud data into BEV representation through two branches.

[0075] YOLOv5-S and attention block embedding mode: embedding CBAM attention block after C3 module and before SPPF module of YOLOv5-S, embedding SE attention block in the Neck layer of the feature pyramid, and the specific channel number is consistent with the channel number of the corresponding layer feature map, preferably, the output of C3 is 80 dimensions, and the compression ratio of the SE attention block is set to 16.

[0076] In the preferred embodiment, for small targets with a size of ≤0.5 m, one 3*3 convolution branch is added to the P2 layer of the feature pyramid to improve the small target feature extraction capability.

[0077] S4, adaptive strategy dynamic adjustment

[0078] Real-time monitoring of environmental parameters, the environmental parameters at least including rainfall, fog degree and illumination intensity; according to a preset mapping relationship between environmental parameters and sensor weights, the contribution weights of the microwave radar and the laser radar in the feature fusion network are dynamically adjusted, wherein in rainy and foggy weather conditions, the weight of the microwave radar is increased and the scanning parameters of the laser radar are adaptively adjusted to compensate for its performance degradation; at the same time, based on local statistical features and environmental context, an adaptive threshold algorithm is used to dynamically update the confidence threshold of target detection to reduce false positives.

[0079] The system continuously monitors the rainfall, fog degree, temperature and other indicators of the weather station and environmental sensors, and dynamically adjusts the sensor weights according to a preset weather mapping table: in rainy and foggy conditions, the weight of the microwave radar is increased to 0.7 and the scanning frequency of the laser radar is increased to 20 Hz to offset the point cloud degradation; in sunny and strong light conditions, the radar weight is reduced and the laser resolution is increased to 14 PPT. The threshold is updated in real time by an adaptive threshold algorithm to prevent false positives. In addition, the attribute weight is fine-tuned by an adaptive updating mechanism based on group consensus decision, so that the weight converges within a few minutes, and combined with 6G low-latency backhaul, the cloud model is quickly iterated and the edge parameters are immediately issued.

[0080] In the preferred embodiment, an optimization algorithm is used.

[0081] The environmental parameter quantitative standard is a rainy and foggy environment with rainfall ≥0.5 mm / h or fog degree ≤500 m, and a sunny and strong light environment with illumination intensity ≥10000 lux, rainfall =0 mm / h and fog degree ≥2000 m.

[0082] The optimization algorithm category size N is dynamically set according to the perimeter length, and N increases by 10 for every 1 km perimeter, the initial N = 20, and it is adapted to 1-2 km perimeter; the maximum N = 40, and it is adapted to perimeter within 4 km; the three consensus indexes of the group consensus decision mechanism are that the false alarm rate decreases by ≥5%, the mAP increases by ≥2%, and the time delay is ≤5 ms, the condition for triggering weight fine-tuning is that ≥2 indexes meet, and the fine-tuning step is = 0.05 x current fitness improvement rate; the 6G backhaul and edge coordination parameters are to use 28 GHz millimeter wave frequency band, and the end-to-end time delay is ≤20 ms, the edge computing unit receives parameters and completes replacement within 1 ms, and the effectiveness is verified within 3 detection periods after replacement, and each period = 0.1 s.

[0083] S5, alarm and closed-loop feedback

[0084] When the target confidence output by the target recognition step exceeds the dynamically updated confidence threshold and the positioning error meets the preset condition, a hierarchical alarm mechanism is triggered; the hierarchical alarm mechanism at least includes local alarm, edge side alarm push and central platform alarm and secondary verification; the system continuously collects alarm results, environmental noise data and false alarm samples, and returns these data as feedback signals to the model training end, triggers the incremental learning and parameter optimization of the model, so as to realize the closed-loop adaptive optimization from perception, decision to model adjustment, and continuously improve the robustness and accuracy of the perimeter monitoring system.

[0085] When the target confidence output by the fusion network is >0.85 and the positioning error is ≤3 m, a three-level alarm is triggered:

[0086] ① Local audible and visual alarm; ② Edge gateway pushes video pop-up window; ③ Central decision platform generates event work order and calls multi-source decision layer fusion algorithm for secondary verification, and re-evaluates the confidence using historical trajectory and environmental parameters. The closed-loop feedback adopts a closed-loop flowchart of parameter optimization, and real-time returns alarm results, environmental noise, false detection rate and other indexes to the model training library, triggers incremental learning and updates weights, realizes adaptive improvement of system robustness, and the entire link is completed within 5 ms, and the end-to-end time delay from detection to alarm meets the high safety demand.

[0087] Preferably, a multi-source decision layer fusion algorithm is constructed: input current target confidence C initial , trajectory continuity and environmental adaptation coefficient of the last 5 detection periods, calculate secondary confidence, and only when C final ≥0.8, confirm the alarm, and reduce the false alarm risk of single confidence determination.

[0088] The secondary confidence calculation adopts a weighted fusion formula, which comprehensively considers the current target detection result, trajectory stability and environmental adaptability, and the formula is as follows:

[0089] ;

[0090] C final is the secondary confidence, that is, the final alarm decision basis; C initial is the initial confidence of the current target output by the AI target recognition module, the value range is [0, 1]; S trace is the trajectory continuity coefficient of the last 5 detection periods, discrete value, 1 when continuous, 0 when discontinuous; K env : environmental adaptation coefficient, 0.8 in rain and fog environment, 1.0 in sunny environment, consistent with the environmental quantization standard of the adaptive strategy dynamic adjustment module.

[0091] Embodiment 2

[0092] The optimization algorithm step is used for parameter adaptive adjustment of the perimeter monitoring system. The optimization algorithm outputs the optimal sensor parameters through chaos initialization, species division, fitness calculation, position update and iterative optimization.

[0093] 1. Chaos initialization species:

[0094] Generate N initial individuals based on the Logistic chaos sequence, N is preferably 20-40, and the value of the jth parameter of each individual i (i.e. search unit i) X i,j is randomly generated in the range [X min ,X max ], ensuring that the species species distribution uniformly covers the search space, enhancing the diversity and ergodicity of the initial species.

[0095] 2. Species gender division:

[0096] Randomly divide the species into first species search unit species X f,j and second species search unit species X m,j in a 1:1 ratio. This division simulates the behavior of gender in the population, which helps to balance exploration and development.

[0097] 3. Initial fitness calculation and optimal value recording:

[0098] Calculate the fitness f(X f,j ) and f(X m,j ) of the first species search unit and the second species search unit, respectively; update the historical optimal position P best of each individual, that is, the position corresponding to the historical optimal fitness of itself and the global optimal value G best , that is, the position corresponding to the optimal fitness of all individuals, set the best food position X food =G best .

[0099] The fitness function is used to evaluate the pros and cons of individuals, and the core association perimeter monitoring detects the accuracy, false alarm rate and response speed:

[0100] ;

[0101] Among them: , preferably , , , focusing on detection effect; FPR is the false alarm rate; mAP is the average precision of multi-target detection; System response delay ≤5ms, =5ms, reference delay threshold.

[0102] 4. Iterative loop:

[0103] a. Calculate temperature threshold and search energy:

[0104] Calculate temperature threshold :

[0105] ;

[0106] Search energy ;

[0107] In the formula: c1=0.5 is a fixed coefficient, t is the current iteration number, t max is the maximum iteration number.

[0108] b. Search energy judgment:

[0109] If Q<0.25, execute global optimization: for each individual, update the position to

[0110] ;

[0111] If not, Q≥0.25, go to step c.

[0112] c. Temperature threshold judgment:

[0113] If T temp >0.6, execute close to search position, update position at t+1 :

[0114] ;

[0115] In the formula: is the target position, "+" indicates random selection of positive or negative to enhance randomness, and c3=0.2 is the default step size coefficient;

[0116] If T temp ≤0.6, go to step d.

[0117] d. Random number determination:

[0118] Generate random number r rand .

[0119] If r rand >0.6, perform a local fine-grained search to avoid local optima, position at time t. Update position at time t+1 for:

[0120] ;

[0121] In the formula: X best The current individual's optimal value P best At the corresponding position, c4=0.1 is the chaotic perturbation coefficient;

[0122] If r rand ≤0.6: Perform a global search, expand the search area, and update the position using the following formula:

[0123] ;

[0124] In the formula: c5=0.3, which is the global search coefficient.

[0125] e. New location fitness update:

[0126] For all updated positions, calculate the new fitness f(X). i,j (t+1)).

[0127] Compare each individual's new fitness with its P best If the fitness is better, then update P. best .

[0128] Compare the new fitness of all individuals with G best If the fitness is better, then update G. best and X food ,X food Always = G best .

[0129] f. Iteration termination judgment:

[0130] If the termination condition is met: t≥t max Or 5 consecutive iterations of G best If the fitness improvement is ≤0.001, exit the loop; if t=t+1 is not satisfied, return to step a to continue iterating.

[0131] 5. Output optimal parameters

[0132] After the iteration terminates, the global optimal solution G is output.best , Corresponding adaptive adjustment parameters: microwave radar weight Wm, laser radar scanning frequency FI, target detection confidence threshold Thresh.

[0133] These optimal parameters are transmitted back to the edge computing unit through 6G low latency to update the sensor operating parameters and detection thresholds, completing the adaptive adjustment of the perimeter monitoring system.

[0134] The optimization algorithm drives the adaptive adjustment of sensor parameters and the 6G low latency backhaul requirement, which needs to be executed according to the following standards and processes to ensure scene adaptability and feasibility:

[0135] 1). Environmental parameter quantification judgment standard

[0136] First, collect rainfall, fog (visibility), and light intensity data through environmental sensors, establish clear quantitative thresholds to divide the scene: rain and fog environment is determined by rainfall ≥ 0.5 mm / h or fog (visibility) ≤ 500 m, in this scene, the microwave radar weight needs to be increased to compensate for the performance degradation of laser; sunny strong light environment is determined by light intensity ≥ 10000 lux and rainfall = 0 mm / h, fog (visibility) ≥ 2000 m, in this scene, the high resolution advantage of laser radar is given priority. The quantification standard provides a scene basis for the optimization algorithm parameter adjustment, avoiding the environmental judgment ambiguity leading to adaptive deviation.

[0137] 2). Optimization algorithm scene adaptation rule

[0138] Based on the above environmental determination results, the key parameters of the optimization algorithm are optimized according to the scene: the size N is dynamically set according to the perimeter length, N increases by 10 every 1 km of perimeter, the initial N = 20, and the maximum N = 40, balancing the optimization accuracy and edge algorithm power consumption; the group consensus decision mechanism takes the decrease of false alarm rate ≥ 5%, the mAP improvement ≥ 2%, and the time delay ≤ 5ms as the three consensus indicators, when ≥ 2 indicators meet, the attribute weight fine-tuning is triggered, the fine-tuning step = 0.05 × current fitness improvement rate, to ensure that the weight adjustment meets the performance requirements of perimeter monitoring.

[0139] 3). 6G backhaul and edge collaboration process

[0140] After the optimization algorithm outputs the optimal parameters, it is landed through cloud 6G edge collaboration: after the cloud model iteration is completed, the 6G base station uses 28GHz millimeter wave frequency band to backhaul parameters, the end-to-end time delay is controlled ≤ 20ms; the edge computing unit receives the parameters and completes the replacement of sensor weight, scanning frequency, etc. within 1ms; after replacement, the effectiveness of the parameters is verified for 3 consecutive detection periods, and the false alarm rate, mAP, and time delay meet the indicators to ensure that the adaptive strategy takes effect in real time. As shown in Table 1, it is the optimization algorithm and perimeter monitoring system adaptation parameter configuration table.

[0141] Table 1 Optimization algorithm and perimeter monitoring system adaptive parameter configuration table

[0142] Parameter name Recommended value range [c1 (fixed coefficient)] 0.5 (fixed value) [c3 (search step coefficient)] 0.1~0.3 [c4 (chaotic disturbance coefficient)] 0.05~0.15 [c5 (global search coefficient)] 0.2~0.4 [c6 (optimization extension coefficient)] 0.3~0.5 t max (maximum number of iterations) 20~40 r rand (random number) [0, 1] uniform distribution chaos (chaotic initial value) (0, 1) and ≠ 0.25, 0.5, 0.75 Microwave radar weight - rain and fog environment 0.7 (fixed initial value) Microwave radar weight - clear and strong light environment 0.3~0.5 Laser radar scanning frequency - rain and fog environment 20 Hz (fixed initial value) Laser radar scanning frequency - clear and strong light environment 10~15 Hz Detection confidence threshold - rain and fog environment 0.85~0.9 Detection confidence threshold - clear and strong light environment 0.8~0.85 (false positive rate, FPR, weight) 0.35 (detection accuracy mAP weight) 0.45 (response latency τ weight) 0.2

[0143] Example 3

[0144] The steps of the multi-source decision layer fusion algorithm are as follows:

[0145] 1. Input data acquisition

[0146] 1.1 Synchronously acquire three types of core data from each module of the system:

[0147] 1.1.1 Initial confidence C initial : directly read the target classification confidence output by the AI-driven feature fusion module, such as the confidence value of the personnel and vehicle recognition result;

[0148] 1.1.2 Target trajectory data: extract the target positioning coordinates of the last 5 detection periods, each period = 0.1s, consistent with the 6G backhaul verification period, and unify them to the world coordinate system;

[0149] 1.1.3 Environmental parameters: call the environmental monitoring module data, determine the environmental type according to the quantitative standard, rain and fog environment: rainfall ≥ 0.5mm / h or fog degree ≤ 500m; sunny environment: illumination intensity ≥ 10000lux and rainfall = 0mm / h, fog degree ≥ 2000m, determine the value of K env .

[0150] 2. Trajectory continuity coefficient S trace calculation

[0151] 2.1 Perform continuity determination based on historical trajectory data:

[0152] 2.1.1 Calculate the target displacement deviation of adjacent detection periods, i.e. the Euclidean distance of coordinates between the i-th period and the i-1-th period;

[0153] 2.1.2 If all adjacent displacement deviations in the last 5 periods are ≤0.5m, determine that the trajectory is continuous, and assign S trace =1;

[0154] 2.1.3 If any adjacent period displacement deviation is >0.5m, such as target sudden disappearance or false target jump, determine that the trajectory is discontinuous, and assign S trace =0.

[0155] 3. Secondary confidence calculation and alarm determination

[0156] 3.1 Substitute the formula to calculate C final , if C final≥0.8, alarm trigger threshold, confirming the target as a valid intrusion target and triggering hierarchical alarms: local audible and light alarms, edge side pushing, and secondary verification on the central platform;

[0157] 3.2 If C final <0.8, determining as a suspected false alarm, not triggering the alarm, and putting the sample into the fuzzy sample pool for incremental learning, with a confidence interval of 0.7-0.85, for subsequent model optimization.

[0158] 3.3 The algorithm integrates multi-dimensional probability information through Bayesian inference, supplements the limitations of single confidence judgment, significantly reduces the false alarm risk caused by rain and fog interference and transient noise, and improves the system robustness.

[0159] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the present application. In the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A microwave and laser radar cooperative perception adaptive perimeter monitoring method, characterized in that, The application relates to a microwave radar and laser radar fusion system for perimeter monitoring. Microwave radar arrays and laser radar arrays are arranged in a to-be-monitored perimeter area; after the system is started, a static environment three-dimensional baseline model is constructed by combining two types of radar data in a self-learning stage; The original microwave radar signal is filtered to extract target distance and speed information, and the laser radar point cloud is denoised and smoothed to output synchronous multi-source data streams; An attention mechanism feature fusion network is constructed, the preprocessed features of the two types of radars are respectively mapped into channel and spatial attention weights and are weighted and fused, and then a lightweight target detection model is used for reasoning to output target classification results, confidence and position information in real time; Real-time environmental parameters are used to adjust the weight of the two types of radars in the fusion network and the scanning parameters of the laser radar according to a preset mapping relationship, and the confidence threshold of target detection is dynamically updated. When the target confidence exceeds the confidence threshold and the positioning error meets the standard, a hierarchical alarm is triggered. Before the hierarchical alarm is triggered, a multi-source decision layer fusion algorithm is constructed to calculate a secondary confidence, and the steps include: Synchronously acquire the current target initial confidence C output by the AI target recognition module initial , target trajectory data and environment parameters of multiple detection periods; S = 1 if the trajectory continuity criterion is met trace : S = 1 if the trajectory continuity criterion is met trace : S = 1 if the trajectory continuity criterion is met trace : S = 1 if the trajectory continuity criterion is met Determine the environmental adaptation coefficient K env If the environment is rain and fog environment, K env = 0.8; if the environment is sunny environment, K env = 1.0; Computing the quadratic confidence C final : ； Only if C final ≥ 0.8, the hierarchical alarm mechanism is triggered.

2. The microwave and lidar co-aware adaptive perimeter monitoring method of claim 1, wherein, An adaptive threshold algorithm is used to dynamically update the confidence threshold of target detection, and the steps include: (1) Rainfall, fog density and light intensity are collected to determine whether the environment is rainy and foggy or sunny and bright; (2) N initial individuals are generated based on a chaotic sequence and are divided into two types; the fitness of the two types is calculated, and the individual optimal value and the global optimal value are updated; a temperature threshold and a search energy Q are calculated, and the position is updated and optimized; after iteration, the global optimal value is output. (3) The global optimal value is returned through a millimeter wave 6G, the edge computing unit is replaced with parameters, and stable operation is realized after a plurality of detection cycles.

3. The microwave and lidar co-aware adaptive perimeter monitoring method of claim 2, wherein, The fitness calculation and optimal value recording steps are as follows: the fitness of the value of the jth parameter of the first type search unit f and the value of the jth parameter of the second type search unit m is calculated respectively, f(X f,j ), f(X m,j ); the individual optimal value P best and the global optimal value G best are updated, and X food =G best ; The search unit searches for a value X of the jth parameter of the ith unit i,j The fitness of the ith unit is calculated as follows: The calculation method is as follows: ； Wherein: FPR is the false positive rate; mAP is multi-object detection average precision; is a system response latency; is a reference latency threshold.

4. The microwave and lidar co-aware adaptive perimeter monitoring method of claim 3, wherein, The fitness function assigns weights to = 0.35, = 0.45, = 0.2, and sets initial parameters according to environmental quantification results: the initial value of the target detection confidence threshold is 0.85-0.9 under a rain and fog environment, and the initial value of the target detection confidence threshold is 0.8-0.85 under a sunny and strong light environment. The iteration termination condition is: current iteration number t≥maximum iteration number t max Or the fitness of the global optimal value G best of consecutive multiple times is less than or equal to 0.001, and the output optimal confidence threshold is adapted in linkage with the microwave radar weight and the laser radar scanning frequency.

5. The microwave and lidar co-sensed adaptive perimeter surveillance method of claim 1, wherein, The feature fusion network adopts a multi-channel interactive attention module, the microwave radar speed feature is mapped into channel weight through 2 layers of multilayer perceptron and Sigmoid, the laser radar point cloud feature is mapped into spatial weight through 3*3 two-dimensional convolution and global pooling, and the two are fused in a feature pyramid cascade; the target detection is embedded with a convolution block attention module after a feature extraction module, an SE attention block is embedded in a fusion layer, and a 3*3 convolution branch is added to the second layer of the feature pyramid for small target detection.

6. The microwave and lidar co-aware adaptive perimeter monitoring method of claim 1, wherein, Microwave radar arrays and laser radar arrays are arranged in a to-be-monitored perimeter area at a preset interval to form a sensor network with overlapping detection areas; after the system is started, a self-learning stage is performed to collect environmental background features, laser radar point cloud data and microwave radar echo statistical data are combined, a three-dimensional baseline model of a static environment is constructed through point cloud voxelization processing, echo multidimensional statistics and a coordinate registration method combining GPS coarse registration and ICP fine registration, and the spatial coordinates of the two types of radars are unified to a world coordinate system.

7. The microwave and lidar co-aware adaptive perimeter monitoring method of claim 1, wherein, Real-time environmental parameters including rainfall, fog density and light intensity are monitored, the contribution weights of the microwave radar and the laser radar in the feature fusion network are dynamically adjusted according to a preset environmental parameter and sensor weight mapping relationship, and the scanning parameters of the laser radar are adaptively adjusted; based on local statistical features and environmental context, an adaptive threshold algorithm is used to dynamically update the confidence threshold of target detection.

Images