Pedestrian detection, light intelligent control method and system

Abstract

The present application relates to the technical field of pedestrian detection, and especially relates to a pedestrian detection and light intelligent control method and system. The present application optimizes the matching cost mechanism of the Hungarian algorithm, fuses the motion cost and deformation compensation cost of the human body point cloud cluster to complete the initial pedestrian matching, and divides to obtain the human body point cloud cluster subsets of matching success and matching failure. The present application configures independent Kalman filters for each human body point cloud cluster, predicts the next frame pedestrian motion state based on the historical motion parameters, combines the pedestrian motion distance, direction similarity and environmental obstacle information to reconstruct the motion cost, and secondarily uses the Hungarian algorithm to check and correct the matching result. Meanwhile, the present application combines the historical trajectory features of pedestrians and the surrounding environment features to adaptively optimize the process noise covariance of Kalman filtering, constructs the pedestrian motion state based on the position, speed and acceleration, effectively improves the pedestrian detection accuracy, and optimizes the office light intelligent linkage control effect.

Description

Technical Field

[0001] This invention relates to the field of pedestrian detection technology, and in particular to a pedestrian detection and intelligent lighting control method and system. Background Technology

[0002] With the increasing demand for intelligent and energy-saving office solutions, office lighting control has evolved from traditional manual switches and timer controls to intelligent linkage where lights turn on when someone is present and turn off when they leave. However, existing technologies still have many pain points that urgently need to be addressed:

[0003] Functional limitations of existing control methods: Current mainstream intelligent office lighting control solutions mainly rely on two technologies: infrared sensing and camera recognition. Infrared sensing is significantly affected by ambient temperature, and in hot summers or heated winters, it is prone to false triggering (lights turning on automatically when no one is present) or missed triggering (lights not turning on when someone is present). Furthermore, it can only detect the presence of moving objects and cannot distinguish between pedestrians and non-target actions such as moving furniture or dropping documents. While camera recognition can accurately detect pedestrians, it poses a serious risk of privacy breaches. Employee work activities and document information can be easily collected, failing to meet the privacy protection requirements of office environments. Moreover, in backlit, shadowed, or multi-person obstruction scenarios, the recognition accuracy drops significantly (typically below 80%). Additionally, cameras require continuous power and data storage, resulting in high hardware and maintenance costs.

[0004] 2D LiDAR has been applied in industrial inspection and low-speed autonomous driving scenarios due to its advantages of strong anti-interference (unaffected by light and temperature), good privacy protection (only outputting distance data, no image information) and moderate cost. However, its application in the field of office lighting control still has obvious technical shortcomings: existing LiDAR pedestrian detection algorithms mostly use the Hungarian algorithm to match the currently detected target with the tracked target. These algorithms generally rely only on the single feature of the target's spatial location distance to construct the matching cost, thereby completing the association and pairing of human point cloud clusters in previous and subsequent frames.

[0005] Relying solely on spatial distance as a single feature to construct matching costs has significant limitations. In office settings, with dense pedestrian traffic, frequent close-range occlusion, and overlapping point cloud clusters corresponding to different pedestrians, mismatches and omissions are common. Simply relying on distance to determine matching relationships easily leads to these problems. Insufficient accuracy in pedestrian detection and target matching directly causes misjudgments in lighting control logic, resulting in issues such as lights turning on unnecessarily, delayed shutdown, or lights not turning on when they should, significantly reducing the actual effectiveness of intelligent office lighting control. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of existing LiDAR pedestrian detection schemes based on the Hungarian algorithm, which only use a single feature to construct the matching cost, are prone to target matching deviation and insufficient pedestrian detection accuracy, thus affecting the effect of intelligent lighting control.

[0007] To solve the above-mentioned technical problems, the present invention provides a pedestrian detection method, comprising: S1: Initialize a=1; where a represents the detection frame index; S2: Obtain the human point cloud cluster set of the target region in frame a and frame a+1. Based on the actual position of each human point cloud cluster in the human point cloud cluster set in frame a+1 and its distance from the nearest obstacle, and the actual position of each human point cloud cluster in the human point cloud cluster set in frame a, calculate the motion cost of each pair of human point cloud clusters. S3: Based on the aspect ratio of the bounding rectangle of each human point cloud cluster in the (a+1)th frame and its distance from the nearest obstacle, and the aspect ratio of the bounding rectangle of each human point cloud cluster in the a-th frame, calculate the deformation compensation cost of each pair of human point cloud clusters. S4: Using the matching cost matrix of the Hungarian algorithm, based on the motion cost and deformation compensation cost of each group of human point cloud clusters, the matching cost of each group of human point cloud clusters is calculated, thereby obtaining the first human point cloud cluster subset that is successfully matched and the second human point cloud cluster subset that is unmatched in the human point cloud cluster set of frame a+1. S5: Update a=a+1 and return to step S2 for real-time pedestrian detection.

[0008] Preferably, the formula for calculating the deformation compensation cost of each group of human point cloud clusters is as follows: , in, The cost of deformation compensation for human body point cloud clusters. Centering human point cloud clusters The width of the bounding rectangle, Centering human point cloud clusters The width of the bounding rectangle, , This represents the total number of human point cloud clusters in frame a. , This represents the total number of human point cloud clusters in frame a+1. Centering human point cloud clusters The length of the circumscribed rectangle, Centering human point cloud clusters The length of the circumscribed rectangle, Centering human point cloud clusters Relative to the human point cloud cluster in frame a The instantaneous rate of change of the aspect ratio of the circumscribed rectangle. , The maximum deformation penalty constant, It is an exponential function with the natural constant as its base. Centering human point cloud clusters The actual position and the distance between its nearest obstacle. This is the threshold for the width of the physical feature.

[0009] Preferably, after obtaining the first successfully matched subset of human point cloud clusters and the second unmatched subset of human point cloud clusters in the (a+1)th frame human point cloud cluster set, the method further includes: Based on the first human point cloud cluster subset in frame a, the position of each human point cloud cluster in frame a and frame a-1, and its velocity and acceleration between the two frames, the predicted position of each human point cloud cluster in frame a+1 and the predicted velocity between frame a and frame a+1 are obtained. Based on the actual positions of each human point cloud cluster in frame a and frame a+1, the instantaneous velocity vector of each human point cloud cluster is calculated. Based on the actual and predicted positions of each human point cloud cluster in frame a+1, the Euclidean distance of each human point cloud cluster is calculated; based on the predicted velocity and instantaneous velocity vector of each human point cloud cluster between frame a and frame a+1, the cosine similarity of the assumed motion direction of each human point cloud cluster is calculated; based on the Euclidean distance, the cosine similarity of the assumed motion direction, and the distance to the nearest obstacle with its actual position, the motion cost of each human point cloud cluster is calculated. Obtain the deformation compensation cost of each human point cloud cluster and its matching human point cloud cluster; Using the matching cost matrix of the Hungarian algorithm, the matching cost of each human point cloud cluster is calculated based on the motion cost and deformation compensation cost of each cluster, thereby verifying the matching result of each cluster.

[0010] Preferably, the process of acquiring the actual position of each human point cloud cluster in frame a includes: The predicted position of each human point cloud cluster in frame a is fused with the actual position to obtain the optimized actual position of each human point cloud cluster in frame a.

[0011] Preferably, the formula for calculating the motion cost of each human point cloud cluster, based on the Euclidean distance of each cluster, the cosine similarity of the assumed motion direction, and the distance to the nearest obstacle to its actual location, is as follows: , ,

[0012] in, For the first The cost of movement for individual human point cloud clusters These are the weighting coefficients. For directional weighting coefficients, , It is an exponential function with the natural constant as its base. To set a safe distance, For the direction weighting coefficient parameter, For the first The Euclidean distance between the actual positions of an individual human point cloud cluster in frame a and in frame (a+1). For the first Cosine similarity of the assumed motion directions of individual human point cloud clusters. for The distance between an individual human point cloud cluster and its nearest obstacle in frame a. For the first The prediction speed of an individual human point cloud cluster between frame a and frame (a+1). For the first The instantaneous velocity vector of an individual human point cloud cluster from frame a to frame a+1. , For time intervals, For the first The actual position of an individual human point cloud cluster in frame a+1. For the first The actual position of an individual human point cloud cluster in frame a. This indicates the calculation of the modulus.

[0013] Preferably, based on the first human point cloud cluster subset in frame a, the method for obtaining the predicted position of each human point cloud cluster in frame a+1 and the predicted velocity between frame a and frame a+1, based on the position of each human point cloud cluster in frame a and frame a-1 and the velocity and acceleration between the two frames, includes: For each human point cloud cluster in the first human point cloud cluster subset of frame a, the Kalman filter corresponding to the human point cloud cluster is used. Based on the position of the human point cloud cluster in frame a and frame a-1, and the velocity and acceleration between the two frames, the predicted position of each human point cloud cluster in frame a+1 and the predicted velocity between frame a and frame a+1 are obtained. The process of obtaining the noise covariance of the Kalman filter for each human point cloud cluster includes: Based on the trajectory curvature and velocity change rate of each human point cloud cluster before the a-th frame, the fusion adjustment factor and noise robustness determination factor are determined. An environmental compensation term is constructed based on the distance from the human point cloud cluster to the nearest obstacle and the density of surrounding targets within a range centered on the human point cloud cluster. Based on the fusion adjustment factor, noise robustness judgment factor, and environmental compensation term, the noise variance of the acceleration process in the x-axis and y-axis directions is calculated. Based on the acceleration process noise variance, acceleration noise weight baseline value, and velocity noise weight baseline value in the x-axis and y-axis directions, the velocity process noise variance in the x-axis and y-axis directions is obtained. By integrating the variances of position, velocity, and acceleration processes along the x and y axes, a process noise covariance is constructed.

[0014] Preferably, the formula for the environmental compensation item is: , in, For environmental compensation items, The obstacle influence coefficient. The target density influence coefficient, It is an exponential function with the natural constant as its base. The distance from the human figure point cloud cluster to the nearest obstacle. Distance threshold The density of surrounding targets within a range centered on the human body point cloud cluster is set; Based on the acceleration process noise variance, acceleration noise weight baseline value, and velocity noise weight baseline value in the x-axis and y-axis directions, the formula for obtaining the velocity process noise variance in the x-axis and y-axis directions is as follows: , , in, The variance of the velocity process noise in the x-axis direction. The velocity noise weighting baseline value in the x-axis direction. Let Variance be the noise variance of the acceleration process in the x-axis direction. The x-axis direction is used as the baseline value for acceleration noise weighting. The variance of the velocity process noise in the y-axis direction. The velocity noise weighting baseline value in the y-axis direction. Let Variance be the noise variance of the acceleration process in the y-axis direction. The acceleration noise weighting baseline value is located in the y-axis direction.

[0015] Preferably, the method for obtaining the human point cloud cluster set of the target region in the a-th frame includes: Use lidar to obtain the original laser points in the target area of ​​frame a; Clustering and segmentation are performed based on the dynamic clustering radius of each original laser point in the target region of frame a to obtain several point cloud clusters, and human point cloud clusters are identified in several point cloud clusters. The formula for calculating the dynamic clustering radius of each original laser point is as follows: , in, The dynamic clustering radius of the original laser points. This is the distance divergence compensation coefficient. This represents the absolute physical distance between the original laser point and the lidar. This is the environmental repulsion weighting coefficient. The obstacle repulsion coefficient of the original laser point. The distance is determined based on the distance between the original laser point and the obstacle with the smallest distance.

[0016] The present invention also provides a method for intelligent lighting control, comprising: The pedestrian detection method described above is used to detect pedestrians in the target area; based on the detection results and real-time illuminance data, the lighting of the lights is controlled.

[0017] The present invention also provides a lighting intelligent control system, comprising: lighting; A natural light sensor, connected to a lighting lamp, is used to acquire real-time illuminance data; LiDAR is used to acquire raw laser points in a target area. The host computer communicates with the lidar and natural light sensor, and is coupled to the lighting lamp to implement the steps of the above-mentioned intelligent lighting control method.

[0018] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0019] This invention discloses a pedestrian detection method. It constructs a motion cost by combining the actual position of the human point cloud clusters in consecutive frames with the distance to the nearest obstacle, and characterizes the real motion correlation characteristics of pedestrians by incorporating environmental spatial constraints. Furthermore, it designs a dedicated, adaptively adjustable deformation compensation cost calculation formula. This formula not only captures the point cloud morphological deformation caused by pedestrian posture and limb movements by utilizing the instantaneous change rate of the length, width, and aspect ratio of the bounding rectangle of the human point cloud clusters in consecutive frames, but also introduces a dynamically adjusted deformation penalty coefficient based on the distance to the nearest obstacle, achieving adaptive correction of morphological features in near-obstacle interference scenarios. By fusing motion features under environmental constraints with dynamically adaptive morphological features to construct a comprehensive matching cost, the accuracy of pedestrian detection is effectively improved, thereby optimizing the office lighting control effect.

[0020] This invention, after completing the initial target matching and grouping, adds a refined secondary matching verification mechanism based on motion state prediction. The initial matching relies solely on basic position and morphological features for coarse matching, while the secondary verification introduces predicted motion states. It constructs a multi-dimensional motion cost by combining spatial distance deviation, cosine similarity of motion direction, and obstacle environment constraints. This cost is then combined with morphological deformation compensation costs to reconstruct the matching cost matrix, achieving accurate verification and correction of the initial matching results. This mechanism fully integrates historical pedestrian movement patterns and real-time movement trends, taking into account positional deviation, consistency of motion direction, and environmental interference. It can effectively correct initial matching mismatches and incorrect matches caused by dense pedestrian traffic, sudden posture changes, close-range occlusion, and environmental interference in office scenarios, filtering out unreasonable matching associations and further optimizing target matching accuracy and tracking stability.

[0021] This invention constructs a complete physical state vector from the position, velocity, and acceleration of human point cloud clusters across consecutive frames. An independent Kalman filter is configured for each human point cloud cluster, and an adaptive process noise covariance is designed to suit pedestrian motion characteristics and office environment. The curvature of the pedestrian's historical trajectory and the rate of velocity change are used to quantify the intensity of motion, determining the fusion adjustment factor and noise robustness judgment factor. Simultaneously, an environmental compensation term incorporating obstacle distance and surrounding target density is constructed. An exponential decay relationship characterizes near-distance obstacle constraints, and density parameters characterize the degree of crowding interference, accurately quantifying the impact of environmental disturbances. Based on this, the multi-dimensional process noise variance of acceleration and velocity is solved hierarchically. Finally, multi-dimensional noise parameters are integrated to construct an adaptive process noise covariance, allowing the filter parameters to dynamically adapt in real-time to pedestrian motion state, scene crowding, and obstacle constraints. This effectively improves the prediction accuracy and tracking stability of pedestrian motion state, suppresses prediction biases caused by point cloud jitter, posture deformation, and scene occlusion, and provides high-precision motion prior information for subsequent secondary matching verification, ensuring the accuracy of pedestrian detection. Attached Figure Description

[0022] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0023] Figure 1 This is a flowchart illustrating a pedestrian detection method according to the present invention.

[0024] Figure 2 This is a flowchart of the preprocessing process for the raw laser points collected by the lidar.

[0025] Figure 3 This is a structural diagram of a smart lighting control system.

[0026] Figure 4 It is a moving overview of the office scene and the corresponding office lighting layout diagram. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0028] Reference Figure 1 As shown, this embodiment provides a pedestrian detection method, including: Step S1: Initialize a=1; where a represents the detection frame index; Step S2: Obtain the human point cloud cluster set of the target region in frame a and frame a+1. Based on the actual position of each human point cloud cluster in the human point cloud cluster set in frame a+1 and its distance from the nearest obstacle, and the actual position of each human point cloud cluster in the human point cloud cluster set in frame a, calculate the motion cost of a pair of human point cloud clusters. In the target region of frame a Each human point cloud cluster is located in the target region of the (a+1)th frame. Individual point cloud clusters are combined in pairs to obtain Group of human body point cloud clusters.

[0029] In this embodiment, preferably, the method for obtaining the human point cloud cluster set of the target region in the a-th frame includes: Use lidar to obtain the original laser points in the target area of ​​frame a; Clustering and segmentation are performed based on the dynamic clustering radius of each original laser point in the target region of frame a to obtain several point cloud clusters, and human point cloud clusters are identified in several point cloud clusters. The formula for calculating the dynamic clustering radius of each original laser point is as follows: , in, The dynamic clustering radius of the original laser points This is the distance divergence compensation coefficient. This represents the absolute physical distance between the original laser point and the lidar. This is the environmental repulsion weighting coefficient. The obstacle repulsion coefficient of the original laser point. The distance is determined based on the distance between the original laser point and the obstacle with the smallest distance.

[0030] like Figure 2 As shown, Figure 2 This is a flowchart of the preprocessing process for the raw laser points collected by the lidar.

[0031] In this embodiment, preferably, the original laser points of the target region in the a-th frame are preprocessed to obtain the original target laser points of the target region in the a-th frame, including: A combined approach of statistical filtering and median filtering is used to filter noise from the original laser points. This includes: firstly, statistical filtering is used to remove isolated noise points from the original laser radar point cloud, specifically based on a preset threshold for the number of neighboring points. If the number of points in a single point's neighborhood is less than five, that point is identified as a noise point and removed. Then, median filtering is used to smooth the point cloud data, selecting the median distance between points within a 3×3 neighborhood to complete the calculation, thereby eliminating random noise interference. After the above processing, the overall signal-to-noise ratio of the point cloud can be improved to over 30dB.

[0032] Static background removal is implemented to filter out point cloud data of fixed obstacles in the environment. LiDAR point cloud data of the target area in an unoccupied state is collected in advance. Point cloud models of fixed obstacles within the area, such as office partitions, desks, and lockers, are created and stored as a background baseline model. During real-time pedestrian detection, the point cloud data collected in the current frame is compared and matched pixel-by-pixel with the pre-stored static obstacle models. Fixed point cloud data that overlaps with the static background model is removed, and only dynamic point cloud data that changes position within the scene is retained, which are the potential pedestrian target point clouds. This processing method effectively avoids false target problems caused by distortion of static obstacle edge point clouds and interference from fixed environmental facilities, preventing static obstacles from being misidentified as pedestrian targets and further improving the accuracy of dynamic human point cloud cluster extraction.

[0033] The point cloud data after static background removal is subjected to coordinate transformation, including converting the distance and angle polar coordinate data output by the lidar into planar rectangular coordinate data. At the same time, a global coordinate system for the target area is built based on the actual installation location of the lidar. The location of one lidar is selected as the origin of the coordinate system. The horizontal coordinate axis is set to be parallel to the long side of the target area and the vertical coordinate axis is parallel to the short side of the target area. The data collected by multiple lidars is fused and processed based on the unified coordinate system.

[0034] Since the coordinates of the point cloud data acquired by lidar are all based on the lidar's own coordinate system, in order to uniformly map all point cloud data to the global coordinate system of the target area, it is necessary to perform coordinate transformation processing on each data point within the point cloud. The coordinate transformation process first performs a coordinate system rotation transformation. The following uses the coordinate transformation of a two-dimensional plane unit vector as an example for illustration, and the relevant coordinate symbols are defined as follows: This invention defines three types of two-dimensional planar coordinate systems: the first is the radar coordinate system, with the physical center of the lidar as the origin, used to characterize the original point cloud coordinates of the lidar; the second is the base coordinate system, with the installation center of the lidar in the monitoring area as the origin, used to characterize the reference coordinates of the radar equipment relative to the installation position; and the third is the world coordinate system, that is, the global fixed coordinate system of the target area, which serves as the reference coordinate system for the unified fusion of point cloud data across the entire field.

[0035] The coordinate system with the center of the lidar as the origin is denoted as the radar coordinate system. ; The coordinate system with the center of the lidar installation location as the origin is denoted as the base coordinate system. ; World coordinate system is denoted as ; The rotation angle from the source coordinate system to the target coordinate system (counterclockwise is positive) is denoted as the global fixed coordinate system rotation angle. ; The offset from the origin of the source coordinate system to the origin of the target coordinate system is denoted as the translation vector. .

[0036] The transformation from the radar coordinate system to the base coordinate system is used to correct for angular and positional deviations caused by radar installation. If the radar is mounted on a robot, there will be an angular deviation from the robot's orientation. Let the radar installation deviation angle be θ. Let represent the fixed rotational deviation of the radar coordinate system relative to the base coordinate system; let the translation vector be a two-dimensional vector composed of the lateral offset and the longitudinal offset, representing the positional deviation of the radar physical center relative to the installation reference center. First, the original point cloud coordinates in the radar coordinate system are rotated to eliminate the influence of the installation angle offset. The rotation transformation formula is as follows: , in, These are the original point cloud coordinates in the radar coordinate system. The x-coordinate of the original point cloud in the radar coordinate system. The original point cloud coordinates in the radar coordinate system. These are the coordinates in the base coordinate system after rotation correction. The x-coordinate of the base coordinate system after rotation correction. The ordinate of the base coordinate system after rotation correction. .

[0037] By superimposing translation compensation on the rotation transformation, we obtain the complete transformation formula from the radar coordinate system to the base coordinate system: , in, This represents the lateral installation offset of the radar coordinate system relative to the base coordinate system. This represents the longitudinal installation offset of the radar coordinate system relative to the base coordinate system. Introducing a homogeneous coordinate system, the homogeneous matrix expression of the above composite transformation is as follows: , It can be simplified to matrix form: ,in, This is the transformation matrix from the radar coordinate system to the base coordinate system. These are homogeneous coordinates in the radar coordinate system. The coordinates are homogeneous coordinates in the base coordinate system.

[0038] After completing the single radar reference calibration, it is necessary to further map all radar reference coordinate systems to the global world coordinate system. Let the global orientation angle of the radar reference (or robot platform) in the world coordinate system be denoted as . The position coordinates are The transformation matrix from the base coordinate system to the world coordinate system is... The format is:

[0039] in, Using the origin as the reference point, the global horizontal coordinates are... The global longitudinal coordinates are based on the origin.

[0040] Through the above multi-level coordinate transformation, the local point cloud data collected by all lidar can be uniformly converted to the global world coordinate system of the target area, completely eliminating the spatial deviation caused by different installation angles and positions of the lidar, ensuring that the point cloud data of multiple lidars and the human point cloud cluster data of previous and subsequent frames have a unified spatial reference, and providing an accurate data foundation for subsequent target matching, trajectory tracking and pedestrian status calculation.

[0041] This invention can use a multi-dimensional identification method that integrates target motion features, contour morphology features, spatial distribution features, and temporal motion stability features to perform layer-by-layer screening and verification of several point cloud clusters generated in the target area, accurately identify and extract human point cloud clusters in the scene, and eliminate various invalid non-pedestrian point cloud interferences.

[0042] In complex office scenarios characterized by high pedestrian activity, numerous static office equipment, and diverse types of clutter, single recognition features are prone to false positives and false negatives. Therefore, this invention combines the pedestrian movement patterns and target contour characteristics specific to office scenarios to construct a hierarchical, multi-dimensional, and accurate judgment system. This enables efficient and accurate screening of human point cloud clusters in office environments. The specific judgment rules are as follows: First, the target's movement speed is determined. Specifically, the target's movement distance is calculated by the positional change of point cloud clusters in two consecutive frames. The real-time movement speed of the point cloud clusters is solved by combining the fixed frame interval of the device. Based on the difference in movement speed, autonomous pedestrian targets and various non-pedestrian dynamic interference targets are distinguished to achieve preliminary screening.

[0043] To address the complex office environment characterized by numerous interferences and fixed target types, this invention precisely defines a general speed threshold based on specific scenarios, adapting it to the unique interference features of the office environment: The effective movement speed range for pedestrians is set to 0.3 m / s to 1.5 m / s. Simultaneously, two types of typical non-target interference speed boundaries are defined: low-speed displacement targets (below 0.2 m / s) correspond to passive movement behaviors such as manually moving desks and chairs or pushing equipment within the office; high-speed instantaneous movement targets (above 2 m / s) correspond to sudden interference behaviors such as falling documents, falling debris, or slipping objects. Only point cloud clusters whose movement speed falls within the specific effective range for office pedestrians can proceed to the next stage of contour feature determination.

[0044] Next, target contour feature determination is performed, distinguishing pedestrians from static office objects based on the planar contour features corresponding to the fixed scanning height of the LiDAR. The LiDAR scanning height is set to 1.8m to 2.2m. At this height, the pedestrian point cloud cluster exhibits a stable rectangular contour shape, thus limiting the pedestrian contour size thresholds: the effective range for contour width is 0.4m to 0.8m, and the effective range for contour length is 0.3m to 0.6m. By solving for the minimum bounding rectangle of the point cloud cluster, the actual contour size of the target is obtained. The size parameters are verified, effectively eliminating static object interference such as desks wider than 1m and chairs narrower than 0.3m that do not conform to the pedestrian contour features. Targets whose size meets the preset thresholds are judged as candidate pedestrian targets.

[0045] Simultaneously, an effective detection distance threshold is set to avoid interference from the radar equipment's own detection blind zone and invalid data at long distances. This invention sets a minimum detection distance of 0.5m to avoid distorted data interference caused by the radar's near-field blind zone; and sets a maximum detection range of 5m, performing final validity judgment only on candidate targets within the 0.5m to 5m effective detection range, further narrowing the effective target selection range and improving detection accuracy.

[0046] To further accurately distinguish between pedestrians and passively moving static interference objects, this invention introduces a dual verification mechanism of motion continuity and contour stability. For candidate dynamic targets after initial screening, verification is performed based on temporal trajectory data from three consecutive frames with a frame interval of 0.1 seconds: if the target's movement trajectory is a straight line or a gentle curve, conforming to the normal walking motion pattern of pedestrians, and the target's contour size shows no significant abrupt changes within consecutive frames, the target is determined to be a valid pedestrian point cloud cluster; if the target's movement trajectory is chaotic and irregular, or if the contour size shows drastic abrupt changes within consecutive frames, it is determined to be a non-pedestrian dynamic interference target such as a transported container or dragged debris, and this type of point cloud cluster is eliminated, ultimately achieving accurate identification and screening of human point cloud clusters.

[0047] It should be noted that, limited by the inability of single-frame point cloud data to provide temporal motion information, this invention, in the initial human point cloud cluster extraction stage, relies solely on contour size features, effective detection distance range, and static background removal to complete the preliminary identification of candidate human point cloud clusters, without relying on temporal features such as velocity, acceleration, and motion trajectory. The target motion velocity determination and the dual verification mechanism of motion continuity and contour stability are both post-matching delayed screening and error correction mechanisms, requiring multiple frames of historical data after the initial association matching of point cloud clusters in consecutive frames has been completed and a continuous tracking trajectory has been formed. Therefore, this invention adopts a layered logic of first coarse identification and then precise verification. It first rapidly extracts potential pedestrian targets through static morphological features, and then uses temporal motion features and adaptive matching costs to complete a secondary calibration of the matching results. This ensures the real-time performance of single-frame detection while solving the problems of false detection and inaccurate matching in single-morphological recognition, significantly improving the overall accuracy of pedestrian detection and tracking in complex office scenarios.

[0048] In this embodiment, preferably, based on the actual position of each human point cloud cluster in the (a+1)th frame and its distance to the nearest obstacle, and the actual position of each human point cloud cluster in the a-th frame, the motion cost of a pair of human point cloud clusters is calculated, including: In the initial inter-frame matching stage, the motion cost is calculated based on the spatial positional relationship of the human point cloud clusters in frame a and frame a+1, combined with environmental constraints. It should be noted that the initial matching relies only on two frames of raw observation data and has not yet formed a stable tracking trajectory or Kalman prediction state. Therefore, it is impossible to obtain information on predicted velocity and motion direction. Consequently, the motion cost of the initial matching does not introduce directional similarity constraints to avoid matching failure due to missing parameters.

[0049] For each pair of human point cloud clusters, this invention constructs a dynamic motion cost calculation logic based on environmental perception modulation. It deeply couples the motion cost weights with the distance to the nearest obstacle, adapting to the scenario characteristics of initial matching without prediction, and achieving adaptive matching in complex office obstacle scenarios. In this stage, motion cost is constructed solely based on two frames of measured position data and environmental obstacle data, without introducing temporal prediction parameters such as speed and direction. The specific formula for calculating the motion cost in the initial matching is as follows:

[0050] in, The cost of movement for human point cloud clusters These are the weighting coefficients. Centering human point cloud clusters Human body point cloud cluster The Euclidean distance between their actual positions. , This represents the total number of human point cloud clusters in frame a. , This represents the total number of human point cloud clusters in frame a+1. , The direction weighting coefficients are set as nonlinear functions modulated by range: , Centering human point cloud clusters The distance to its nearest obstacle To set a safe distance, These are the directional weighting coefficient parameters.

[0051] When human point cloud clusters are aligned with human point cloud clusters Get close to obstacles (such as ), When the distance is less than the safe distance, The modulation coefficient decays exponentially, reducing the penalty weight for positional distance deviation and adapting to the real motion characteristics of pedestrians who need to avoid obstacles and have small positional shifts. When pedestrians move away from obstacles, the modulation coefficient approaches its maximum value, using positional distance deviation as the core matching criterion to ensure matching accuracy in open scenes.

[0052] The initial matching motion cost design is fully adaptable to initial matching scenarios without historical trajectories or predicted states. It completes matching evaluation solely based on two frames of measured positions and real-time environmental obstacle information, ensuring the feasibility of single-frame initial matching. Furthermore, by dynamically modulating the cost weights based on obstacle distances, it addresses the issue of mismatches that easily occur when pedestrians are close to partitions or moving between dense workstations. In the subsequent matching verification stage, with the availability of multiple frames of historical trajectories and Kalman prediction speeds, cosine similarity of motion directions is introduced to achieve high-precision secondary error correction, forming a hierarchical matching system characterized by "coarse matching without parameter constraints and fine verification with multi-dimensional constraints."

[0053] Step S3: Based on the aspect ratio of the bounding rectangle of each human point cloud cluster in the (a+1)th frame and its distance to the nearest obstacle, and the aspect ratio of the bounding rectangle of each human point cloud cluster in the a-th frame, calculate the deformation compensation cost for a pair of human point cloud clusters. The formula is as follows: , in, The cost of deformation compensation for human body point cloud clusters. Centering human point cloud clusters The width of the bounding rectangle, Centering human point cloud clusters The width of the bounding rectangle, , This represents the total number of human point cloud clusters in frame a. , This represents the total number of human point cloud clusters in frame a+1. Centering human point cloud clusters The length of the circumscribed rectangle, Centering human point cloud clusters The length of the circumscribed rectangle, Centering human point cloud clusters Relative to the human point cloud cluster in frame a The instantaneous rate of change of the aspect ratio of the circumscribed rectangle. , The maximum deformation penalty constant, It is an exponential function with the natural constant as its base. Centering human point cloud clusters The actual position and the distance between its nearest obstacle. The physical feature width threshold. , The aspect ratio of the bounding rectangle of the human body point cloud cluster.

[0054] This invention combines the physical characteristics of pedestrian obstacle avoidance in office scenarios to achieve dynamic contour matching compensation. Traditional matching methods relying solely on the minimum bounding rectangle are prone to matching failures due to abrupt changes in contour shape when pedestrians sideways to avoid obstacles. Therefore, this invention introduces the contour deformation rate as a constraint term, combined with the distance between the target and the obstacle. It can determine the area where human point cloud clusters are located, especially when pedestrians are in narrow passageways, When the value is small, Approaching 0. At this point, the system judges the sudden change in contour size as a normal deformation caused by compliant obstacle avoidance actions such as side-stepping and turning, weakens the penalty weight corresponding to contour changes, and effectively solves the problem of excessive rigidity in conventional circumscribed rectangle matching rules.

[0055] Step S4: Using the matching cost matrix of the Hungarian algorithm, based on the motion cost and deformation compensation cost of each group of human point cloud clusters, calculate the matching cost of each group of human point cloud clusters, thereby obtaining the first successfully matched human point cloud cluster subset and the second unmatched human point cloud cluster subset in the human point cloud cluster set of frame a+1; complete the association matching of human point cloud clusters in consecutive frames. The successfully matched consecutive frame point cloud clusters are continuous frame point clouds of the same pedestrian, which can continue the historical trajectory of the pedestrian and update and generate a complete pedestrian motion trajectory in real time. Based on this, the first human point cloud cluster subset with completed trajectory continuation and successful matching in frame a+1 is obtained; the unmatched human point cloud clusters in frame a+1 are new pedestrians or interference targets, constituting the second unmatched human point cloud cluster subset, waiting for subsequent frame association verification to achieve continuous pedestrian tracking; This invention employs the Hungarian algorithm to achieve globally optimal association matching of human point cloud clusters between consecutive frames. For the point cloud cluster pair formed by the tracked human point cloud cluster in frame a and the real-time detected human point cloud cluster in frame a+1, a dimension-adapted matching cost matrix is ​​constructed. This invention overcomes the limitation of traditional algorithms that only use a single distance feature to construct the cost, by fusing motion features and morphological features to construct a two-dimensional composite matching cost. The specific matching cost is obtained by weighted fusion of motion cost and deformation compensation cost; based on the motion cost of each pair of human point cloud clusters... Deformation compensation cost The matching cost of each group of human point cloud clusters is calculated using the following formula: , in, For human body point cloud clusters Human body point cloud cluster The matching cost of the human body point cloud clusters. As the weight of the cost of movement, The weight is the cost of deformation compensation.

[0056] Determine whether the matching cost of each human point cloud cluster pair is less than a set threshold. If it is less, add it to the first human point cloud cluster subset that has been successfully matched in the human point cloud cluster set of frame a+1. Otherwise, add it to the second human point cloud cluster subset that has failed to match.

[0057] In this embodiment, preferably, a unique identifier lifecycle management is performed on the detected human point cloud clusters to achieve stable tracking and state updates of pedestrian targets in consecutive frames, specifically as follows: During the initialization phase, a unique tracking ID is assigned to each valid pedestrian target detected for the first time in the target area. This ID is used to bind historical feature information such as the position, shape, and motion trajectory of the corresponding human point cloud cluster. In subsequent frame matching processes, if the human point cloud cluster calculated using the Hungarian algorithm affects the matching cost... Below the preset cost threshold If the target in consecutive frames is determined to be the same pedestrian, the location information and motion feature parameters bound to the unique tracking ID are updated to continuously iterate the target state. If a human point cloud cluster corresponding to a certain tracking ID fails to match for two consecutive frames, it is determined that the human point cloud cluster is temporarily occluded or temporarily out of the scanning field of view, marked as temporarily lost, and its motion position is continuously predicted using Kalman filtering to maintain tracking continuity. If the tracking ID fails to complete a valid match for five consecutive frames, it is determined that the human point cloud cluster has left the target area, the human point cloud cluster is confirmed to have disappeared, and the corresponding unique tracking ID is released, completing the full lifecycle management of the pedestrian target. This mechanism can effectively solve the tracking breakage and ID jump problems caused by short-term pedestrian occlusion and entry / exit from the field of view, further improving the tracking stability and detection continuity in multi-pedestrian scenarios.

[0058] This invention addresses the problems of predicted trajectories penetrating static obstacles and prediction results diverging during the Kalman continuous extrapolation prediction phase when the target is temporarily lost. It adds a prediction coordinate truncation and anchoring mechanism based on the topological boundary of the static environment, which is only activated during the prediction period when the target is temporarily lost. Let the trajectory prediction coordinates of the human point cloud cluster be... Static obstacles occupy space as The set of boundaries of the legal passage area is Space collision detection verification was performed to obtain the corrected trajectory prediction coordinates. : , in, for The coordinates of the point on the [the map].

[0059] Determine whether the predicted position of the human point cloud cluster in frame a in the world coordinate system of frame a+1 is within the space occupied by static obstacles. If the predicted position falls within the static obstacle area, the original Kalman prediction coordinates are retained; if the predicted position intrudes into the obstacle area, the predicted trajectory is determined to be inconsistent with the physical passage rules, and anchoring correction is required. Based on the pre-constructed static point cloud map, extract the topological boundary of the passable area outside the obstacle. Traverse all discrete boundary points on the boundary, calculate the Euclidean distance between each boundary point and the original predicted position, and select the discrete boundary point with the smallest Euclidean distance as the updated target predicted position.

[0060] When the occlusion is removed and the target is detected again in subsequent frames, the system uses the corrected topological boundary anchoring position as a reference to participate in the inter-frame matching cost calculation. It combines motion cost and deformation compensation cost to complete target re-identification, effectively solving the problems of trajectory divergence and matching misalignment caused by long-term occlusion, and significantly improving the tracking stability and re-identification accuracy in complex office occlusion scenarios.

[0061] In this embodiment, for a scenario of cross-occlusion caused by multiple pedestrians crossing each other, when the point cloud clusters of different people overlap and the point cloud overlap rate is greater than 50%, the instantaneous velocity vector before the occlusion occurs is used as the basis for the Kalman filter output. and Complete the point cloud assignment for the overlapping region. For any laser cluster point within the overlapping region... Calculate the predicted centroid of the point relative to the two human point cloud clusters respectively. and Based on the projection length in their respective velocity directions, the point cloud is assigned to a specific tracking ID corresponding to the human point cloud cluster with the highest projection matching degree, determined by the projection matching degree. Under conditions where the occlusion duration does not exceed 1 second, combining the aforementioned predicted coordinate anchoring with measured data joint matching method, the tracking continuity of the human point cloud cluster can be maintained above 90%; among which, and These are the Kalman instantaneous velocity vectors corresponding to the two sets of interfering human point cloud clusters before the cross-occlusion occurs. and These are the Kalman predicted centroid coordinates of two human point cloud clusters in the current frame; Step S5: Update a=a+1, iterate to complete the point cloud cluster matching and trajectory update of continuous video frames, and realize real-time pedestrian tracking and detection.

[0062] In this embodiment, preferably, after obtaining the first human point cloud cluster subset that successfully matches and the second human point cloud cluster subset that fails to match in the human point cloud cluster set of the (a+1)th frame, the method further includes: Based on the first human point cloud cluster subset in frame a, the position of each human point cloud cluster in frame a and frame a-1, and its velocity and acceleration between the two frames, the predicted position of each human point cloud cluster in frame a+1 and the predicted velocity between frame a and frame a+1 are obtained. Based on the actual positions of each human point cloud cluster in frame a and frame a+1, the instantaneous velocity vector of each human point cloud cluster is calculated. In this embodiment, preferably, the process of obtaining the actual position of each human point cloud cluster in frame a includes: The predicted position of each human point cloud cluster in frame a is fused with the actual position to obtain the optimized actual position of each human point cloud cluster in frame a, for example, by using Kalman filtering for fusion.

[0063] Based on the actual and predicted positions of each human point cloud cluster in frame a+1, the Euclidean distance of each human point cloud cluster is calculated; based on the predicted velocity and instantaneous velocity vector of each human point cloud cluster between frame a and frame a+1, the cosine similarity of the assumed motion direction of each human point cloud cluster is calculated; based on the Euclidean distance, the cosine similarity of the assumed motion direction, and the distance to the nearest obstacle with its actual position, the motion cost of each human point cloud cluster is calculated. Unlike the initial matching which relies solely on the two original observation positions to calculate the cost, the secondary verification has complete historical trajectory and motion state information, and can fuse the predicted state and the measured state to achieve more accurate target association.

[0064] The formula for calculating the motion cost of each human point cloud cluster is: , , in, For the first The cost of movement for individual human point cloud clusters These are the weighting coefficients. For directional weighting coefficients, , It is an exponential function with the natural constant as its base. To set a safe distance, For the direction weighting coefficient parameter, For the first The Euclidean distance between the actual positions of an individual human point cloud cluster in frame a and in frame (a+1). For the first Cosine similarity of the assumed motion directions of individual human point cloud clusters. for The distance between an individual human point cloud cluster and its nearest obstacle in frame a. For the first The prediction speed of an individual human point cloud cluster between frame a and frame (a+1). For the first The instantaneous velocity vector of an individual human point cloud cluster from frame a to frame a+1. , For time intervals, For the first The actual position of an individual human point cloud cluster in frame a+1. For the first The actual position of an individual human point cloud cluster in frame a. This indicates the calculation of the modulus.

[0065] When a pedestrian is detected to be extremely close to a physical barrier... The matching weights decrease exponentially, automatically reducing the cosine similarity of the assumed motion direction, while relatively increasing the confidence of the Euclidean distance. This mechanism breaks the logical deadlock of "range radar has no speed in a single frame" and "trajectory matching strongly depends on speed" from a physical level, perfectly adapting to the mismatch problem caused by pedestrians making minor adjustments to their direction to avoid obstacles when moving between dense workstations.

[0066] Obtain the deformation compensation cost of each human point cloud cluster and its matching human point cloud cluster; After the initial matching is completed, the aforementioned formula for calculating deformation compensation cost is used for the paired human point cloud clusters. The deformation penalty weight is dynamically adjusted by combining the instantaneous change rate of the length, width, and aspect ratio of the bounding rectangle of the human point cloud cluster in the previous and next frames with the distance between the human point cloud cluster and surrounding obstacles. This is used to characterize the degree of change in the target contour shape and serve as the evaluation basis for the secondary verification.

[0067] Using the matching cost matrix of the Hungarian algorithm, the matching cost of each human point cloud cluster is calculated based on the motion cost and deformation compensation cost of each cluster. This allows for the verification of the matching results of each cluster, the identification of invalid pairings, the correction of associations, and the completion of the target matching result verification.

[0068] In this embodiment, preferably, based on the first human point cloud cluster subset in frame a, the position of each human point cloud cluster in frame a and frame a-1, and its velocity and acceleration between the two frames, the method for obtaining the predicted position of each human point cloud cluster in frame a+1 and the predicted velocity between frame a and frame a+1 includes: For each human point cloud cluster in the first human point cloud cluster subset of frame a, the Kalman filter corresponding to the human point cloud cluster is used. Based on the position of the human point cloud cluster in frame a and frame a-1, and the velocity and acceleration between the two frames, the predicted position of each human point cloud cluster in frame a+1 and the predicted velocity between frame a and frame a+1 are obtained. The Kalman filter algorithm used in this invention is used to accurately predict the motion state of the human point cloud cluster in frame a+1. An initial state vector is constructed based on the two-dimensional position and motion velocity of the human point cloud cluster in frame a. The motion state of the next frame is estimated through the state equation. Then, the state parameters are iteratively updated and the prediction error is corrected by combining the actual detection position of the human point cloud cluster in the next frame. This ensures the prediction accuracy of pedestrians in uniform and slowly changing motion states, and the position prediction error can be controlled within 0.1m.

[0069] In response to the characteristics of pedestrian movement in office scenarios, such as low movement speed, smooth trajectory, and multiple small-amplitude obstacle avoidance fine-tuning, this invention specifically optimizes the process noise covariance matrix (Q matrix) of Kalman filtering, weakens the weight of speed change, effectively suppresses inter-frame prediction fluctuations, and significantly improves the tracking stability of human point cloud clusters in complex office scenarios.

[0070] Kalman filtering is the core state prediction and smoothing module of the human point cloud cluster tracking system in this invention. Based on the motion model of the human point cloud cluster, it predicts the ideal motion state of subsequent frames based on the motion state of historical frames, effectively filtering out inherent measurement noise and random errors in the lidar point cloud acquisition process. Compared to traditional matching methods that rely solely on single-frame point cloud features, this solution significantly improves the stability of the human point cloud cluster tracking ID through temporal motion state prediction, eliminating problems such as ID jumps and target loss caused by fluctuations in single-frame point cloud data or short-term data loss.

[0071] Furthermore, to adapt to the complex working conditions of office scenarios with frequent occlusion, pedestrians making small starts, stops, and turns, and point cloud edge jitter, this invention reconstructs a high-dimensional state motion model adapted for human point cloud cluster tracking. Unlike traditional two-dimensional LiDAR tracking algorithms that only use position and velocity to construct a four-dimensional state vector, this invention constructs a six-dimensional state vector. , This represents the x-coordinate of the two-dimensional position of the human body point cloud cluster in the global world coordinate system. This represents the ordinate of the two-dimensional position of the human body point cloud cluster in the global world coordinate system. This represents the velocity component of the human body point cloud cluster along the x-axis. This represents the velocity component of the human body point cloud cluster along the y-axis. This represents the acceleration component of the human body point cloud cluster along the x-axis. This represents the acceleration component of the human body point cloud cluster along the y-axis.

[0072] Traditional four-dimensional state models can only capture changes in position and velocity, failing to adapt to the micro-motion characteristics of pedestrians in office scenarios, such as obstacle avoidance, starting and stopping, and slight turns. They are prone to misjudging low-speed, slightly moving human point cloud clusters as static noise due to point cloud edge jitter and radar scan cycle delays, resulting in target loss. This invention introduces an acceleration dimension parameter, which can accurately map the physical motion characteristics of pedestrians, such as obstacle avoidance, temporary pauses, and slight turns. It can capture subtle changes in human motion posture one radar scan cycle in advance, fundamentally solving the tracking failure problem caused by minute displacement changes in low-speed pedestrian targets.

[0073] This invention selects an approximate uniform / uniform acceleration motion model adapted to the motion characteristics of pedestrians in office settings as the basic motion model. At the same time, it combines the motion law of pedestrians with low speed and low acceleration changes, and performs fine adaptive tuning of the noise covariance matrix (Q matrix) of the Kalman filter process to match various motion modes of human point cloud clusters in office scenarios, thereby further improving the accuracy and robustness of inter-frame motion state prediction.

[0074] The process of obtaining the noise covariance of the Kalman filter for each human point cloud cluster includes: Based on the trajectory curvature and velocity change rate of each human point cloud cluster before the a-th frame, the fusion adjustment factor and noise robustness determination factor are determined. The formula for calculating the fusion adjustment factor is: , in, To integrate adjustment factors, To optimize weighting, the following selection criteria are used. , , The mean curvature of the trajectory of the human body point cloud cluster over multiple consecutive frames. This represents the maximum value of the trajectory curvature. The average velocity change rate of the human body point cloud cluster over multiple consecutive frames. This represents the maximum value of the rate of change of velocity.

[0075] The mean trajectory curvature of a human point cloud cluster over multiple consecutive frames, for example, selecting 5 consecutive frames to cover the mean trajectory curvature over 0.5 seconds, and the process of obtaining the mean velocity change rate of a human point cloud cluster over multiple consecutive frames includes: extracting the motion velocity of the human point cloud cluster over multiple consecutive frames, calculating the velocity change rate between adjacent frames frame by frame, and then taking the arithmetic mean of the velocity change rates of all valid frames to obtain the mean velocity change rate of the human point cloud cluster over multiple consecutive frames.

[0076] For office scenarios, based on a large amount of trajectory statistics... for If the value exceeds this limit, it is considered atypical motion; for .

[0077] Fusion Adjustment Factor It achieves a continuous and smooth transition of adjustment items from 0 to 1, accurately matching the gradual movement changes of pedestrians from steady walking to slow turning, and eliminating the lag of step-like adjustments.

[0078] A trajectory smoothness adjustment term is constructed by integrating adjustment factors. , To adjust the gain, the value is set to 1.0-1.5.

[0079] In this embodiment, to suppress trajectory miscorrection and prediction fluctuations caused by single-frame point cloud drift, random noise, and contour jitter in lidar, a noise robustness determination factor based on multi-frame motion characteristics of human point cloud clusters is constructed. By verifying the trajectory smoothness adjustment term through dual continuous frames using curvature and velocity change characteristics, controllable activation and deactivation are achieved. It is a binary switch quantity, with values ​​of only 0 or 1: This indicates that the movement is considered realistic human motion, and the trajectory smoothness adjustment option is allowed to take effect. If the signal is determined to be radar noise or invalid jitter, the trajectory fine-tuning item is forcibly masked to avoid predictive oscillations caused by pseudo-motion.

[0080] This invention employs a two-layer decision logic consisting of continuous frame amplitude verification and multi-frame trend consistency verification. The specific decision rules are as follows: If the instantaneous trajectory curvature of the human body point cloud cluster across multiple consecutive frames With instantaneous rate of change Any value exceeding its preset maximum limit, and the trend of trajectory curvature change of the human point cloud cluster. Trend of velocity change rate If all values ​​are greater than 0, the noise robustness determination factor is set to 1; otherwise, it is 0.

[0081] The formulas for calculating the trends of trajectory curvature and velocity rate of change of a human point cloud cluster are as follows: , , in, For The curvature of the trajectory at any given moment. The time is the time corresponding to the a-th frame. for Trajectory velocity at any given moment For The time before the end Mean curvature of the frame trajectory For The time before the end The trend of frame rate of change To set the frame rate.

[0082] An environmental compensation term is constructed based on the distance from the human point cloud cluster to the nearest obstacle and the density of surrounding targets within a defined range centered on the human point cloud cluster; the formula for the environmental compensation term is: , in, For environmental compensation items, This is the obstacle influence coefficient, with a value ranging from 0.8 to 1.0. The target density influence coefficient has a value range of 0.5 to 0.7. It is an exponential function with the natural constant as its base. The distance from the human figure point cloud cluster to the nearest obstacle. The preferred distance threshold is 1.0m. The density of surrounding targets is set within a range centered on the human body point cloud cluster.

[0083] Surrounding target density The method for obtaining the data is as follows: taking the two-dimensional position of the human point cloud cluster as the center, a circular determination area is constructed with a predetermined physical radius, the number of other human point cloud clusters that fall into the circular determination area at the same time is counted, and the number is divided by the area of ​​the circular determination area.

[0084] When pedestrians are very close to physical barriers (such as...) When the population is dense, regardless of the first floor Whether the system exhibits a stable performance or not, the speed noise weight will be forcibly increased through an exponentially decaying term.

[0085] Based on the fusion adjustment factor, noise robustness judgment factor, and environmental compensation term, the variance of acceleration process noise in the x-axis and y-axis directions is calculated using the following formula: , , in, Let Variance be the noise variance of the acceleration process in the x-axis direction. Let Variance be the noise variance of the acceleration process in the y-axis direction. This is a fusion adjustment factor, with a value ranging from 0 to 1. To adjust the gain, the value should be between 1.0 and 1.5. For environmental compensation items, The x-axis direction is used as the baseline value for acceleration noise weighting. The y-axis direction acceleration noise weighting baseline value. This is a noise robustness determination factor (a binary switch quantity, with a value of 0 or 1), used to control whether the trajectory smoothness fine-tuning item is effective.

[0086] The above formula for the total control of noise variance in the acceleration process includes the internal motion tendency (time dimension). ), noise verification switch ( ) and external environmental pressures (spatial dimension) Linear weighting and multiplication are combined. When a real adjustment or turning of the human body is detected ( ), triggering internal smoothing adjustment item ( When pedestrians are extremely close to physical barriers or in high-density crowds, regardless of the smoothness of the initial movement, the system will use an exponentially decaying environmental term. The forced excitation increases the noise variance during the speed process, thereby significantly improving the adaptive convergence speed and tracking robustness of the Kalman filter under dynamic obstacle avoidance conditions.

[0087] This invention establishes a hierarchical complementary prediction and correction system through a two-layer adaptive adjustment mechanism and a front-end environmental constraint mechanism. The first layer, a fusion adjustment factor, and the second layer, a noise robustness determination factor, constitute a dual-layer motion adaptive optimization kernel, specifically addressing prediction bias issues in different human motion states. Specifically, the fusion adjustment factor, constructed based on motion and deformation features, forms the first-layer optimization mechanism. This mechanism adapts to the autonomous motion of pedestrians moving at low speeds and making minor posture adjustments, effectively resolving prediction inaccuracies caused by small deformations and slight deviations in posture during slow human movement. The noise robustness determination factor, based on dual verification of trajectory curvature and velocity change rate, forms the second-layer constraint mechanism. This mechanism dynamically distinguishes between genuine pedestrian motion and radar spurious jitter, accurately differentiating effective motion features of pedestrians actively changing speed and making small turns from point cloud noise interference, thus avoiding prediction fluctuations caused by ineffective trajectory correction. This dual-layer motion optimization kernel, combined with an outer constraint mechanism based on static environmental topological boundaries, effectively adapts to pedestrian passive avoidance scenarios under external environmental pressures such as obstacle approach and narrow passage. This hierarchical prediction correction model completely overcomes the trajectory prediction oscillation defects that traditional tracking algorithms are prone to in near-obstacle areas, and stably controls the human point cloud cluster position prediction error within 0.1m in complex office occlusion and dynamic obstacle avoidance scenarios. Simultaneously, to ensure the stability and consistency of the temporal state prediction of the human point cloud cluster, this invention adapts and optimizes the motion noise weights, and performs square root dimensionality reduction fine-tuning of the velocity noise weights following the changes in acceleration weights. This achieves synergistic suppression of multi-dimensional motion parameter noise in velocity and acceleration, further improving the overall stability and accuracy of human point cloud cluster tracking and prediction in complex office scenarios.

[0088] This invention abandons the conventional design that separates trajectory smoothness, measurement noise filtering, and the influence of the physical environment. For complex office scenarios, it pioneers a "spatiotemporal dual-dimensional adaptive" Q-matrix dynamic adjustment model. This model deeply integrates internal motion tendency (time dimension), noise verification switch, and external environmental pressure (spatial dimension), achieving refined calculation of the noise variance of the acceleration process through a single overall control formula. Based on the acceleration process noise variance, acceleration noise weight baseline value, and velocity noise weight baseline value in the x-axis and y-axis directions, the velocity process noise variance in the x-axis and y-axis directions is obtained using the following formula: , , in, The variance of the velocity process noise in the x-axis direction. The velocity noise weighting baseline value in the x-axis direction. Let Variance be the noise variance of the acceleration process in the x-axis direction. The x-axis direction is used as the baseline value for acceleration noise weighting. The variance of the velocity process noise in the y-axis direction. The velocity noise weighting baseline value in the y-axis direction. Let Variance be the noise variance of the acceleration process in the y-axis direction. The y-axis direction acceleration noise weighting baseline value. The acceleration noise weighting benchmark value for smooth motion in an office setting ( As a benchmark value When the judgment Location noise weights Simultaneously increased by 1.2 times.

[0089] By integrating the variances of position, velocity, and acceleration processes along the x and y axes, a process noise covariance is constructed.

[0090] This second embodiment provides a method for intelligent lighting control, including: The pedestrian detection method described above is used to detect pedestrians in the target area; based on the detection results and real-time illuminance data, the lighting of the lights is controlled.

[0091] In this embodiment, optionally, controlling the lighting of the lamps based on the detection results and real-time illuminance data includes: Based on the detection results and real-time illuminance data, control the on / off state and brightness of the lighting fixtures; If the real-time illuminance is lower than the set brightness threshold and there are pedestrians in the human point cloud cluster subset that is successfully matched in the (a+1)th frame, then the lighting is turned on; otherwise, the lighting is turned off.

[0092] Existing LiDAR pedestrian detection algorithms are mostly designed for open outdoor or large-space industrial scenarios, failing to consider the characteristics of office environments, such as "dense static obstacles (e.g., partitions, desks, lockers), slow pedestrian movement speed (0.3-1.5 m / s), and frequent intersections of multiple targets." This leads to LiDAR data easily misidentifying the edges of desks and chairs as pedestrians in office settings, or losing tracking targets due to pedestrians being obscured by partitions, resulting in detection accuracy generally below 75%. Furthermore, existing technologies lack deep linkage logic between LiDAR detection and tracking results and lighting control, only achieving simple on / off control of "whether there are pedestrians in a single area." They cannot achieve refined control such as zoned brightness adjustment or predictive lighting based on the number of pedestrians, dwell time, and movement trajectory, resulting in poor energy efficiency and user experience.

[0093] In summary, existing intelligent office lighting control solutions have three core problems: First, insufficient detection accuracy and scene adaptability, making it unable to cope with complex obstacles and multiple pedestrian crossings in the office; second, difficulty in balancing privacy and cost, with camera solutions infringing on privacy and infrared solutions having low accuracy, and both types of solutions struggling to balance cost and performance; and third, simplistic control logic, lacking the ability to finely adjust to match pedestrian behavior characteristics, resulting in limited energy-saving effects (existing solutions typically achieve energy savings of less than 30%) and poor user experience (such as lights not turning on promptly when pedestrians enter the area, or lights turning off immediately after a brief departure).

[0094] This invention overcomes the technical shortcomings of existing intelligent office lighting control solutions, such as low detection accuracy, insufficient privacy protection, and simple control logic. It provides an intelligent lighting control method that achieves pedestrian detection accuracy ≥95% and multi-target tracking continuity ≥90% (unobstructed scenarios) in office environments, with tracking recovery time ≤1 second in obstructed scenarios. It avoids collecting any image information, fully complying with privacy protection requirements. It establishes refined lighting control logic based on the number of pedestrians, dwell time, and movement trajectory, achieving an energy saving rate ≥45% and a lighting adjustment response delay ≤0.5 seconds. It is compatible with existing office lighting systems, reducing deployment costs and facilitating large-scale promotion.

[0095] This third embodiment provides a smart lighting control system, including: lighting; A natural light sensor, connected to a lighting lamp, is used to acquire real-time illuminance data; LiDAR is used to acquire raw laser points in a target area. The host computer communicates with the lidar and natural light sensor, and is coupled to the lighting lamp to implement the steps of the above-mentioned intelligent lighting control method.

[0096] The LiDAR uses a 2D LiDAR and must possess core performance characteristics suitable for office scenarios: detection range covering a single office area or two adjacent cubicles (effective detection range meeting the requirements for accurate close-range perception); scanning frequency ensuring real-time capture of slow-moving pedestrians to avoid missed detections; angular resolution meeting high-precision differentiation standards, effectively identifying pedestrians and static obstacles such as tables and chairs; minimum target size meeting the differentiation requirements for human bodies and common office facilities; low-power design suitable for long-term continuous operation in office scenarios; and operating voltage compatible with conventional power supply systems; communication interface supporting efficient connection to the processor, with data transmission rate meeting the timeliness requirements of real-time data processing.

[0097] LiDAR deployment follows a zoned coverage principle: in open-plan offices, deployment density is rationally planned according to area to ensure no blind spots; installation employs layered height to avoid obstructions as much as possible, while ensuring simultaneous scanning and detection of employee positions and trajectories in both standing and sitting postures, accurately covering the height range of pedestrians in standing and sitting positions; the installation method uses a planar interval deployment to ensure that employees in every workstation are scanned, and that the scanning range covers the area outside the cubicle. In larger open areas, such as conference rooms and public office areas, which contain multiple lighting zones, an additional set or array of 2D LiDARs is required.

[0098] The host computer possesses high-performance data processing and intelligent control capabilities. It can receive multiple channels of LiDAR point cloud data and ambient lighting data in parallel, efficiently completing all algorithm and data processing tasks, including LiDAR data analysis, human point cloud detection, adaptive Kalman filtering calculation, target matching verification, and intelligent lighting control decisions. This ensures the real-time performance and accuracy of pedestrian detection and lighting control. The host computer supports full-range brightness adjustment logic and two-way data interaction. Based on real-time pedestrian distribution, movement status, and natural light intensity, it adaptively outputs light switch and brightness adjustment commands, and receives feedback on the operating status of lighting equipment, achieving closed-loop precise control of office lighting. The LiDAR and host computer use a direct connection transmission method to ensure stable and low-latency human point cloud data transmission. The natural light sensor is deployed in the office's natural light-sensitive areas, establishing communication with the host computer via a bus to collect and upload ambient light data in real time, providing a reliable environmental basis for adaptive lighting brightness compensation.

[0099] This invention can be implemented using either the above-mentioned integrated centralized control method with a host computer, or a distributed hardware architecture with an independent processor and a lighting controller.

[0100] like Figure 3 As shown, Figure 3This is a structural diagram of an intelligent lighting control system. The processor handles radar data processing and algorithm calculations, while the lighting controller performs dimming, on / off control, and status feedback, all fulfilling the intelligent control functions of this invention. The lighting controller supports bus communication and PWM dimming, covering a dimming range of 0-100%, with a fast response time for rapid control signal response. A single controller must meet the needs of multiple lighting connections, and the communication protocol supports bidirectional data interaction with the processor, enabling lighting status feedback and precise control. A natural light sensor, with a wide range of illuminance measurement capabilities and measurement accuracy meeting the lighting judgment requirements of office scenarios, is deployed in natural light-sensitive areas such as near office windows to collect ambient light intensity data in real time, providing data support for intelligent correction of lighting brightness. The communication module uses a direct connection between the lidar and the processor to ensure stable and efficient data transmission. The processor, lighting controller, and natural light sensor are connected via a bus, with a transmission distance adaptable to the wiring needs of small and medium-sized offices, supporting bidirectional data transmission to achieve precise transmission of device status feedback and control commands.

[0101] The hardware system adopts a layered architecture design: perception layer (2D LiDAR, natural light sensor) → data processing layer (ORIN NX processor) → control execution layer (light controller). The specific connection logic is as follows: the LiDAR transmits raw point cloud data to the processor via USB, and the natural light sensor transmits illuminance data to the processor via RS485 bus; the processor runs detection and tracking algorithms and control logic, generating PWM dimming signals or switching signals, which are sent to the light controller via RS485 bus; after receiving the signals, the light controller controls the switching and brightness adjustment of the corresponding circuit lights, and feeds back the light operating status (such as current brightness, whether there is a fault) to the processor, forming a closed-loop control.

[0102] Based on the aforementioned pedestrian detection results, lighting linkage control is implemented, and normal vectors are set between different sub-areas within the office area. Passing the benchmark Virtual boundary detection lines; Extracting the current centroid position of the human point cloud cluster Instantaneous velocity vector (Obtained by dividing the difference in the centroid positions of two consecutive human point cloud clusters by the time interval), using the formula: Calculate the estimated time for pedestrians to arrive at the virtual boundary detection line. ; Set time threshold The value range is 0.5s to 1.0s to match the hardware PWM dimming duration. In advance, a zone switching signal is issued to the next zone that the pedestrian is about to enter, and a light-on command is sent to achieve the pre-on of the target zone lights. After the pedestrian has completely crossed the virtual boundary detection line, the lights in the original zone are controlled to turn off after a delay, thus completing a smooth transition of the lighting status.

[0103] When three or more pedestrians are simultaneously moving in close intersections within a scene, a three-level progressive conflict resolution algorithm is used to ensure pedestrian tracking accuracy. First, the angle between the instantaneous velocity vectors of different human point cloud clusters is calculated. Prioritize matching velocity angle Targets moving in the same direction and running in parallel with a velocity amplitude difference of less than 10%; for targets that did not complete matching in the first round, the motion cost function is used. Secondary matching is performed by comparing the similarity of trajectory predictions; for targets with a high degree of conflict, spatial location features and deformation cost functions are integrated. A full cost matrix is ​​constructed to complete the joint solution. This hierarchical matching method can effectively avoid the problem of tracking ID confusion in high-density cross-scenes such as conference room entrances, and maintain the pedestrian tracking accuracy in cross-scenes at a stable level of over 85%. Relying on highly reliable pedestrian detection and tracking results, the stable operation of the entire intelligent lighting control system is further guaranteed.

[0104] Based on the pedestrian detection and tracking results of the aforementioned human point cloud clusters, this embodiment combines real-time ambient illuminance data to construct an intelligent office lighting control logic that features multi-condition triggering, fine-tuning, and fault tolerance, thereby enabling smooth start-up and shutdown of lighting equipment, adaptive dimming, and zone priority control.

[0105] The overall office environment is divided into several independent lighting control zones, each covering approximately 20 square meters. The lighting status is determined through a multi-dimensional set of constraints. For pedestrian count triggering, the number of valid human point cloud clusters within a single zone is counted in real time. When the number of pedestrians in the zone is greater than or equal to one, the lighting preparation logic is triggered. When the number of pedestrians in the zone is zero, a delayed light-off timing mechanism is activated to avoid frequent light-on / off fluctuations caused by momentary emptiness. For dwell time triggering, a 2-second zone dwell threshold is set. The dwell time of human point cloud clusters corresponding to the same tracking ID is monitored. The corresponding zone lights are only turned on when the target's continuous dwell time in the zone is not less than 2 seconds. If a pedestrian passes by momentarily and the dwell time is less than 2 seconds, the light-on command is not triggered, effectively filtering short-term pedestrian interference. For zone priority triggering, a zone priority ranking is set according to the functional attributes of the office environment, with the priority order being: meeting room > office area > corridor. When multiple zones simultaneously detect human point cloud clusters, the lighting brightness baseline of the higher priority zone is prioritized, with the meeting room baseline brightness not less than 80% and the regular office area baseline brightness not less than 60%.

[0106] Traditional zoned lighting control relies solely on the instantaneous signal of a pedestrian crossing a physical boundary to switch lights on and off. Limited by hardware communication delays and the physical ramp-up characteristics of PWM dimming, the 0.5 to 1 second hardware response time easily creates the impression of a delayed light cycle, where the lights only gradually illuminate after a pedestrian enters a dark area. This invention breaks through the traditional fixed zoned switching control mode by combining the Kalman timing prediction capability of human point cloud clusters to construct a dynamic follow-up lighting mechanism based on motion feedforward prediction.

[0107] At the predictive feedforward triggering level, the system reads the instantaneous velocity vector and global absolute coordinates of the human point cloud clusters output by the Kalman filter in real time, and calculates the estimated time TETA for the target to reach the boundary of the adjacent lighting zone in real time. When TETA is less than the preset PWM hardware slow-up time threshold, a cross-zone dimming command is issued in advance to realize the pre-on of the lights, completely eliminating the lighting lag problem caused by hardware response delay.

[0108] At the level of generating the motion-guided lighting envelope, the crude control method of global illumination is abandoned. Instead, a local dynamic highlight lighting area is constructed based on the predicted trajectory endpoint of the human point cloud cluster. Centered on the predicted target location, a circular motion-guided lighting highlight envelope with a radius of approximately 2 meters is mapped in physical space. The luminaires in the central area of ​​the envelope operate at the baseline maximum brightness, while the PWM duty cycle of the edge luminaires smoothly transitions to the ambient low-light brightness according to a Gaussian decay law. This method only provides high-brightness illumination to the effective area around the pedestrian, while the luminaires in other areas maintain low-power dormant low-light conditions. This achieves energy-saving control while creating a smooth halo effect that follows the movement of the pedestrian, completely eliminating the problems of abrupt changes in light and discontinuous brightness in traditional zoned lighting.

[0109] At the specific dimming calculation level, the continuous two-dimensional Gaussian light field distribution is discretized and mapped to each physical luminaire node. Let the coordinates of the endpoint of the predicted trajectory of the human point cloud cluster output by the Kalman filter be... The first in the office setting The center coordinates of the street light circuit are Solve for the Euclidean distance between the current lighting fixture and the predicted center of the human body point cloud cluster. : , The processor calculates and sends the data to the first... Real-time PWM duty cycle of a lighting controller The formula for (i.e., light brightness) is as follows: , in, The target maximum brightness requirement (e.g., 80%) is calculated under the current natural light conditions. Set the background light to sleep brightness (usually set to about 10% for safe pathfinding background light). This is the control parameter for the light spot envelope radius.

[0110] In a real-world office setting, This allows for a distance from the prediction center The brightness of the luminaires at a given location is reduced to approximately 25% of their maximum brightness, thus creating a soft, gradually fading, invisible halo of searchlight light with a radius of about 2 meters in the physical space. This mapping formula breaks the traditional rigid on-off partitioning limitation, allowing the PWM dimming signal to flow mathematically smoothly between multiple luminaire circuits, effectively eliminating abrupt changes in light within the field of view.

[0111] If a radar malfunctions, the following procedure applies: If a radar fails to transmit data for 10 consecutive seconds (or the data packet loss rate is >50%), it is determined to be a radar malfunction. The processor immediately initiates area coverage compensation: If there are adjacent radars in the area where the malfunctioning radar is located, the area is covered by the detection range of the adjacent radars (reducing the detection accuracy to ≥85%); if there are no adjacent radars, the lights in the area are switched to "constant light mode" (brightness 50%), and a fault alarm signal is sent to the administrator terminal through the processor's RS485 interface. If a communication failure occurs, the following procedures will be followed: If communication between the processor and the lighting controller is interrupted (no feedback is received from the controller for 5 consecutive seconds), the processor will store the last sent control signal locally and resend it after communication is restored; if the communication interruption lasts for more than 30 seconds, the lighting controller will automatically switch to emergency mode (all lights will be constantly on at 50% brightness) until communication is restored.

[0112] like Figure 4 As shown, Figure 4 This is a general overview of an office scene and its corresponding lighting layout. Below, we use a small to medium-sized open-plan office (45㎡, including 6 workstations and 1 small meeting room) as an example, combined with... Figure 4 The complete implementation process of the overall detection, tracking and intelligent lighting control of this invention is described in detail. Figure 4 The specific layout and effective coverage area of ​​the two 2D LiDARs, one processor, one lighting controller, and two natural light sensors are shown.

[0113] The office setting selected in this embodiment has an overall size of 9m long and 5m wide. On the left side of the office area, there are two rows of six workstations, spaced 1.2m apart, with 1.5m high partitions. On the right side is a small 12㎡ meeting room with a 0.9m wide doorway. A 6m × 1.8m window is installed on the south side of the office to provide natural light. A total of 12 dimmable LED lights are installed indoors, with 8 lights in the office area controlled independently in two groups, and 4 lights in the meeting room controlled as a single group. All lights support 0~100% continuous PWM stepless dimming.

[0114] The hardware is deployed according to the principle of zoned coverage and cross-coverage: two 2D LiDARs are deployed on the ceiling in the southwest and northeast corners of the office, respectively, at a height of 2m and a downward tilt angle of 8°, covering the south-facing workstations, the entrance area of ​​the meeting room, the north-facing workstations, and the interior of the meeting room, achieving full-area, blind-spot-free scanning perception. The core processor is an NVIDIA Jetson ORIN NX, located in the west-facing equipment cabinet, which establishes data transmission with the two LiDARs via USB 2.0 and interacts with the lighting controller and natural light sensors via an RS485 bus. The lighting controller is an RS485 bus-type dimming controller, independently controlling three sets of indoor LED lighting circuits. Two BH1750 natural light sensors are symmetrically deployed on both sides of the south-facing window, collecting real-time indoor natural light intensity data and uploading it to the processor.

[0115] After powering on, the system first completes static environment modeling and parameter configuration. In an unmanned, dark environment, two LiDARs continuously acquire multiple frames of raw point cloud data. After filtering and noise reduction, a static point cloud map containing obstacles such as office partitions, workstations, desks, and chairs is generated and stored locally. Simultaneously, global parameters such as pedestrian detection size threshold, movement speed threshold, Kalman tracking parameters, Hungarian matching threshold, as well as lighting zone priority, delayed light-off duration, and brightness benchmark threshold are uniformly configured on the processor. Natural light sensor illuminance calibration is also completed, providing parameter benchmarks and environmental prior models for subsequent real-time operation.

[0116] The system acquires raw point cloud data from dual radars in real time at a frequency of 10Hz, removes outlier noise and invalid point clouds through a joint filtering algorithm, and then converts the data to a global coordinate system to complete the fusion of the dual radar point clouds, thus obtaining real-time point cloud data for the entire office area.

[0117] The system performs differential processing between the real-time point cloud and the pre-stored static environment model, stripping away the static obstacle point cloud and retaining only the dynamic motion point cloud. Based on the adaptive clustering method described earlier, which adapts to the long-range divergence characteristics of radar and environmental topological constraints, the clustering radius is adaptively adjusted to effectively solve the problems of human point clouds adhering to the environment and long-distance target fragmentation in workstation partition scenarios, accurately extracting independent human point cloud clusters. Valid pedestrian targets are selected by combining target contour size and motion speed features, generating a set of candidate human point cloud clusters for the current frame.

[0118] Based on the detected human point cloud clusters, an improved six-dimensional state Kalman filter is used to predict inter-frame motion states. A Hungarian algorithm incorporating motion and deformation costs is then used to achieve optimal target matching between consecutive frames. The system manages the entire lifecycle, including new target ID allocation, normal frame state updates, temporary prediction for short-term occlusion, and release of long-term departure IDs. Furthermore, for scenarios with dense, overlapping occlusion by multiple people, a multi-level conflict resolution strategy is employed to ensure that targets do not have duplicate IDs or are lost in high-density crowds, maintaining stable tracking.

[0119] The system reads illuminance data from natural light sensors in real time and combines this data with the number of human point cloud clusters, dwell time, and area priority in each region to execute multi-condition joint triggering logic. Based on the real-time location of pedestrians and Kalman predictive motion state, it achieves cross-regional light prediction feedforward triggering. Combined with a dynamic Gaussian follower light envelope dimming model, it generates refined PWM dimming commands that adapt to the current number of people, their location, movement trends, and ambient light.

[0120] The lighting controller receives dimming commands from the processor and performs smooth PWM adjustment of the brightness of the lights in each zone, achieving pre-on, delayed off, and gradual follow-up lighting effects. Simultaneously, it transmits the current lighting status and fault information back to the processor in real time, completing a full control loop within a single frame and continuously iterating.

[0121] The core operating parameters of this embodiment are set as follows: LiDAR scanning frequency 10Hz, effective detection distance 0.5~5m; effective pedestrian movement speed range 0.3~1.5m / s, contour size threshold adapted to office pedestrian characteristics; Kalman position prediction accuracy at the tracking level ≤0.1m, multi-target ID confusion rate <2%; lighting control response delay ≤0.5s, independent configuration of light-off delay duration for each zone, and accurate brightness adjustment resolution with no step-like abrupt changes.

[0122] The system combines pedestrian count, dwell time, movement trajectory, and real-time illuminance information from human point cloud clusters within the area. It utilizes zone priority logic to trigger light start / stop and fine-grained brightness adjustment. Combined with time-sharing correction coefficients and differentiated light-off delay strategies, it constructs a complete closed-loop lighting control system. This invention uses 300 lx natural illuminance as the brightness adjustment threshold, matching corresponding baseline lighting brightness based on pedestrian count ranges within the office area (1-2 people, 3-5 people, ≥6 people). Furthermore, considering different daytime working hours (8:00-10:00, 10:00-16:00, 16:00-18:00) and weekend / holiday conditions, it uses an adaptive correction coefficient in the range of 0.6~1.2 to calibrate and optimize the baseline brightness, achieving continuous stepless brightness adjustment across the entire range from 0~100%.

[0123] Meanwhile, this invention adds a hardware fault tolerance mechanism to improve system stability and office lighting reliability. When the lidar has no data output for 10 consecutive seconds or the data packet loss rate is greater than 50%, the system determines that the sensing module is abnormal, automatically activates the area lighting coverage compensation strategy or the light constant mode, and simultaneously uploads the device abnormality alarm signal; when the communication interruption between the processor and the lighting controller lasts for more than 30 seconds, the lighting controller actively switches to a 50% brightness emergency working mode to avoid area darkness due to equipment or communication abnormalities, ensuring continuous, safe, and stable operation of office scene lighting.

[0124] Real-world 24-hour continuous operation tests show that the present invention achieves a pedestrian detection accuracy of over 96% under complex conditions such as multiple people working normally, back-and-forth movement, and partial occlusion. Tracking continuity in typical unobstructed scenarios is no less than 92%, and short-term occlusion can be quickly recovered within 0.8 seconds. Relying on predictive feedforward dimming and dynamic follow-up light zone mechanisms, the accuracy of light prediction and control reaches 98%, completely eliminating the lag, abrupt changes, and fragmented lighting problems inherent in traditional lighting. Compared to traditional fixed-duration constant lighting, this system achieves an average daily energy saving rate of 48%, significantly exceeding preset energy-saving targets. Simultaneously, the system only collects laser ranging point cloud data, without image or video acquisition, ensuring the privacy and security of office workers throughout the process and adapting to the needs of routine office intelligent lighting.

[0125] Traditional 2D LiDAR in office applications is highly susceptible to point cloud frame drift caused by reflections from desks and chairs and slow movement of people, leading to frequent tracking interruptions. This invention innovatively reconstructs the state space vector of the Kalman filter (introducing physical acceleration) and pioneers a dynamic adaptive adjustment formula for the Q-matrix based on the distance to physical obstacles. This technical solution directly compensates for the physical detection blind spots of 2D LiDAR in environments with dense obstacles. Combined with Hungarian algorithm matching that fuses multi-dimensional spatiotemporal physical features, the system maintains accuracy in trajectory splitting and tracking even in extreme physical scenarios such as severe cross-occlusion (point cloud overlap rate > 50%) and extremely close parallel walking. Real-world testing shows that the tracking recovery time in occluded scenarios is ≤ 1 second, and the multi-target ID confusion rate is reduced to < 2%. This fundamentally solves the technical bottleneck of traditional radar perception solutions—"ID jumps and lighting miscontrols due to prediction divergence"—with a response latency of less than 0.5 seconds.

[0126] This invention offers practical benefits while balancing energy saving, privacy, and office experience. Based on precise adjustments of pedestrian numbers, dwell time, and natural light intensity, combined with predictive control and area priority management, it achieves an office lighting energy saving rate of ≥45%. Calculated for a small to medium-sized office (12 LED lights, 18W per light, 8 hours of work per day), it can save approximately 1200 yuan in electricity costs annually. At the same time, it reduces the usage time of the lights, extends their lifespan (from 3 years to 5 years), and lowers maintenance costs. This invention uses only 2D LiDAR to collect distance data and does not involve any image or video information, thus avoiding the privacy leakage risk of camera solutions, eliminating employees' concerns about privacy leakage, and improving team trust. This invention features predictive lighting to prevent people from entering areas in the dark, tiered brightness adjustment to suit different numbers of office workers, a fixed color temperature of 4000K to avoid visual fatigue, and an anomaly tolerance mechanism to ensure stable operation of the lights. It eliminates problems such as frequent switching on and off and sudden brightness changes, resulting in a user experience satisfaction rate of ≥95%, thus improving office comfort and efficiency.

[0127] This invention boasts high application value, low cost, and easy promotion. It is adaptable to multiple scenarios, compatible with existing systems, reduces deployment costs, supports RS485 bus communication, and can be directly connected to existing office lighting equipment such as LED lights and fluorescent lights without the need to replace the lamps. The hardware selection (YDLIDAR T-mini Plus radar, NVIDIA Jetson ORIN NX processor) is moderately priced, with a deployment cost of approximately 5,000 yuan per office (45㎡), which is only 40%-50% of the cost of traditional camera-based smart lighting solutions (approximately 12,000 yuan), lowering the barrier to entry for enterprise transformation. This invention has strong adaptability and wide application scope: it is not only suitable for small and medium-sized open offices, but can also be adapted to different scenarios such as independent office cubicles (10-15㎡ / radar), co-working spaces (30-40㎡ / radar), and large office floors (deployed in zones) by adjusting the number of radars and their installation positions. Moreover, it does not require major modification of the algorithm logic and has a short adaptation cycle (parameter debugging for a certain scenario can be completed in 1-2 days).

[0128] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0130] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0131] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0132] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A pedestrian detection method, characterized in that, include: S1: Initialize a=1; where a represents the detection frame index; S2: Obtain the human point cloud cluster set of the target region in frame a and frame a+1. Based on the actual position of each human point cloud cluster in the human point cloud cluster set in frame a+1 and its distance from the nearest obstacle, and the actual position of each human point cloud cluster in the human point cloud cluster set in frame a, calculate the motion cost of each pair of human point cloud clusters. S3: Based on the aspect ratio of the bounding rectangle of each human point cloud cluster in the (a+1)th frame and its distance from the nearest obstacle, and the aspect ratio of the bounding rectangle of each human point cloud cluster in the a-th frame, calculate the deformation compensation cost of each pair of human point cloud clusters. S4: Using the matching cost matrix of the Hungarian algorithm, based on the motion cost and deformation compensation cost of each group of human point cloud clusters, the matching cost of each group of human point cloud clusters is calculated, thereby obtaining the first human point cloud cluster subset that is successfully matched and the second human point cloud cluster subset that is unmatched in the human point cloud cluster set of frame a+1. S5: Update a=a+1 and return to step S2 for real-time pedestrian detection.

2. The pedestrian detection method according to claim 1, characterized in that, The formula for calculating the deformation compensation cost of each human point cloud cluster pair is as follows: ， in, The cost of deformation compensation for human body point cloud clusters. Centering human point cloud clusters The width of the bounding rectangle, Centering human point cloud clusters The width of the bounding rectangle, , This represents the total number of human point cloud clusters in frame a. , This represents the total number of human point cloud clusters in frame a+1. Centering human point cloud clusters The length of the circumscribed rectangle, Centering human point cloud clusters The length of the circumscribed rectangle, Centering human point cloud clusters Relative to the human point cloud cluster in frame a The instantaneous rate of change of the aspect ratio of the circumscribed rectangle. , The maximum deformation penalty constant, It is an exponential function with the natural constant as its base. Centering human point cloud clusters The actual position and the distance between its nearest obstacle. This is the threshold for the width of the physical feature.

3. The pedestrian detection method according to claim 1, characterized in that, After obtaining the first successfully matched human point cloud cluster subset and the second unmatched human point cloud cluster subset from the human point cloud cluster set in frame a+1, it also includes: Based on the first human point cloud cluster subset in frame a, the position of each human point cloud cluster in frame a and frame a-1, and its velocity and acceleration between the two frames, the predicted position of each human point cloud cluster in frame a+1 and the predicted velocity between frame a and frame a+1 are obtained. Based on the actual positions of each human point cloud cluster in frame a and frame a+1, the instantaneous velocity vector of each human point cloud cluster is calculated. Based on the actual and predicted positions of each human point cloud cluster in frame a+1, the Euclidean distance of each human point cloud cluster is calculated; based on the predicted velocity and instantaneous velocity vector of each human point cloud cluster between frame a and frame a+1, the cosine similarity of the assumed motion direction of each human point cloud cluster is calculated; based on the Euclidean distance, the cosine similarity of the assumed motion direction, and the distance to the nearest obstacle with its actual position, the motion cost of each human point cloud cluster is calculated. Obtain the deformation compensation cost of each human point cloud cluster and its matching human point cloud cluster; Using the matching cost matrix of the Hungarian algorithm, the matching cost of each human point cloud cluster is calculated based on the motion cost and deformation compensation cost of each cluster, thereby verifying the matching result of each cluster.

4. The pedestrian detection method according to claim 3, characterized in that, The process of obtaining the actual position of each human point cloud cluster in frame a includes: The predicted position of each human point cloud cluster in frame a is fused with the actual position to obtain the optimized actual position of each human point cloud cluster in frame a.

5. The pedestrian detection method according to claim 3, characterized in that, Based on the Euclidean distance of each human point cloud cluster, the cosine similarity of the assumed motion direction, and the distance to the nearest obstacle with its actual position, the formula for calculating the motion cost of each human point cloud cluster is as follows: ， ， in, For the first The cost of movement for individual human point cloud clusters These are the weighting coefficients. For directional weighting coefficients, , It is an exponential function with the natural constant as its base. To set a safe distance, For the direction weighting coefficient parameter, For the first The Euclidean distance between the actual positions of an individual human point cloud cluster in frame a and in frame (a+1). For the first Cosine similarity of the assumed motion directions of individual human point cloud clusters. for The distance between an individual human point cloud cluster and its nearest obstacle in frame a. For the first The prediction speed of an individual human point cloud cluster between frames a and a+1. For the first The instantaneous velocity vector of an individual human point cloud cluster from frame a to frame a+1. , For time intervals, For the first The actual position of an individual human point cloud cluster in frame a+1. For the first The actual position of an individual human point cloud cluster in frame a. This indicates the calculation of the modulus.

6. The pedestrian detection method according to claim 3, characterized in that, Based on the first human point cloud cluster subset in frame a, the method for obtaining the predicted position of each human point cloud cluster in frame a+1 and the predicted velocity between frames a and a+1 includes the following: For each human point cloud cluster in the first human point cloud cluster subset of frame a, the Kalman filter corresponding to the human point cloud cluster is used. Based on the position of the human point cloud cluster in frame a and frame a-1, and the velocity and acceleration between the two frames, the predicted position of each human point cloud cluster in frame a+1 and the predicted velocity between frame a and frame a+1 are obtained. The process of obtaining the noise covariance of the Kalman filter for each human point cloud cluster includes: Based on the trajectory curvature and velocity change rate of each human point cloud cluster before the a-th frame, the fusion adjustment factor and noise robustness determination factor are determined. An environmental compensation term is constructed based on the distance from the human point cloud cluster to the nearest obstacle and the density of surrounding targets within a range centered on the human point cloud cluster. Based on the fusion adjustment factor, noise robustness judgment factor, and environmental compensation term, the noise variance of the acceleration process in the x-axis and y-axis directions is calculated. Based on the acceleration process noise variance, acceleration noise weight baseline value, and velocity noise weight baseline value in the x-axis and y-axis directions, the velocity process noise variance in the x-axis and y-axis directions is obtained. By integrating the variances of position, velocity, and acceleration processes along the x and y axes, a process noise covariance is constructed.

7. The pedestrian detection method according to claim 6, characterized in that, The formula for the environmental compensation item is: ， in, For environmental compensation items, The obstacle influence coefficient. The target density influence coefficient, It is an exponential function with the natural constant as its base. The distance from the human figure point cloud cluster to the nearest obstacle. Distance threshold The density of surrounding targets within a range centered on the human body point cloud cluster is set; Based on the acceleration process noise variance, acceleration noise weight baseline value, and velocity noise weight baseline value in the x-axis and y-axis directions, the formula for obtaining the velocity process noise variance in the x-axis and y-axis directions is as follows: ， ， in, The variance of the velocity process noise in the x-axis direction. The velocity noise weighting baseline value in the x-axis direction. Let Variance be the noise variance of the acceleration process in the x-axis direction. The x-axis direction is used as the baseline value for acceleration noise weighting. The variance of the velocity process noise in the y-axis direction. The velocity noise weighting baseline value in the y-axis direction. Let Variance be the noise variance of the acceleration process in the y-axis direction. The acceleration noise weighting baseline value is located in the y-axis direction.

8. The pedestrian detection method according to claim 1, characterized in that, Methods for obtaining the human point cloud cluster set of the target region in frame a include: Use lidar to obtain the original laser points in the target area of ​​frame a; Clustering and segmentation are performed based on the dynamic clustering radius of each original laser point in the target region of frame a to obtain several point cloud clusters, and human point cloud clusters are identified in several point cloud clusters. The formula for calculating the dynamic clustering radius of each original laser point is as follows: ， in, The dynamic clustering radius of the original laser points. This is the distance divergence compensation coefficient. This represents the absolute physical distance between the original laser point and the lidar. This is the environmental repulsion weighting coefficient. The obstacle repulsion coefficient of the original laser point. The distance is determined based on the distance between the original laser point and the obstacle with the smallest distance.

9. A method for intelligent lighting control, characterized in that, include: The pedestrian detection method according to any one of claims 1 to 8 is used to detect pedestrians in a target area; The lighting of the lamps is controlled based on the detection results and real-time illuminance data.

10. A lighting intelligent control system, characterized in that, include: lighting; A natural light sensor, connected to a lighting lamp, is used to acquire real-time illuminance data; LiDAR is used to acquire raw laser points in a target area. The host computer is connected to the lidar and the natural light sensor, and coupled to the lighting lamp, to implement the steps of the intelligent lighting control method described in claim 9.

Images