Intelligent lighting management system and method based on multi-source data

Abstract

The invention discloses an intelligent lighting management system and method based on multi-source data, and relates to the technical field of intelligent lighting management, and the system comprises a multi-source data collection module, an environment feature anchoring module, a dynamic error modeling and calibration module, an edge-cloud collaborative decision module and a lighting strategy execution module. The multi-source data acquisition module fuses environment sensor, visual perception and laser radar data to construct a'sensor original data + environment characteristic data 'dual-input system; the environment feature anchoring module extracts stable features from the visual / laser radar data to serve as a calibration reference; the dynamic error modeling and calibration module quantifies the error in real time based on the reference and generates calibration parameters; the edge-cloud collaborative decision-making module and cloud long-term optimization form a hierarchical collaborative architecture through edge node real-time control; and the lighting strategy execution module dynamically adjusts lighting parameters according to the calibration data and feeds back an effect to form a closed loop.

Description

Technical Field

[0001] The present invention relates to the technical field of smart lighting management, and in particular to a smart lighting management system and method based on multi-source data. Background Art

[0002] As smart city development progresses, smart lighting systems, a crucial component of urban infrastructure, are becoming increasingly intelligent, directly impacting public safety, energy efficiency, and user experience. Traditional smart lighting systems primarily rely on single environmental sensors (such as illuminance, temperature, and humidity) to collect data and adjust lighting parameters using preset thresholds or timed calibration strategies. However, these systems face multiple technical bottlenecks in complex environments, necessitating breakthrough solutions.

[0003] Traditional systems use a single sensor for environmental perception, resulting in insufficient data reliability. For example, illuminance sensors are susceptible to temperature drift, and condensation forms on the sensor surface when humidity exceeds 80% RH, further exacerbating data deviations. Vibration accelerometers are affected by the mechanical vibration of the lamp, and their cumulative displacement error increases with equipment aging, making them unable to meet the requirements of high-precision lighting control. Furthermore, existing systems rely on fixed thresholds or scheduled calibration (e.g., daily), lacking real-time dynamic calibration mechanisms. When faced with temporary events (such as road construction, crowds) or extreme weather (heavy rain, smog), sudden changes in environmental parameters can cause sensor data anomalies, which the system cannot identify and adjust in a timely manner. This often results in excessive or insufficient lighting, even posing safety risks.

[0004] When it comes to calibration benchmarks, existing technologies rely on the sensor's inherent stability or manually preset parameters, failing to establish a physical benchmark using inherent environmental characteristics. For example, the geometric dimensions and reflectivity of fixed structures like traffic signs and building facades are subject to national standards. However, traditional systems fail to incorporate these characteristics into their calibration systems, resulting in a lack of a calibration benchmark and an objective basis for error correction.

[0005] Traditional smart lighting systems generally adopt a centralized "sensor-cloud-actuator" architecture with weak edge computing capabilities. This leads to two core issues: First, they lack real-time performance. Sensor data needs to be uploaded to the cloud for processing, resulting in round-trip delays of seconds, making them unable to meet millisecond-level response requirements. Second, their reliability relies on network connectivity. When the network is disconnected, the system fails immediately. For example, in scenarios such as mountainous areas or tunnels, communication failures can cause lamps to lose control, and manual intervention to adjust the response can be delayed for hours, seriously affecting lighting continuity.

[0006] Furthermore, existing systems lack scene-specific adaptability and rigid calibration strategies. For example, the reflectivity of reflective strips on tunnel walls decreases with age, but the system doesn't dynamically update the reflectivity benchmark based on LiDAR 3D point cloud data, leading to increasing illumination calibration errors year by year. Temporary activity areas rely on manually configured lighting parameters, failing to use visual perception to detect crowds in real time and automatically adjust them, resulting in low efficiency and prone to errors.

[0007] In summary, existing smart lighting systems suffer from fundamental flaws in data collection, calibration mechanisms, and architectural design: single sensors have weak anti-interference capabilities, calibration strategies are static, and system architectures rely on the cloud and lack local intelligence. These issues lead to high error rates, delayed responses, and high maintenance costs in complex environments, making them unable to meet the high-precision, high-reliability lighting requirements of smart cities.

[0008] To address the above-mentioned issues, the present invention proposes a smart lighting management system and method based on multi-source data. Through multimodal data collection from environmental sensors, visual perception, and lidar, a dynamic calibration system is constructed in which "environmental characteristics serve as calibration benchmarks". Edge computing is used to achieve real-time error modeling and scenario-based strategy adjustment, while the cloud performs long-term model optimization through federated learning. This architecture breaks through the single data dependence and centralized bottlenecks of traditional systems, achieving a fully closed-loop control of "self-perception-self-calibration-self-optimization", providing a new technical path for smart lighting technology. Summary of the Invention

[0009] The purpose of the present invention is to provide an intelligent lighting management system and method based on multi-source data to solve the problems raised in the prior art.

[0010] To achieve the above object, the present invention provides the following technical solutions: A smart lighting management system based on multi-source data is characterized by: including a multi-source data acquisition module, an environmental feature anchoring module, a dynamic error modeling and calibration module, an edge-cloud collaborative decision-making module and a lighting strategy execution module; the multi-source data acquisition module is responsible for real-time acquisition of environmental data and calibration of anchoring features, and constructing a dual-input system of sensor raw data + environmental feature data; the environmental feature anchor module extracts stable and unchanging environmental features from visual / lidar data as a physical benchmark for sensor calibration; the dynamic error modeling and calibration module is used to calculate sensor data deviations and generate calibration parameters in real time based on the environmental feature benchmark; in the edge-cloud collaborative decision-making module, the edge nodes and the cloud platform work together, the edge nodes are responsible for real-time calibration and local control, and the cloud is responsible for long-term model optimization, forming a layered collaborative architecture between the edge and the cloud; the lighting strategy execution module dynamically adjusts the lighting parameters according to the calibrated data, and feeds back the calibration effect to form a closed loop.

[0011] The multi-source data acquisition module includes an environmental sensor unit, a visual perception unit and a laser radar unit; The environmental sensor unit is responsible for deploying illuminance, temperature and humidity, and vibration acceleration sensors to collect illuminance, temperature and humidity, and vibration acceleration data in real time, and uses anti-vibration packaging and signal filtering technology to reduce the impact of external interference on the data; The visual perception unit is equipped with a camera to collect fixed features of the road and its surroundings in real time. Fixed features include but are not limited to traffic signs, building facade textures, and zebra crossings. A lightweight neural network is used to detect fixed features in real time and a background difference method is used to eliminate dynamic targets. Static visual features are extracted as anchor point data for sensor calibration, providing a visual basis for scene recognition. The laser radar unit obtains three-dimensional point cloud data through a solid-state laser radar, combines it with the static features extracted by the visual perception unit, uses a voxel filtering algorithm to extract the contour features and geometric dimensions of static objects, and aligns the laser radar coordinate system with the visual coordinate system to provide baseline geometric data of the environment for subsequent construction of a three-dimensional spatial model of the environment.

[0012] The environmental feature anchoring module includes a static feature recognition unit and a spatial geometry modeling unit; The static feature recognition unit is responsible for screening out long-term stable fixed features from visual images and eliminating dynamic interference. Specifically, for feature stability judgment, the image of the same area is continuously observed. If the position and shape of a feature do not change in multiple shots, it is judged as a static feature. Conversely, if the position or outline changes continuously, it is marked as a dynamic target and filtered out. For feature semantic annotation, semantic labels are assigned to the features based on their visual morphology and combined with the built-in standard feature library. Through feature stability judgment and feature semantic annotation, semantic data including feature type, two-dimensional position, and stability is formed to provide a visual reference for subsequent spatial positioning. The spatial geometric modeling unit first separates objects through density analysis, then determines the size through extreme value calculation, and finally performs coordinate transformation and associated vision to convert the lidar point cloud into physical reference data that can be used for calibration, which is used to construct a three-dimensional spatial model of static features and determine their geometric size and spatial coordinates.

[0013] The spatial geometric modeling unit specifically includes the following contents: Using LiDAR point cloud data, a 3D spatial model of static features is constructed to determine their geometric dimensions and spatial coordinates. Specifically: First, by analyzing the density and distribution of the point cloud returned by the lidar, a point cloud density threshold is set, high-density areas are identified as object areas, and low-density areas are identified as background. The spatial distance between adjacent points is used to determine whether the point cloud belongs to the same connected domain, thus separating independent objects. Then, the least squares method is used to fit the ground plane, retaining points in the point cloud whose vertical distance from the ground plane is greater than a preset threshold and excluding ground points. Finally, vegetation is identified based on the height distribution of the point cloud to exclude unstructured background. Then calculate the three-dimensional space coordinates and geometric dimensions. For the three-dimensional coordinate calculation, set the point cloud set P = {(x i ,y i ,z i )}, calculate the center of mass coordinates as the three-dimensional position of the object, the calculation formula is as follows: ; Where n is the number of point clouds, and the calculated coordinates represent the spatial center position of the object; For geometric dimension calculation, find the extreme value in the direction of the object point cloud coordinate axis and calculate the length, width and height. The formula is as follows: ; where x max , x min The maximum / minimum coordinate of the point cloud in the x-axis direction, and the same applies to the y and z axes; The principal axis of the point cloud is determined by principal component analysis, and the average distance from the point cloud to the axis is calculated as the radius. The calculation formula is as follows: ; Height is the extreme value difference of the point cloud in the z-axis direction: Height = z max -z min ; Furthermore, cuboid / rectangular objects are identified. When the size ratios of the point cloud in the three coordinate axes satisfy |{length / width} - 1| < 0.2 and |{height / width} - 5| < 0.5, the point cloud is identified as a rectangular plane object. Combined with semantic annotation, the standard size of the rectangular symbol in the national standard library is matched to verify whether the measured size error exceeds the preset error threshold. Cylinders / rods are judged by calculating the standard deviation of the distance from the point cloud to the main axis. r , if r If the point cloud is less than 2cm, that is, it is distributed around the axis, it is judged to be a cylinder. The diameter of the cylinder / rod is D = 2r, which must comply with the national standard rod diameter range. Finally, the external parameters of the visual coordinate system and the lidar coordinate system are calibrated. The conversion relationship between the two is established through the rotation matrix R and the translation vector T. The two-dimensional pixel coordinates (u, v) in the visual image are converted to the three-dimensional coordinates of the lidar. The conversion formula is as follows: ; Where (Xc, Yc, Zc) is the three-dimensional coordinate in the camera coordinate system, which is calculated from the pixel coordinates using the camera intrinsic parameter matrix. The calculation formula is as follows: ; Where K is the camera intrinsic parameter matrix, fx / fy is the focal length, and (u0, v0) is the principal point coordinate; By separating objects into extreme value calculations through the above density analysis and then determining their sizes through coordinate conversion, the logic of associated vision is used to transform static objects in the environment into computable calibration scales.

[0014] The dynamic error modeling and calibration module includes an error modeling unit and a calibration strategy execution unit; The error modeling unit uses the static feature physical benchmark output by the environmental feature anchoring module to calculate the theoretical value of the sensor data, and compares it with the measured value of the environmental sensor unit to quantify the error. Specifically: first, the theoretical value is calculated, and the theoretical value is divided into the theoretical value of illumination and the theoretical value of vibration displacement; For the theoretical value of illuminance, the theoretical light intensity E is calculated by the illuminance calculation formula based on the geometric size and reflectivity of the static feature and the spatial coordinates of the distance sensor. 理论 , the formula is as follows: ; in is the luminous flux of the light source, is the reflectivity, the reflectivity is the reflectivity of the national standard reflective film, A is the characteristic surface area, and d is the distance between the sensor and the characteristic; For the theoretical value of vibration displacement, the three-dimensional coordinate stability of the street light pole detected by the laser radar is used to determine that the theoretical value of the sensor vibration displacement is 0, which is compared with the measured value of the integral displacement of the vibration acceleration sensor; Then calculate the error between the theoretical value and the measured value. The error types are divided into absolute error and relative error; The absolute error is used to directly correct the sensor output and is calculated as follows: ; The relative error is used to determine whether the error exceeds the threshold range. The calculation formula is as follows: ; The calibration strategy execution unit generates sensor calibration parameters based on the output of the error modeling calculation unit, triggers the lighting strategy adjustment, and provides feedback on the calibration effect to optimize the model; Specifically: For calibration parameter generation, a linear correction method is used to directly compensate for absolute errors, and a sliding filter processing method is used to average high-frequency vibration noise filtering. The generated calibration parameters need to be corrected for sensor parameters and control lighting parameters through the hardware interface of the edge computing unit to ensure the change of calibration parameters from digital signals to physical control; then, based on the scene recognition results of the environmental feature anchoring module, the calibration strategy is dynamically adjusted: low-frequency threshold trigger calibration is used for conventional scenarios to correct errors exceeding ±5%; temporary activity scenes are switched to high-frequency real-time calibration, with 100% theoretical value correction; tunnel scenes use a mixed correction with 80% of the theoretical value combined with the fixed reflective strip feature to improve adaptability to complex environments. Finally, a dimming instruction is sent to the lighting strategy execution module, and the calibration instruction is fed back to the edge computing unit at the same time.

[0015] The edge-cloud collaborative decision-making module includes an edge computing unit and a cloud platform control unit; The edge computing unit is responsible for receiving the raw data from the multi-source data acquisition module in real time, performing time synchronization calibration and noise filtering, and generating a standardized data set with spatiotemporal labels; then calling the static feature semantic labels and three-dimensional geometric parameters of the environmental feature anchoring module to build a localized environmental feature benchmark library; then running the algorithm of the dynamic error modeling and calibration module, calculating the sensor theoretical value based on the benchmark library, generating calibration parameters, and performing sensor parameter correction and lighting control operations through the hardware interface; the sensor parameter correction writes the compensation coefficient to the vibration sensor through the built-in interface to correct the original vibration data in real time; the lighting control sends brightness / color temperature instructions to the lighting driver module through the PWM signal; finally, executing the edge-end closed-loop control operation, specifically: monitoring data fluctuations within a preset time after calibration, and automatically triggering secondary calibration when the error exceeds ±3% for three consecutive times to ensure the real-time nature of local control; then storing the environmental feature models and calibration parameters for the last seven days, so that calibration and lighting strategies can continue to be executed based on local data when the network is interrupted, avoiding system failures caused by cloud dependence; The cloud platform control unit receives desensitized statistical data uploaded by each edge computing unit, constructs a global error distribution map and scene feature library; and adopts federated learning technology to aggregate multi-node data to train the global feature recognition model without transmitting the original data; then periodically generates model update packages and pushes them to the edge computing units to support offline updates.

[0016] The lighting strategy execution module includes a parameter parsing unit and an execution control unit; The parameter parsing unit receives the calibration parameters sent by the edge computing unit and parses them into control signals that can be recognized by the hardware. Specifically, the digital instructions of the edge computing unit are converted into a protocol supported by the lamp driver module. Then, the control parameters are double-checked in combination with the scene recognition results of the environmental feature anchoring module. The execution control unit drives the lighting equipment to execute control instructions through the hardware interface to achieve brightness adjustment, color temperature switching and emergency response; specifically: for brightness adjustment, digital PWM signals are output through PWM to control the output power of the LED driver power supply; for color temperature adjustment, color temperature instructions are sent through the DALI protocol to adapt to the lighting needs of different scenarios; for emergency scene response, by receiving the emergency trigger signal from the edge computing unit, it automatically switches to the emergency lighting mode and controls the backup power supply access through the relay; when the equipment is in operation, the voltage and current data of the lamp are monitored in real time, and the equipment operation status is fed back to the system in real time to prevent equipment failure.

[0017] A smart lighting management method based on multi-source data includes the following steps: S1: Real-time environmental status data is acquired through illumination, temperature, humidity, and vibration acceleration sensors, while anti-vibration packaging and signal filtering technologies are used to reduce noise interference. A camera is used to capture images of fixed road features, and a lightweight neural network is used to detect static features and eliminate dynamic targets. LiDAR is used to acquire 3D point cloud data, and density analysis is used to separate independent objects and eliminate background. A unified timestamp is added to the collected environmental data, visual image data, and LiDAR point cloud data, and the sensor coordinates are associated with the feature space position to generate a standardized dataset with spatiotemporal labels. S2. Determine the stability of features in visual images, identify features that are continuously observed without displacement as static features, and perform semantic annotation based on the national standard library; perform centroid calculation, extreme value analysis, and shape determination on the LiDAR point cloud data, unify the visual and LiDAR coordinate systems through external calibration, and establish a mapping relationship between feature semantics, spatial coordinates, and geometric dimensions; S3. Calculate theoretical illuminance using a formula based on the sign's reflectivity, size, and sensor distance. Determine theoretical vibration displacement values ​​based on the stability of the light pole's coordinates as detected by the LiDAR. Compare these theoretical values ​​with the sensor's measured values, calculate absolute and relative errors, and determine whether the errors exceed a threshold. S4: Linear compensation is performed for absolute errors exceeding the threshold, and sliding filtering is performed on the vibration data. Calibration strategies are developed for different scenarios. In conventional scenarios, calibration is triggered when the error is greater than ±5%. In temporary event scenarios, real-time calibration is performed when crowds are detected, and 100% theoretical value correction is mandatory. In tunnel scenarios, a mixed correction with 80% theoretical value is used, taking into account the characteristics of wall reflective strips. S5. Write a compensation coefficient to the vibration sensor to correct the original data. Generate a PWM signal or DALI color temperature command based on the corrected illuminance, and drive the lamp dimming through the hardware interface. After calibration, continuously monitor the data within a preset time range. If the error is greater than ±3% three times in a row, a second calibration is triggered. The feature model and calibration parameters for the past seven days are stored to support continuous control based on local data when the network is disconnected. S6: Feedback on the status and control effect of the lighting fixture, generating feedback data in JSON format including the device ID, execution result, and exception code; the cloud aggregates the desensitized data of each edge node, builds a global error map and scene feature library, uses federated learning to optimize the feature recognition model, periodically generates update packages and pushes them to the edge computing unit through a secure channel, and supports offline update calibration strategies.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. To address the problem of traditional systems relying on a single sensor, which leads to significant data deviation due to interference from temperature, humidity, and vibration, this invention integrates three types of data: environmental sensors, visual perception, and lidar. Cameras identify static features such as traffic signs and building textures, while lidar acquires three-dimensional geometric dimensions. Combined with real-time data from environmental sensors, this invention extracts inherently standardized features in the environment (such as national standard traffic signs and building components) as calibration benchmarks. The physical parameters of these features, such as geometric dimensions and reflectivity, are defined by national standards, forming an objective and reliable calibration reference. This fundamentally addresses the shortcomings of traditional systems that rely on inherent sensor stability or manually preset parameters, improving data reliability from the source. This invention establishes a self-calibration system where "environmental features serve as calibration benchmarks," addressing the core issues of weak interference resistance and the lack of calibration benchmarks associated with single sensors.

[0019] 2. To address the existing system's lack of a real-time calibration mechanism and its delayed response to scenarios like temporary events, this invention uses dynamic error modeling to calculate the deviation between the sensor's measured and theoretical values ​​in real time. It then dynamically adjusts the calibration strategy based on the scene recognition results: Low-frequency threshold-triggered calibration is used for routine scenarios, while temporary events or extreme weather scenarios automatically switch to high-frequency real-time calibration. A hybrid correction strategy is implemented in conjunction with fixed features for special scenarios like tunnels. This "real-time error quantification + scenario-based strategy" enables the system to respond to environmental changes in milliseconds and automatically adapt to complex scenarios (such as crowds, heavy rain, and aging tunnel reflective strips) without manual intervention. This addresses the static calibration strategies and long response delays of traditional systems, significantly improving the real-time and precision of lighting control.

[0020] 3. To address the issues of traditional systems relying on the cloud, resulting in insufficient real-time performance and failure during network disconnections, this invention adopts an edge-cloud layered collaborative architecture: the edge computing unit is integrated into the smart lamp, processing multi-source data in real time, generating calibration parameters, and driving lamp adjustment. It supports the continuous operation of the locally stored feature model during network disconnections, ensuring millisecond-level response and offline reliability; the cloud platform aggregates desensitized data from each node, optimizes the global model through federated learning, and pushes update packages to achieve long-term strategy evolution. This architecture not only solves the latency and single point of failure problems of centralized architectures, but also improves the system's generalization capabilities through long-term cloud optimization, forming a closed loop of "local real-time control + cloud-based global evolution", reducing reliance on human intervention and significantly improving the system's reliability and intelligence in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a system workflow diagram of a smart lighting management system based on multi-source data in the present invention; Figure 2 This is a system organizational structure diagram of a smart lighting management system based on multi-source data in the present invention. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0023] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution. A smart lighting management system based on multi-source data, characterized by: including a multi-source data acquisition module, an environmental feature anchoring module, a dynamic error modeling and calibration module, an edge-cloud collaborative decision-making module and a lighting strategy execution module; the multi-source data acquisition module is responsible for real-time acquisition of environmental data and calibration of anchoring features, and constructing a dual-input system of sensor raw data + environmental feature data; the environmental feature anchor module extracts stable and unchanging environmental features from visual / lidar data as a physical benchmark for sensor calibration; the dynamic error modeling and calibration module is used to calculate sensor data deviations and generate calibration parameters in real time based on the environmental feature benchmark; in the edge-cloud collaborative decision-making module, the edge node and the cloud platform work together, the edge node is responsible for real-time calibration and local control, and the cloud is responsible for long-term model optimization, forming a layered collaborative architecture between the edge and the cloud; the lighting strategy execution module dynamically adjusts the lighting parameters according to the calibrated data, and feeds back the calibration effect to form a closed loop; such as Figure 1 、 Figure 2 shown.

[0024] The multi-source data acquisition module includes an environmental sensor unit, a visual perception unit and a laser radar unit; The environmental sensor unit is responsible for deploying illuminance, temperature and humidity, and vibration acceleration sensors to collect illuminance, temperature and humidity, and vibration acceleration data in real time, and uses anti-vibration packaging and signal filtering technology to reduce the impact of external interference on the data; Through anti-vibration packaging and signal filtering technology, the high-frequency noise of the vibration acceleration sensor is reduced by more than 60%, and the drift error of the temperature and humidity sensor in a high-humidity environment is reduced from ±10% to ±3%, ensuring the reliability of the original environmental data and providing a stable input for subsequent error modeling; the real-time collection of light intensity, temperature, humidity, and vibration data forms a multi-dimensional environmental status perception, solving the problem of weak anti-interference ability of traditional single sensors.

[0025] The visual perception unit is equipped with a camera to collect fixed features of the road and its surroundings in real time. Fixed features include but are not limited to traffic signs, building facade textures, and zebra crossings. A lightweight neural network is used to detect fixed features in real time and a background difference method is used to eliminate dynamic targets. Static visual features are extracted as anchor point data for sensor calibration, providing a visual basis for scene recognition. By combining a lightweight neural network with a background difference method, the accuracy of dynamic target (vehicles, pedestrians) rejection is improved, and the detection delay of static features (traffic signs, building textures) is reduced to <20ms, providing real-time and accurate visual anchor data for scene recognition.

[0026] The laser radar unit obtains three-dimensional point cloud data through a solid-state laser radar, combines it with the static features extracted by the visual perception unit, uses a voxel filtering algorithm to extract the contour features and geometric dimensions of static objects, and aligns the laser radar coordinate system with the visual coordinate system to provide baseline geometric data of the environment for subsequent construction of a three-dimensional spatial model of the environment.

[0027] A voxel filtering algorithm is used to improve the accuracy of separating independent objects (such as streetlight poles and traffic signs), the coordinate error of the 3D point cloud, and the geometric dimension measurement error (for example, the measured error of a 60cm×40cm sign is less than 3cm), providing an accurate spatial geometric benchmark for anchoring environmental features. Coordinate system alignment technology ensures that the spatial error between visual images and lidar data is less than 1 cm, achieving accurate registration of multimodal data and providing a unified spatiotemporal benchmark for subsequent error modeling.

[0028] The environmental feature anchoring module includes a static feature recognition unit and a spatial geometry modeling unit; The static feature recognition unit is responsible for screening out long-term stable fixed features from visual images and eliminating dynamic interference. Specifically, for feature stability judgment, the image of the same area is continuously observed. If the position and shape of a feature do not change in multiple shots, it is judged as a static feature. Conversely, if the position or outline changes continuously, it is marked as a dynamic target and filtered out. For feature semantic annotation, semantic labels are assigned to the features based on their visual morphology and combined with the built-in standard feature library. Through feature stability judgment and feature semantic annotation, semantic data including feature type, two-dimensional position, and stability is formed to provide a visual reference for subsequent spatial positioning. Through the stability judgment mechanism (no displacement for 30 consecutive minutes is judged as a static feature), dynamic targets (such as the misjudgment rate of moving vehicles is less than 2%) are filtered out. Combined with the semantic annotation of the national standard library (such as the matching accuracy of GB 5768 traffic signs is greater than 95%), semantic data containing "feature type + two-dimensional position + stability" is generated, providing stable morphological feature input for the calibration benchmark. After eliminating dynamic interference, the amount of effective anchor point data of static features is reduced by 80% (only long-term stable features are retained), improving the efficiency of subsequent spatial modeling and benchmark quality.

[0029] The spatial geometric modeling unit first separates objects through density analysis, then determines the size through extreme value calculation, and finally performs coordinate transformation and associated vision to convert the lidar point cloud into physical reference data that can be used for calibration, which is used to construct a three-dimensional spatial model of static features and determine their geometric size and spatial coordinates.

[0030] The spatial geometric modeling unit specifically includes the following contents: Using LiDAR point cloud data, a 3D spatial model of static features is constructed to determine their geometric dimensions and spatial coordinates. Specifically: First, by analyzing the density and distribution of the point cloud returned by the lidar, a point cloud density threshold is set, high-density areas are identified as object areas, and low-density areas are identified as background. The spatial distance between adjacent points is used to determine whether the point cloud belongs to the same connected domain, thus separating independent objects. Then, the least squares method is used to fit the ground plane, retaining points in the point cloud whose vertical distance from the ground plane is greater than a preset threshold and excluding ground points. Finally, vegetation is identified based on the height distribution of the point cloud to exclude unstructured background. Then calculate the three-dimensional space coordinates and geometric dimensions. For the three-dimensional coordinate calculation, set the point cloud set P = {(x i ,y i ,z i )}, calculate the center of mass coordinates as the three-dimensional position of the object, the calculation formula is as follows: ; Where n is the number of point clouds, and the calculated coordinates represent the spatial center position of the object; For geometric dimension calculation, find the extreme value in the direction of the object point cloud coordinate axis and calculate the length, width and height. The formula is as follows: ; where x max , x min The maximum / minimum coordinate of the point cloud in the x-axis direction, and the same applies to the y and z axes; The principal axis of the point cloud is determined by principal component analysis, and the average distance from the point cloud to the axis is calculated as the radius. The calculation formula is as follows: ; Height is the extreme value difference of the point cloud in the z-axis direction: Height = z max -z min ; Furthermore, cuboid / rectangular objects are identified. When the size ratios of the point cloud in the three coordinate axes satisfy |{length / width} - 1| < 0.2 and |{height / width} - 5| < 0.5, the point cloud is identified as a rectangular plane object. Combined with semantic annotation, the standard size of the rectangular symbol in the national standard library is matched to verify whether the measured size error exceeds the preset error threshold. Cylinders / rods are judged by calculating the standard deviation of the distance from the point cloud to the main axis. r , if r If the point cloud is less than 2cm, that is, it is distributed around the axis, it is judged to be a cylinder. The diameter of the cylinder / rod is D = 2r, which must comply with the national standard rod diameter range. Finally, the external parameters of the visual coordinate system and the lidar coordinate system are calibrated. The conversion relationship between the two is established through the rotation matrix R and the translation vector T. The two-dimensional pixel coordinates (u, v) in the visual image are converted to the three-dimensional coordinates of the lidar. The conversion formula is as follows: ; Where (Xc, Yc, Zc) is the three-dimensional coordinate in the camera coordinate system, which is calculated from the pixel coordinates using the camera intrinsic parameter matrix. The calculation formula is as follows: ; Where K is the camera intrinsic parameter matrix, fx / fy is the focal length, and (u0, v0) is the principal point coordinate; By separating objects into extreme value calculations through the above density analysis and then determining their sizes through coordinate conversion, the logic of associated vision is used to transform static objects in the environment into computable calibration scales.

[0031] Independent objects are separated through density analysis and connected domain detection (for example, the accuracy of separating street light pole point clouds is >98%), and ground points are excluded by fitting the ground plane using the least squares method (error rate is <1%). Vegetation is filtered out in combination with height distribution to ensure that only structured objects (such as traffic signs and building components) are retained, providing pure 3D geometric data for error modeling.

[0032] Based on centroid calculation (three-dimensional coordinate error <5cm), extreme value analysis (length, width and height measurement error <3%) and shape judgment (cylinder / rectangular block judgment accuracy >95%), a high-precision three-dimensional spatial model is constructed (for example, the three-dimensional coordinate error of a speed limit sign is <2cm). After associating it with visual semantic labels, a composite calibration benchmark of "semantics-geometry-position" is formed to reduce the calculation error of the sensor's theoretical value.

[0033] The dynamic error modeling and calibration module includes an error modeling unit and a calibration strategy execution unit; The error modeling unit uses the static feature physical benchmark output by the environmental feature anchoring module to calculate the theoretical value of the sensor data, and compares it with the measured value of the environmental sensor unit to quantify the error. Specifically: first, the theoretical value is calculated, and the theoretical value is divided into the theoretical value of illumination and the theoretical value of vibration displacement; For the theoretical value of illuminance, the theoretical light intensity E is calculated by the illuminance calculation formula based on the geometric size and reflectivity of the static feature and the spatial coordinates of the distance sensor. 理论 , the formula is as follows: ; in is the luminous flux of the light source, is the reflectivity, the reflectivity is the reflectivity of the national standard reflective film, A is the characteristic surface area, and d is the distance between the sensor and the characteristic; For the theoretical value of vibration displacement, the three-dimensional coordinate stability of the street light pole detected by the laser radar is used to determine that the theoretical value of the sensor vibration displacement is 0, which is compared with the measured value of the integral displacement of the vibration acceleration sensor; Then calculate the error between the theoretical value and the measured value. The error types are divided into absolute error and relative error; The absolute error is used to directly correct the sensor output and is calculated as follows: ; The relative error is used to determine whether the error exceeds the threshold range. The calculation formula is as follows: ; Theoretical values ​​are calculated based on environmental characteristic benchmarks (e.g., the theoretical illuminance error is less than 5%, and the theoretical vibration displacement deviation is less than 0.3 cm). Deviations are quantified through absolute error (direct correction of sensor output) and relative error (threshold judgment), thereby improving the calibration and correction accuracy of the sensor's actual measured values ​​(e.g., the measured illuminance of 550 lux is close to the theoretical value of 600 lux after correction, with the error reduced from -50 lux to -10 lux). Multi-physical quantity coupling modeling (joint calibration of illuminance and vibration displacement) addresses the one-sidedness of single parameter calibration. For example, the drift error of temperature and humidity sensors in high-vibration scenarios is simultaneously corrected, reducing the combined error rate of multiple sensors from 20% to below 8%.

[0034] The calibration strategy execution unit generates sensor calibration parameters based on the output of the error modeling calculation unit, triggers the lighting strategy adjustment, and provides feedback on the calibration effect to optimize the model; Specifically: For calibration parameter generation, a linear correction method is used to directly compensate for absolute errors, and a sliding filter processing method is used to average high-frequency vibration noise filtering. The generated calibration parameters need to be corrected for sensor parameters and control lighting parameters through the hardware interface of the edge computing unit to ensure the change of calibration parameters from digital signals to physical control; then, based on the scene recognition results of the environmental feature anchoring module, the calibration strategy is dynamically adjusted: low-frequency threshold trigger calibration is used for conventional scenarios to correct errors exceeding ±5%; temporary activity scenes are switched to high-frequency real-time calibration, with 100% theoretical value correction; tunnel scenes use a mixed correction with 80% of the theoretical value combined with the fixed reflective strip feature to improve adaptability to complex environments. Finally, a dimming instruction is sent to the lighting strategy execution module, and the calibration instruction is fed back to the edge computing unit at the same time.

[0035] High-frequency vibration noise is processed through linear correction and sliding filtering (for example, the fluctuation amplitude of vibration data is reduced by 40%). Combined with scenario-based strategies (low-frequency calibration for conventional scenarios, high-frequency calibration for temporary activities, and mixed correction for tunnel scenarios), the calibration response speed in complex environments is improved (for example, the calibration delay for temporary activity scenarios is shortened from 30 seconds to 1 second). The generated calibration parameters (such as vibration compensation coefficient and lighting correction value) directly drive the hardware interface of the edge computing unit to ensure the physical implementation accuracy of the calibration instructions (lamp brightness adjustment error is less than 2%, and color temperature switching delay is less than 50ms).

[0036] The edge-cloud collaborative decision-making module includes an edge computing unit and a cloud platform control unit; The edge computing unit is responsible for receiving the raw data from the multi-source data acquisition module in real time, performing time synchronization calibration and noise filtering, and generating a standardized data set with spatiotemporal labels; then calling the static feature semantic labels and three-dimensional geometric parameters of the environmental feature anchoring module to build a localized environmental feature benchmark library; then running the algorithm of the dynamic error modeling and calibration module, calculating the sensor theoretical value based on the benchmark library, generating calibration parameters, and performing sensor parameter correction and lighting control operations through the hardware interface; the sensor parameter correction writes the compensation coefficient to the vibration sensor through the built-in interface to correct the original vibration data in real time; the lighting control sends brightness / color temperature instructions to the lighting driver module through the PWM signal; finally, executing the edge-end closed-loop control operation, specifically: monitoring data fluctuations within a preset time after calibration, and automatically triggering secondary calibration when the error exceeds ±3% for three consecutive times to ensure the real-time nature of local control; then storing the environmental feature models and calibration parameters for the last seven days, so that calibration and lighting strategies can continue to be executed based on local data when the network is interrupted, avoiding system failures caused by cloud dependence; By processing multi-source data in real time (data preprocessing delay <50ms), building a local feature benchmark library (storing data from the last 7 days), and supporting offline calibration during network disconnection (for example, in tunnel scenarios, calibration can still be performed at a frequency of 1 minute after network disconnection, with a stable error rate within ±5%), the single point failure problem of traditional systems relying on the cloud is solved; sensor compensation coefficients are then directly written through the hardware interface (for example, the real-time correction delay of the vibration sensor is <10ms), and lighting control commands are generated (PWM signal duty cycle error is <1%), achieving millisecond-level control response (for example, emergency lighting switching delay is <50ms).

[0037] The cloud platform control unit receives desensitized statistical data uploaded by each edge computing unit, constructs a global error distribution map and scene feature library; and adopts federated learning technology to aggregate multi-node data to train the global feature recognition model without transmitting the original data; then periodically generates model update packages and pushes them to the edge computing units to support offline updates.

[0038] Federated learning technology is used to aggregate desensitized data from multiple nodes (such as regional error distribution and calibration success rate), optimize global feature recognition models (for example, the accuracy of sign detection in rainstorm scenes has increased from 85% to 98%), and periodically push updates (such as the addition of fluorescent sign reflectivity parameters), improving edge scene recognition capabilities by 30%. A global error map and scene feature library are constructed (for example, dynamic updates of the mean reflectivity of tunnel reflective strips). Long-term strategy optimization has reduced the annual increase in the system's error rate from 15% to 3% in equipment aging scenarios, reducing manual maintenance intervention.

[0039] The lighting strategy execution module includes a parameter parsing unit and an execution control unit; The parameter parsing unit receives the calibration parameters sent by the edge computing unit and parses them into control signals that can be recognized by the hardware. Specifically, the digital instructions of the edge computing unit are converted into a protocol supported by the lamp driver module. Then, the control parameters are double-checked in combination with the scene recognition results of the environmental feature anchoring module. The compatibility of protocol conversion (such as PWM signals) is greater than 95%, supporting the integration of new and old lighting equipment. The secondary verification mechanism (combined with scene tags) reduces the control parameter error rate to less than 0.5% (for example, the error rate of color temperature commands in tunnel scenes is reduced from 5% to 0.2%), ensuring the accurate implementation of calibration parameters.

[0040] The execution control unit drives the lighting equipment to execute control instructions through the hardware interface to achieve brightness adjustment, color temperature switching and emergency response; specifically: for brightness adjustment, digital PWM signals are output through PWM to control the output power of the LED driver power supply; for color temperature adjustment, color temperature instructions are sent through the DALI protocol to adapt to the lighting needs of different scenarios; for emergency scene response, by receiving the emergency trigger signal from the edge computing unit, it automatically switches to the emergency lighting mode and controls the backup power supply access through the relay; when the equipment is in operation, the voltage and current data of the lamp are monitored in real time, and the equipment operation status is fed back to the system in real time to prevent equipment failure.

[0041] The execution control unit improves brightness adjustment accuracy (e.g., the measured error of the target 70% brightness is less than 0.7%), color temperature switching speed is less than 200ms, and emergency response delay (e.g., vibration exceeding the threshold triggers full brightness mode) is less than 50ms, thereby enhancing the accuracy and safety of lighting control. At the same time, it monitors lamp voltage and current in real time (with anomaly detection accuracy of >98%) and generates device status feedback (e.g., "Abnormal current of lamp A - fault code 001"), increasing the equipment fault warning advance rate to 90%, thereby reducing maintenance costs.

[0042] A smart lighting management method based on multi-source data includes the following steps: S1: Real-time environmental status data is acquired through illumination, temperature, humidity, and vibration acceleration sensors, while anti-vibration packaging and signal filtering technologies are used to reduce noise interference. A camera is used to capture images of fixed road features, and a lightweight neural network is used to detect static features and eliminate dynamic targets. LiDAR is used to acquire 3D point cloud data, and density analysis is used to separate independent objects and eliminate background. A unified timestamp is added to the collected environmental data, visual image data, and LiDAR point cloud data, and the sensor coordinates are associated with the feature space position to generate a standardized dataset with spatiotemporal labels. S2. Determine the stability of features in visual images, identify features that are continuously observed without displacement as static features, and perform semantic annotation based on the national standard library; perform centroid calculation, extreme value analysis, and shape determination on the LiDAR point cloud data, unify the visual and LiDAR coordinate systems through external calibration, and establish a mapping relationship between feature semantics, spatial coordinates, and geometric dimensions; S3. Calculate theoretical illuminance using a formula based on the sign's reflectivity, size, and sensor distance. Determine theoretical vibration displacement values ​​based on the stability of the light pole's coordinates as detected by the LiDAR. Compare these theoretical values ​​with the sensor's measured values, calculate absolute and relative errors, and determine whether the errors exceed a threshold. S4: Linear compensation is performed for absolute errors exceeding the threshold, and sliding filtering is performed on the vibration data. Calibration strategies are developed for different scenarios. In conventional scenarios, calibration is triggered when the error is greater than ±5%. In temporary event scenarios, real-time calibration is performed when crowds are detected, and 100% theoretical value correction is mandatory. In tunnel scenarios, a mixed correction with 80% theoretical value is used, taking into account the characteristics of wall reflective strips. S5. Write a compensation coefficient to the vibration sensor to correct the original data. Generate a PWM signal or DALI color temperature command based on the corrected illuminance, and drive the lamp dimming through the hardware interface. After calibration, continuously monitor the data within a preset time range. If the error is greater than ±3% three times in a row, a second calibration is triggered. The feature model and calibration parameters for the past seven days are stored to support continuous control based on local data when the network is disconnected. S6: Feedback on the status and control effect of the lighting fixture, generating feedback data in JSON format including the device ID, execution result, and exception code; the cloud aggregates the desensitized data of each edge node, builds a global error map and scene feature library, uses federated learning to optimize the feature recognition model, periodically generates update packages and pushes them to the edge computing unit through a secure channel, and supports offline update calibration strategies.

[0043] Example: The implementation scenario was selected for a city arterial road, where smart streetlights with integrated edge computing units were deployed. These were accompanied by traffic signs (60km / h speed limit signs complying with the GB5768 standard, 60cm×40cm in size, and 80% reflectivity), solid-state LiDAR (accuracy ±2cm), and high-definition cameras (resolution 1920×1080). The system achieved dynamic calibration and precise adjustment of lighting parameters through multi-source data fusion. First, data collection is performed, including real-time collection of illuminance (measured at 550 lux), temperature and humidity (25°C, 70% RH), and vibration acceleration (0.3 m / s²). After anti-vibration packaging and sliding filtering, the vibration data noise is reduced by 60%, and the temperature and humidity drift error is controlled within ±2%.

[0044] The visual perception unit camera collects road images in real time and detects speed limit signs using a lightweight neural network (MobileNetV3). It then uses background subtraction to remove dynamic vehicles and confirms that their positions have not changed for 30 consecutive minutes. These features are then identified as static and a semantic label is output (type: prohibition sign, coordinates: (100, 200) pixels).

[0045] The LiDAR unit acquires 3D point cloud data and separates the marker pole (cylinder, 30 cm in diameter, 5 m in height) through density analysis. After alignment with the coordinate system (error < 1 cm), the 3D coordinates of the marker are determined to be (5, 3, 2) meters, with dimensions of 0.6 m × 0.4 m × 0.05 m. Then, environmental features are anchored and benchmarks are constructed, long-term stable features are screened (no displacement for 10 consecutive observations), and semantic annotation is performed by matching the national standard library to generate structured data containing feature type, location, and stability (such as "Speed ​​Limit Sign-ID001-Continuous Stability").

[0046] Spatial geometric modeling constructs a three-dimensional spatial model of the sign through point cloud density analysis, centroid calculation (error < 5 cm), and shape determination (rectangular block determination accuracy > 95%). This verifies that the error between the measured dimensions and the national standard is less than 3%, forming a "semantic-geometric-position" composite benchmark (e.g., reflectivity 80%, three-dimensional coordinates (5,3,2) meters). Based on the reflectivity, size, and sensor distance of the sign, the theoretical value is calculated to be 600 lux (assuming a luminous flux of 1000 lm). The measured value is 550 lux, with an absolute error of -50 lux (relative error of -8.3%, exceeding the ±5% threshold).

[0047] The LiDAR detection pole coordinates are stable, with a theoretical value of 0cm, a measured integrated displacement of 0.5cm, and an absolute error of -0.5cm. Calibration is triggered every 5 minutes when the error exceeds ±5%. Linear compensation (corrected to 600 lux illumination) and sliding filtering (40% noise reduction in vibration data) are used.

[0048] Temporary event scenarios (such as a crowd of more than 10 people): Switch to real-time calibration (1 time / second) and force the use of 100% theoretical value correction.

[0049] Then, the system receives multi-source data, builds a localized benchmark library (including the reflectivity, size, and coordinates of the ID001 marker), generates calibration parameters (vibration compensation coefficient +0.5cm, PWM dimming signal duty cycle 70%), and writes these parameters to the sensor and lamp driver modules through the hardware interface. After correction, the vibration displacement error is reduced to 0cm, and the lamp brightness is adjusted to 600lux (measured error <1%).

[0050] Closed-loop verification and offline support monitor data within 10 minutes after calibration. If the error is greater than ±3% (e.g., below 570 lux) for three consecutive times, a second calibration is triggered. The system stores the last seven days of data and continuously controls data according to the local benchmark when the network is disconnected (e.g., the calibration frequency in tunnel scenarios is once every minute, with a stable error within ±5%). The cloud receives desensitized statistical data for data aggregation and model training (such as a regional average error rate of 5% and a calibration success rate of 98%), builds a global error map and scene feature library (such as an average reflectivity of 80% for main road signs), and uses federated learning to optimize the sign detection model. The recognition accuracy of heavy rain scenes has increased from 85% to 95%.

[0051] Strategy updates and push notifications periodically generate model update packages (such as adding slippery road reflection correction parameters), which are pushed to the edge through a secure channel and support offline updates, reducing the annual increase in the system's long-term error rate from 15% to 3%.

[0052] Parameter parsing and hardware driver The parameter parsing unit converts edge commands (such as "brightness 70%, color temperature 4000K") into DALI protocol signals, and combines them with scene tags for secondary verification (such as the color temperature of conventional roads ≥ 3000K) to ensure command accuracy (control parameter error rate <0.5%).

[0053] The execution control unit adjusts the lamp brightness to 70% (measured error <0.7%) through PWM signals, and monitors the lamp current of 0.5A (rated 0.55A) in real time. If there is no abnormality, it will feedback "execution successful"; Full process logical closed loop and technical effects The system follows the closed-loop process of "data collection → feature anchoring → error modeling → edge calibration → strategy execution → cloud optimization": Precise calibration: illumination error corrected from -8.3% to ±1%, vibration displacement error returned to zero; Real-time response: calibration delay for regular scenarios is less than 50ms, and response time for temporary activity scenarios is within seconds. Reliable operation: The edge device operates independently during network outages, providing a 90% early warning rate for equipment failures and reducing maintenance costs by 40%.

[0054] Through multi-source data fusion and edge-cloud collaboration, this embodiment verifies the system's high-precision control capabilities in complex environments and provides a replicable technical solution for smart city lighting.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. An intelligent lighting management system based on multi-source data, characterized by: It includes a multi-source data acquisition module, an environmental feature anchoring module, a dynamic error modeling and calibration module, an edge-cloud collaborative decision-making module and a lighting strategy execution module; the multi-source data acquisition module is responsible for real-time acquisition of environmental data and calibration of anchoring features, and constructing a dual-input system of sensor raw data + environmental feature data; the environmental feature anchoring module extracts stable and unchanging environmental features from visual / lidar data as a physical benchmark for sensor calibration; the dynamic error modeling and calibration module is used to calculate sensor data deviations and generate calibration parameters in real time based on the environmental feature benchmark; in the edge-cloud collaborative decision-making module, the edge nodes work together with the cloud platform, the edge nodes are responsible for real-time calibration and local control, and the cloud is responsible for long-term model optimization, forming a layered collaborative architecture between the edge and the cloud; the lighting strategy execution module dynamically adjusts the lighting parameters according to the calibrated data, and feeds back the calibration effect to form a closed loop.

2. The intelligent lighting management system based on multi-source data according to claim 1, characterized in that: The multi-source data acquisition module includes an environmental sensor unit, a visual perception unit and a laser radar unit; The environmental sensor unit is responsible for deploying illuminance, temperature and humidity, and vibration acceleration sensors to collect illuminance, temperature and humidity, and vibration acceleration data in real time, and uses anti-vibration packaging and signal filtering technology to reduce the impact of external interference on the data; The visual perception unit is equipped with a camera to collect fixed features of the road and its surroundings in real time. Fixed features include but are not limited to traffic signs, building facade textures, and zebra crossings. A lightweight neural network is used to detect fixed features in real time and a background difference method is used to eliminate dynamic targets. Static visual features are extracted as anchor point data for sensor calibration, providing a visual basis for scene recognition. The laser radar unit obtains three-dimensional point cloud data through a solid-state laser radar, combines it with the static features extracted by the visual perception unit, uses a voxel filtering algorithm to extract the contour features and geometric dimensions of static objects, and aligns the laser radar coordinate system with the visual coordinate system to provide baseline geometric data of the environment for subsequent construction of a three-dimensional spatial model of the environment.

3. The intelligent lighting management system based on multi-source data according to claim 1, characterized in that: The environmental feature anchoring module includes a static feature recognition unit and a spatial geometry modeling unit; The static feature recognition unit is responsible for screening out long-term stable fixed features from the visual image and eliminating dynamic interference. Specifically, for feature stability judgment, the image of the same area is continuously observed. If the position and shape of a feature do not change in multiple shots, it is determined to be a static feature. Conversely, if the position or outline continuously changes, it is marked as a dynamic target and filtered. For feature semantic annotation, semantic labels are assigned based on the visual form of the feature and combined with the built-in standard feature library. Through feature stability judgment and feature semantic annotation, semantic data including feature type, two-dimensional position, and stability is formed, providing a visual reference for subsequent spatial positioning. The spatial geometric modeling unit first separates objects through density analysis, then determines the size through extreme value calculation, and finally performs coordinate transformation and associated vision to convert the lidar point cloud into physical reference data that can be used for calibration, which is used to construct a three-dimensional spatial model of static features and determine their geometric size and spatial coordinates.

4. The intelligent lighting management system based on multi-source data according to claim 3, characterized in that: The spatial geometric modeling unit specifically includes the following contents: Using LiDAR point cloud data, a 3D spatial model of static features is constructed to determine their geometric dimensions and spatial coordinates. Specifically: First, by analyzing the density and distribution of the point cloud returned by the lidar, a point cloud density threshold is set, high-density areas are identified as object areas, and low-density areas are identified as background. The spatial distance between adjacent points is used to determine whether the point cloud belongs to the same connected domain, thus separating independent objects. Then, the least squares method is used to fit the ground plane, retaining points in the point cloud whose vertical distance from the ground plane is greater than a preset threshold and excluding ground points. Finally, vegetation is identified based on the height distribution of the point cloud to exclude unstructured background. Then calculate the three-dimensional space coordinates and geometric dimensions. For the three-dimensional coordinate calculation, set the point cloud set P = {(x i ,y i ,z i )}, calculate the center of mass coordinates as the three-dimensional position of the object, the calculation formula is as follows: ； Where n is the number of point clouds, and the calculated coordinates represent the spatial center position of the object; For geometric dimension calculation, find the extreme value in the direction of the object point cloud coordinate axis and calculate the length, width and height. The formula is as follows: ； where x max , x min The maximum / minimum coordinate of the point cloud in the x-axis direction, and the same applies to the y and z axes; The principal axis of the point cloud is determined by principal component analysis, and the average distance from the point cloud to the axis is calculated as the radius. The calculation formula is as follows: ； Height is the extreme value difference of the point cloud in the z-axis direction: Height = z max -z min ; Furthermore, cuboid / rectangular objects are identified. When the size ratios of the point cloud along the three coordinate axes satisfy |{length / width} - 1| < 0.2 and |{height / width} - 5| < 0.5, the point cloud is identified as a rectangular plane object. Combined with semantic annotation, the standard size of rectangular symbols in the national standard library is matched to verify whether the measured size error exceeds the preset error threshold. Cylinders / rods are judged by calculating the standard deviation of the distance from the point cloud to the main axis. r , if r If the point cloud is less than 2cm, that is, the point cloud is distributed around the axis, it is judged to be a cylinder. The diameter of the cylinder / rod D = 2r, which must comply with the national standard rod diameter range. Finally, the external parameters of the visual coordinate system and the lidar coordinate system are calibrated. The conversion relationship between the two is established through the rotation matrix R and the translation vector T. The two-dimensional pixel coordinates (u, v) in the visual image are converted to the three-dimensional coordinates of the lidar. The conversion formula is as follows: ； Where (Xc, Yc, Zc) is the three-dimensional coordinate in the camera coordinate system, which is calculated from the pixel coordinates using the camera intrinsic parameter matrix. The calculation formula is as follows: ； Where K is the camera intrinsic parameter matrix, fx / fy is the focal length, and (u0, v0) is the principal point coordinate; By separating objects into extreme value calculations through the above density analysis and then determining their sizes through coordinate conversion, the logic of associated vision is used to transform static objects in the environment into computable calibration scales.

5. The intelligent lighting management system based on multi-source data according to claim 1, characterized in that: The dynamic error modeling and calibration module includes an error modeling unit and a calibration strategy execution unit; The error modeling unit uses the static feature physical benchmark output by the environmental feature anchoring module to calculate the theoretical value of the sensor data, and compares it with the measured value of the environmental sensor unit to quantify the error. Specifically: first, the theoretical value is calculated, and the theoretical value is divided into the theoretical value of illumination and the theoretical value of vibration displacement; For the theoretical value of illuminance, the theoretical light intensity E is calculated by the illuminance calculation formula based on the geometric size and reflectivity of the static feature and the spatial coordinates of the distance sensor. 理论 , the formula is as follows: ； in is the luminous flux of the light source, is the reflectivity, the reflectivity is the reflectivity of the national standard reflective film, A is the characteristic surface area, and d is the distance between the sensor and the characteristic; For the theoretical value of vibration displacement, the three-dimensional coordinate stability of the street light pole detected by the laser radar is used to determine that the theoretical value of the sensor vibration displacement is 0, which is compared with the measured value of the integral displacement of the vibration acceleration sensor; Then calculate the error between the theoretical value and the measured value. The error types are divided into absolute error and relative error; The absolute error is used to directly correct the sensor output and is calculated as follows: ； The relative error is used to determine whether the error exceeds the threshold range. The calculation formula is as follows: ； The calibration strategy execution unit generates sensor calibration parameters based on the output of the error modeling calculation unit, triggers the lighting strategy adjustment, and provides feedback on the calibration effect to optimize the model; Specifically, for calibration parameter generation, a linear correction method is used to directly compensate for absolute errors, and a sliding filter method is used to average high-frequency vibration noise filtering. The generated calibration parameters are then used through the hardware interface of the edge computing unit to correct sensor parameters and control lighting parameters, ensuring the change of calibration parameters from digital signals to physical control. Then, based on the scene recognition results of the environmental feature anchoring module, the calibration strategy is dynamically adjusted: low-frequency thresholds are used to trigger calibration in conventional scenarios to correct errors exceeding ±5%; temporary activity scenarios switch to high-frequency real-time calibration, with 100% theoretical value correction. The tunnel scene combines the fixed reflective strip feature with a hybrid correction of 80% of the theoretical value to improve adaptability to complex environments. Finally, a dimming instruction is sent to the lighting strategy execution module, and a calibration instruction is fed back to the edge computing unit at the same time.

6. The intelligent lighting management system based on multi-source data according to claim 1, characterized in that: The edge-cloud collaborative decision-making module includes an edge computing unit and a cloud platform control unit; The edge computing unit is responsible for receiving the raw data from the multi-source data acquisition module in real time, performing time synchronization calibration and noise filtering, and generating a standardized data set with spatiotemporal labels. It then calls the static feature semantic labels and three-dimensional geometric parameters of the environmental feature anchoring module to build a localized environmental feature reference library. It then runs the algorithm of the dynamic error modeling and calibration module, calculates the sensor theoretical value based on the reference library, generates calibration parameters, and performs sensor parameter correction and lighting control operations through the hardware interface. The sensor parameter correction writes compensation coefficients to the vibration sensor via a built-in interface, correcting the raw vibration data in real time. The lamp control sends brightness / color temperature instructions to the lamp driver module via PWM signals. Finally, edge-end closed-loop control operations are performed. Specifically, data fluctuations are monitored within a preset time after calibration. If the error exceeds ±3% three times in a row, a secondary calibration is automatically triggered to ensure real-time local control. The environmental feature model and calibration parameters for the past seven days are then stored. In the event of a network interruption, calibration and lighting strategies can continue to be executed based on local data, avoiding system failures caused by cloud dependence. The cloud platform control unit receives the desensitized statistical data uploaded by each edge computing unit, builds a global error distribution map and scene feature library, and uses federated learning technology to aggregate multi-node data to train a global feature recognition model without transmitting the original data. Then, the model update package is generated periodically and pushed to the edge computing unit to support offline updates.

7. The intelligent lighting management system based on multi-source data according to claim 1, characterized in that: The lighting strategy execution module includes a parameter parsing unit and an execution control unit; The parameter parsing unit receives the calibration parameters sent by the edge computing unit and parses them into control signals that can be recognized by the hardware. Specifically, the digital instructions of the edge computing unit are converted into a protocol supported by the lamp driver module. Then, the control parameters are double-checked in combination with the scene recognition results of the environmental feature anchoring module. The execution control unit drives the lighting equipment to execute control instructions through the hardware interface to achieve brightness adjustment, color temperature switching and emergency response; specifically: for brightness adjustment, digital PWM signals are output through PWM to control the output power of the LED driver power supply; for color temperature adjustment, color temperature instructions are sent through the DALI protocol to adapt to the lighting needs of different scenarios; for emergency scene response, by receiving the emergency trigger signal from the edge computing unit, it automatically switches to the emergency lighting mode and controls the backup power supply access through the relay; when the equipment is in operation, the voltage and current data of the lamp are monitored in real time, and the equipment operation status is fed back to the system in real time to prevent equipment failure.

8. A smart lighting management method based on multi-source data, applied to the smart lighting management system based on multi-source data according to claims 1-7, characterized in that: The following steps are involved: S1: Real-time environmental status data is acquired through illumination, temperature, humidity, and vibration acceleration sensors, while anti-vibration packaging and signal filtering technologies are used to reduce noise interference. A camera is used to capture images of fixed road features, and a lightweight neural network is used to detect static features and eliminate dynamic targets. LiDAR is used to acquire 3D point cloud data, and density analysis is used to separate independent objects and eliminate background. A unified timestamp is added to the collected environmental data, visual image data, and LiDAR point cloud data, and the sensor coordinates are associated with the feature space position to generate a standardized dataset with spatiotemporal labels. S2. Determine the stability of features in visual images, identify features that are continuously observed without displacement as static features, and perform semantic annotation based on the national standard library; perform centroid calculation, extreme value analysis, and shape determination on the LiDAR point cloud data, unify the visual and LiDAR coordinate systems through external calibration, and establish a mapping relationship between feature semantics, spatial coordinates, and geometric dimensions; S3. Calculate theoretical illuminance using a formula based on the sign's reflectivity, size, and sensor distance. Determine theoretical vibration displacement values ​​based on the stability of the light pole's coordinates as detected by the LiDAR. Compare these theoretical values ​​with the sensor's measured values, calculate absolute and relative errors, and determine whether the errors exceed a threshold. S4. Perform linear compensation for the absolute error exceeding the threshold and perform sliding filtering on the vibration data; Develop corresponding calibration strategies based on different scenarios. In normal scenarios, calibration is triggered when the error is greater than ±5%; If a crowd is detected in a temporary event scene, real-time calibration is performed and 100% theoretical value correction is mandatory; In tunnel scenes, a hybrid correction with 80% of the theoretical value is used in combination with the characteristics of wall reflective strips. S5. Write the compensation coefficient to the vibration sensor to correct the original data, generate a PWM signal or DALI color temperature command according to the corrected illuminance, and drive the lamp to dim through the hardware interface; After calibration, data is continuously monitored within a preset time range. If the error is greater than ±3% for three consecutive times, a second calibration is triggered. The characteristic model and calibration parameters for the last seven days are stored to support continuous control based on local data when the network is disconnected. S6. Feedback on the lamp status and control effect, generating JSON format feedback data including device ID, execution result, and exception code; The cloud aggregates desensitized data from each edge node, constructs a global error map and scene feature library, uses federated learning to optimize the feature recognition model, periodically generates update packages and pushes them to the edge computing unit through a secure channel, and supports offline update calibration strategies.

Images