Multi-sensor fusion sensing precision adaptive system

Abstract

The invention discloses a multi-sensor fusion sensing precision self-adaptive system, and relates to the technical field of radio positioning and multi-sensor fusion sensing, the system integrates multi-source sensing units such as a laser radar and a GPS, establishes a unified reference by means of a space-time synchronization module to complete data alignment, and optimizes data quality through preprocessing means such as filtering and denoising; a dynamic weight fusion module allocates weights according to the real-time quality scores, and a unified sensing result is output by fusing the positioning and speed measuring results; through multi-dimensional index evaluation precision grade, self-adaptive adjustment of sensor parameters and fusion strategies, and balance of precision and energy consumption, a result is output in a standardized format, an operation state is fed back, and an external system is linked. According to the invention, through multi-sensing fusion and dynamic adaptive adjustment, the sensing precision and stability are significantly improved; the method can intelligently adapt to multiple scenes, resists interference, has fault-tolerant capability, adapts to new requirements through self-learning evolution, and provides reliable perception support for automatic driving, unmanned aerial vehicle surveying and mapping and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radio positioning and multi-sensor fusion perception, and in particular to a multi-sensor fusion perception accuracy adaptive system. BACKGROUND

[0002] With the rapid development of automatic driving, intelligent robots, unmanned aerial vehicle surveying and mapping, etc., the precision, stability and adaptability of environmental perception are put forward strict requirements. Single sensor device is difficult to cope with complex and changeable application scenarios: GPS positioning module is easy to be interrupted in urban high-rise shelter, tunnel and other scenes, and the positioning accuracy is greatly reduced; laser radar can provide high-precision three-dimensional point cloud data, but is easily affected by rain, fog and same frequency interference, and the data validity is reduced; IMU inertial measurement unit can realize short-time motion state blind filling, but long-term work will accumulate drift error; visual camera can capture rich texture features, but is sensitive to light changes and its performance is limited in weak or strong light environment. Therefore, multi-sensor fusion has become the core technology direction to break through the limitations of single sensor and improve the performance of perception.

[0003] The existing multi-sensor fusion system has many technical shortcomings and is difficult to meet the needs of high-precision adaptive perception. At the data fusion level, most systems use fixed weight allocation strategy, which cannot dynamically adjust according to the real-time quality and environmental adaptability of each sensor data, resulting in that the sensors with high data quality cannot fully play their role, while the sensors with performance degradation introduce redundant errors; the time and space synchronization accuracy is insufficient, and the collection delay and spatial coordinate deviation of different sensor units are not effectively corrected, which directly affects the accuracy of the fusion result. In terms of environmental adaptability, the existing system lacks intelligent identification and adaptation ability for scene types, and uses a unified fusion strategy in different scenes such as open, sheltered and dynamic interference, resulting in large fluctuations in perception accuracy; although some systems have simple parameter adjustment function, the adjustment logic lacks data support and cannot realize the dynamic balance between precision and energy consumption.

[0004] In addition, the existing system lacks fault tolerance and anti-interference performance: when a sensor unit fails completely or its performance decays partially, there is a lack of effective redundancy mechanism and fault self-diagnosis capability, which easily leads to the paralysis of the whole perception system; in complex electromagnetic environment, laser radar, GPS and other sensors are easily affected by same frequency interference, multipath reflection, etc., and the data reliability decreases, while the existing anti-interference means is single and difficult to resist multi-dimensional interference comprehensively. At the same time, most systems do not have self-learning and remote upgrading capabilities, and the perception performance cannot be continuously optimized with the use time and scene accumulation, which is difficult to adapt to the changing application requirements. These technical limitations lead to the insufficient precision, stability and adaptability of the existing multi-sensor fusion system, which restricts the intelligent development of related fields. SUMMARY

[0005] This invention proposes a multi-sensor fusion sensing accuracy adaptive system to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-sensor fusion sensing accuracy adaptive system, comprising the following modules: Multi-source sensing unit: integrates LiDAR, GPS positioning module, IMU inertial measurement unit and vision camera; LiDAR calculates the target's three-dimensional coordinates and velocity, GPS positioning module outputs positioning data with timestamps, IMU inertial measurement unit collects angular velocity and acceleration data, and vision camera captures target texture features; Spatiotemporal synchronization module: Establishes a unified time reference based on timestamps, unifies spatial coordinate transformation through sensor extrinsic parameter matrices, and uses linear interpolation algorithms to correct the data transmission delay differences between different sensor units, aligning multi-sensor data to the same spatiotemporal coordinate system; Data preprocessing module: LiDAR point cloud data is filtered to remove noise points and downsampled by voxel gridding to retain core features; GPS data is smoothed using the Kalman smoothing algorithm to eliminate random errors; IMU data is corrected for drift errors using the zero-bias estimation algorithm; visual images are denoised using bilateral filtering and enhanced for contrast using histogram equalization. Dynamic weight fusion calculation module: Constructs a multimodal fusion framework, dynamically allocates fusion weights based on real-time quality scoring of sensor data, fuses localization, ranging and velocity measurement results through an improved extended Kalman filter algorithm, introduces a CNN-LSTM hybrid model to adaptively adjust the fusion strategy, and outputs unified perception results; Real-time perception accuracy assessment module: Establishes a multi-dimensional assessment index system, classifies accuracy levels according to preset scenario-based thresholds, and analyzes historical data using a sliding window algorithm to predict accuracy change trends; Adaptive parameter adjustment module: Triggers corresponding strategies based on accuracy assessment results. When the accuracy is low, it increases the sensor sampling and update frequency, optimizes the compensation weight, and enhances feature extraction. When the accuracy is standard, it maintains the core parameters and fine-tunes the fusion weight. When the accuracy is high, it reduces the load on non-critical units to balance accuracy and energy consumption. Output and feedback control module: Outputs the fused perception results in a standardized format, while providing real-time feedback on the operating status and perception accuracy level of each module, and notifying external systems to take security measures.

[0007] Furthermore, it also includes a weighted intelligent optimization unit, which constructs a weight calculation model based on sensor data quality, environmental adaptability, and historical errors. The optimization model expression is as follows: ,in For the first The fusion weights of individual sensor data For data quality weighting coefficients, For environment adaptation weighting coefficients, For the first Real-time quality score of sensor data For the first Historical error attenuation coefficient of individual sensor data, For the first The current environmental adaptability of individual sensor data For the first Redundancy contribution of individual sensor data The number of sensors participating in the fusion This is the sensor serial number.

[0008] Furthermore, it also includes an environmental state perception module. This module analyzes environmental feature parameters to identify the current environment type. For each environment type, a threshold range for feature parameters is set, and an environment type and sensor adaptation rule base is established. When an occluded scene is identified, the fusion weight of the IMU and visual camera is increased and the data buffering time is extended. A multi-frame data association algorithm is used to compensate for the information loss caused by some sensor signal occlusion. When a dynamic interference scene is identified, the velocity filtering algorithm of the lidar and the anti-multipath interference processing of GPS are strengthened. When a complex terrain scene is identified, the 3D ranging accuracy of the lidar and the attitude estimation frequency of the IMU are improved, and the terrain adaptation fusion strategy is optimized.

[0009] Furthermore, it also includes a sampling frequency adaptive adjustment unit, which dynamically adjusts the sampling frequency of each sensing unit based on the sensing accuracy requirements and environmental complexity. The adjustment model expression is as follows: ,in For the first The sensing unit in the first Target sampling frequency under such conditions For the first The reference sampling frequency of each sensing unit This is the frequency adjustment coefficient. For the first The current accuracy requirement coefficient for each sensing unit. For the first The complexity coefficient of the environment.

[0010] Furthermore, it also includes a sensor fault self-diagnosis and redundancy mechanism. The fault self-diagnosis unit identifies the fault type by monitoring data continuity, signal strength, and error fluctuations. When a sensor completely fails, the redundancy mechanism automatically activates the complementary functions of the remaining sensors, reallocates the fusion weights, and adjusts the preprocessing algorithm parameters. When the performance of a sensor degrades, the fusion weight of that sensor is reduced, while the acquisition intensity and data processing accuracy of other sensors are enhanced. When a sensor exhibits intermittent anomalies, a data verification mechanism is activated, using cross-validation of multi-source data to eliminate abnormal data segments and employing interpolation algorithms to supplement missing data.

[0011] Furthermore, it also includes an edge computing processing module, which adopts a heterogeneous computing architecture of CPU and GPU. The CPU uses a multi-core processor to handle logic control and data scheduling, while the GPU is configured with a dedicated computing core to handle parallel data processing, deploying computationally intensive tasks to be executed at the edge. The deep learning fusion model is pruned and quantized through model lightweighting technology to remove redundant network layers and adjust the parameter accuracy to a preset range. A hierarchical data caching strategy is adopted to prioritize storing key data related to target perception in a high-speed cache, while non-key data is processed asynchronously, so that the system perception result output latency is controlled within the millisecond range required by the application scenario.

[0012] Furthermore, it also includes a scenario-based accuracy threshold configuration module. This module has a preset library of accuracy threshold parameters for various application scenarios, including autonomous driving highway scenarios, robotic warehousing scenarios, drone mapping scenarios, and indoor positioning scenarios. Each scenario corresponds to different positioning error, ranging deviation, and speed measurement stability thresholds. Users can select the corresponding configuration or customize threshold parameters according to the actual application scenario, and the system dynamically adjusts the adaptive strategy based on the configured thresholds. It also supports online updates and offline calibration of threshold parameters, and optimizes threshold settings by importing actual scenario test data.

[0013] Furthermore, it includes a visualization and interaction module for sensing results. This module uses a Cartesian three-dimensional coordinate system to display the target position, motion trajectory, and sensing coverage in real time, and combines numerical panels and dynamic charts to present accuracy indicators. It provides a graphical interactive interface that allows users to view the working status of each sensing unit, the fusion weight allocation, and adaptive adjustment records. It allows users to manually intervene in the fusion strategy and parameter configuration, and supports custom fusion weights and sampling frequencies. It also supports historical playback and analysis of sensing data, allowing users to trace the sensing process and accuracy changes along a timeline.

[0014] Furthermore, it includes remote upgrade and self-learning update modules. The remote upgrade module receives system firmware and algorithm model update packages through a TLS encrypted communication link, supports online upgrades and breakpoint resume, retains the original configuration data during upgrades, and automatically rolls back to the original version if the upgrade fails. The self-learning update module continuously adjusts the dynamic weight fusion model, environmental recognition rules, and adaptive adjustment strategies based on massive historical sensing data and accuracy feedback, and regularly updates the sensor fault diagnosis model and accuracy threshold parameter library. The adjustment process is executed asynchronously in the background and does not affect the normal operation of the system.

[0015] Furthermore, it includes a multi-dimensional anti-interference enhancement module. This module uses pseudo-random code pulse code modulation technology to resist co-frequency interference for the lidar, and achieves signal recognition and separation with a unique code design. The GPS module integrates an anti-multipath interference antenna and an adaptive signal filtering algorithm to suppress the impact of multipath reflection signals on positioning accuracy. The IMU unit is equipped with an electromagnetic shielding structure and vibration isolation device. The vision camera adopts adaptive exposure control and anti-glare processing technology. At the same time, an interference detection algorithm is introduced at the data fusion level. By analyzing data consistency and error distribution, it identifies the interfered sensor data and reduces the impact of interference on sensing accuracy by using multi-source data redundancy verification.

[0016] Compared with existing technologies, the beneficial effects of this invention are: The multi-source sensing unit integrates LiDAR, GPS, IMU, and a vision camera, with each sensor complementing the others. Combined with standardized interfaces and data caching design, it ensures the compatibility and integrity of data transmission. The spatiotemporal synchronization module, based on high-precision timestamps and extrinsic parameter matrices, achieves accurate spatiotemporal alignment of multi-sensor data, reducing the impact of synchronization errors on perception accuracy from the source. The data preprocessing module adopts differentiated processing strategies for different sensor data characteristics, effectively removing noise, correcting drift, and enhancing features, providing a high-quality data foundation for subsequent fusion.

[0017] The dynamic weight fusion calculation module dynamically allocates fusion weights through an intelligent optimization model, enabling sensors with high data quality and good environmental adaptability to play a leading role. Combined with an improved extended Kalman filter and a deep learning model, it significantly improves the accuracy and reliability of fused sensing. The real-time sensing accuracy evaluation module establishes a multi-dimensional index system, accurately classifies accuracy levels, and predicts changing trends, providing a scientific basis for adaptive adjustment. The adaptive parameter adjustment module dynamically optimizes sensing parameters and fusion strategies according to the accuracy level, strengthening data acquisition and processing intensity in low-precision scenarios and balancing accuracy and energy consumption in high-precision scenarios, thereby achieving dynamic optimization of sensing performance.

[0018] The system's additional modules further expand its core advantages: the environmental state perception module enables intelligent scene type recognition and precise adaptation of fusion strategies, improving adaptability in complex scenarios; the sensor fault self-diagnosis and redundancy mechanism ensures perception continuity when some sensors fail, enhancing the system's fault tolerance; the edge computing processing module meets the requirements of high real-time applications through heterogeneous computing and model lightweighting; the multi-dimensional anti-interference enhancement module comprehensively resists interference from the hardware to the algorithm level, improving the system's stability in complex environments; and the remote upgrade and self-learning update module enables the system's perception performance to continuously evolve over time, adapting to new scenarios and new requirements without manual intervention.

[0019] Overall, this invention constructs a fully closed-loop perception system through multi-module collaborative design, encompassing data acquisition, preprocessing, synchronization, fusion, evaluation, and adaptive adjustment. This significantly improves perception accuracy, stability, and environmental adaptability, effectively solving problems such as fixed weights, poor environmental adaptability, weak fault tolerance, and insufficient anti-interference in existing systems. It can accurately adapt to various application scenarios such as autonomous driving, intelligent robots, and UAV mapping, providing reliable perception technology support for the intelligent upgrading of related fields and possessing broad application value. Attached Figure Description

[0020] Figure 1 This is a schematic block diagram of the multi-sensor fusion sensing accuracy adaptive system proposed in this invention; Figure 2 A comparison chart of perception accuracy in different scenarios; Figure 3 A comparison chart of system output frequencies under different data volumes; Figure 4 This is a comparison chart of the system's anti-interference performance under different interference intensities. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0024] Reference Figures 1 to 4 A multi-sensor fusion sensing accuracy adaptive system includes the following modules: Multi-source sensing unit: integrates LiDAR, GPS positioning module, IMU inertial measurement unit and vision camera. The LiDAR uses pulse time-of-flight ranging technology to emit nanosecond-level laser pulses. By measuring the pulse round-trip time, it calculates the target's three-dimensional coordinates and relative velocity. The GPS positioning module receives signals from multiple satellites and uses differential positioning technology to improve absolute position accuracy, outputting positioning data with timestamps. The IMU inertial measurement unit has a built-in three-axis gyroscope and a three-axis accelerometer. It collects angular velocity and acceleration data in real time through zero-bias calibration and temperature compensation algorithms for motion state blind spot compensation and short-term positioning assistance. The vision camera uses a high frame rate global shutter sensor to capture target texture features, assisting in target identification and scene semantic understanding. Each sensing unit is equipped with a standardized data cache module and a universal interface adapter unit to ensure data transmission compatibility with subsequent modules. Spatiotemporal synchronization module: Based on the high-precision timestamp output by the GPS positioning module, a unified time reference is established. The spatial coordinates of each sensor are uniformly transformed through the pre-calibrated sensor extrinsic parameter matrix. The linear interpolation algorithm is used to correct the data transmission delay differences of different sensing units. The data collected by LiDAR, GPS, IMU and vision camera are accurately aligned to the same spatiotemporal coordinate system, and the synchronization error is controlled within the preset millisecond range. Data preprocessing module: Differentiated processing strategies are adopted for different sensor data characteristics. For LiDAR point cloud data, discrete noise points are removed by statistical filtering algorithm, and effective point clouds are selected by setting the number of neighboring points and distance thresholds. Then, voxel gridding downsampling is used to reduce data redundancy and retain the core geometric features of the target. For GPS data, Kalman smoothing algorithm is used to eliminate random errors caused by satellite signal obstruction and improve the stability of positioning data. For IMU data, drift error is corrected in real time by zero bias estimation algorithm, and temperature compensation model is combined to reduce the impact of ambient temperature on measurement accuracy. For visual images, Gaussian noise is removed by bilateral filtering algorithm and edge details are preserved. Then, histogram equalization algorithm is used to enhance texture contrast and improve the effectiveness of subsequent feature extraction. Dynamic weight fusion calculation module: Constructs a multimodal fusion framework that combines data-level and feature-level approaches. It dynamically allocates fusion weights based on the real-time quality scores of each sensor data. It fuses positioning, ranging, and velocity measurement results through an improved extended Kalman filter algorithm. It introduces a CNN-LSTM hybrid deep learning model to learn the fusion rules under different scenarios. This model is trained with massive scene samples and can adaptively adjust the fusion strategy to optimize the accuracy of target motion trajectory and state estimation. Finally, it outputs the unified perception result after fusion. Real-time perception accuracy assessment module: Establishes a multi-dimensional assessment index system that includes positioning error, ranging deviation, speed measurement stability and data continuity. Each index is divided into three levels: high accuracy, standard accuracy and low accuracy based on preset scenario-based thresholds. The thresholds are determined through a large number of real vehicle tests and simulation experiments. At the same time, it combines the sliding window algorithm to analyze historical assessment data, predict the trend of accuracy changes, and provide a predictive basis for subsequent parameter adjustments. Adaptive parameter adjustment module: Triggers corresponding adjustment strategies based on accuracy assessment results. When at a low accuracy level, it increases the LiDAR sampling frequency and GPS data update frequency, optimizes the IMU compensation weight coefficient, and enhances feature extraction intensity by increasing the number of visual image feature point detections and strengthening the feature matching algorithm. When at a standard accuracy level, it maintains the stability of core sensing parameters and dynamically fine-tunes the fusion weights of each sensor data based on real-time quality scores. When at a high accuracy level, it appropriately reduces the sampling frequency and data transmission bandwidth of non-critical sensing units to balance sensing accuracy and system energy consumption. Output and Feedback Control Module: Outputs the fused perception results in a standardized format, supporting real-time calls from external control systems such as autonomous driving and robot navigation. The output data includes the target's three-dimensional coordinates, relative speed, attitude information, and perception accuracy level. Simultaneously, it provides real-time feedback on the operating status and perception accuracy level of each module. When the accuracy is below a preset threshold for several consecutive cycles, an abnormal alarm mechanism is triggered, redundant sensing units are activated, and external systems are notified to take safety measures.

[0025] This invention also includes a fusion weight intelligent optimization unit. This unit constructs a weight calculation model based on sensor data quality, environmental adaptability, and historical errors, and dynamically optimizes the fusion weight allocation of each sensor data point. The optimization model expression is as follows: ,in For the first The fusion weights of individual sensor data This is the data quality weighting coefficient, with a value ranging from 0.5 to 0.7. The environmental adaptation weight coefficient ranges from 0.3 to 0.5. and The sum is 1. For the first The real-time quality score of each sensor data point is calculated by comprehensively considering the fluctuation range of signal strength data integrity error. For the first The historical error decay coefficient of each sensor data point is dynamically updated over time. For the first The current environmental adaptability of each sensor data point is determined by matching environmental state parameters with sensor adaptation rules. For the first The redundancy contribution of a single sensor data point reflects the degree to which that data complements other sensor data. The number of sensors participating in the fusion Using sensor serial numbers, dynamic weight allocation allows sensors with high data quality and good environmental adaptability to play a leading role in the fusion process, improving the reliability and stability of fusion perception results in complex scenarios.

[0026] This invention also includes an environmental state perception module. This module analyzes environmental feature parameters such as lidar point cloud density, GPS satellite signal strength, IMU vibration amplitude, and visual image contrast to identify the current environment type, including open scenes, occluded scenes, dynamic interference scenes, and complex terrain scenes. Each environment type has a clearly defined threshold range for feature parameters, and an environment type and sensor adaptation rule base is established. When an occluded scene is identified, the fusion weights of the IMU and visual camera are increased, and the data buffering time is extended. A multi-frame data association algorithm is used to compensate for information loss caused by partial sensor signal occlusion. When a dynamic interference scene is identified, the lidar velocity filtering algorithm and GPS anti-multipath interference processing are strengthened to reduce the impact of dynamic interference on the perception results. When a complex terrain scene is identified, the 3D ranging accuracy of the lidar and the attitude estimation frequency of the IMU are improved, and the terrain adaptation fusion strategy is optimized to ensure accurate matching between the fusion strategy and the environmental state, avoiding a decrease in perception accuracy due to limitations in the performance of a single sensor.

[0027] This invention also includes a sampling frequency adaptive adjustment unit, which dynamically adjusts the sampling frequency of each sensing unit based on the sensing accuracy requirements and environmental complexity. The adjustment model expression is as follows: ,in For the first The sensing unit in the first Target sampling frequency under such conditions For the first The reference sampling frequency of each sensing unit is determined by hardware performance and basic accuracy requirements. This is the frequency adjustment coefficient, ranging from 0.2 to 0.4. For the first The current accuracy requirement coefficient of each sensing unit is dynamically mapped from the sensing accuracy level. For the first The complexity coefficient of the environment is calculated by weighting environmental characteristic parameters. By dynamically adjusting the sampling frequency, sufficient data acquisition is ensured in high-precision or high-complexity environments, while system energy consumption is reduced in low-precision or simple environments, thus achieving a dynamic balance between accuracy and energy consumption.

[0028] This invention also includes a sensor fault self-diagnosis and redundancy mechanism. The fault self-diagnosis unit identifies sensor fault types by monitoring the continuity of data output, signal strength, and error fluctuation range of each sensor unit. Data output continuity is judged by the number of valid data frames within a preset time. Signal strength and error fluctuation range are both set with clear thresholds. Fault types include complete failure, partial performance degradation, and intermittent anomalies. When a complete failure of a sensor is detected, the redundancy mechanism automatically activates the complementary functions of the remaining sensors, reallocates the fusion weights, and adjusts the preprocessing algorithm parameters to ensure that the core sensing function is not affected. When partial performance degradation is detected, the fusion weight of that sensor is reduced, and the acquisition intensity and data processing accuracy of other sensors are enhanced. When intermittent anomalies are detected, the data verification mechanism is activated to eliminate abnormal data segments through multi-source data cross-verification and to supplement missing data using an interpolation algorithm, ensuring that the system can still maintain the preset sensing accuracy when some sensors fail.

[0029] This invention also includes an edge computing processing module, which adopts a heterogeneous computing architecture of CPU and GPU. The CPU uses a multi-core processor to handle logic control and data scheduling, while the GPU is configured with a dedicated computing core to handle parallel data processing. Computationally intensive tasks such as data preprocessing, spatiotemporal synchronization, and dynamic weight fusion are deployed and executed at the edge. The deep learning fusion model is pruned and quantized using model lightweighting technology to remove redundant network layers and adjust the parameter accuracy to a preset range, thereby reducing computational latency. A hierarchical data caching strategy is adopted, prioritizing the storage of key data related to target perception in a high-speed cache, while non-key data is processed asynchronously. This ensures that the system perception result output latency is controlled within the millisecond range required by the application scenario, meeting the requirements of high real-time application scenarios such as autonomous driving and drone navigation.

[0030] This invention also includes a scenario-based accuracy threshold configuration module. This module has a preset library of accuracy threshold parameters for various typical application scenarios, including autonomous driving high-speed scenarios, robotic warehousing scenarios, UAV mapping scenarios, and indoor positioning scenarios. Each scenario corresponds to different positioning error, ranging deviation, and speed stability thresholds. For example, the positioning error threshold for autonomous driving high-speed scenarios is lower than that for low-speed scenarios, and the ranging deviation threshold for UAV mapping scenarios is higher than that for ordinary navigation scenarios. Users can select the corresponding configuration or customize the threshold parameters according to the actual application scenario. The system dynamically adjusts the adaptive strategy based on the configured thresholds to ensure that the perception accuracy accurately matches the scenario requirements. At the same time, it supports online updates and offline calibration of threshold parameters, and optimizes the threshold settings by importing actual scenario test data to improve the system's scenario adaptability.

[0031] This invention also includes a visualization and interaction module for sensing results. This module uses a Cartesian three-dimensional coordinate system to display the target position, motion trajectory, and sensing coverage in real time. It presents accuracy indicators such as positioning error and ranging deviation in a combination of numerical panels and dynamic charts. It provides a graphical interactive interface that allows users to view the working status of each sensing unit, the fusion weight allocation, and adaptive adjustment records. The working status includes running status, signal strength, and data transmission rate. It allows users to manually intervene in the fusion strategy and parameter configuration, and supports custom fusion weights and sampling frequencies. It also supports historical playback and analysis of sensing data, allowing users to trace the sensing process and accuracy changes along a timeline, providing data support for system optimization and troubleshooting, and improving the system's operability and maintainability.

[0032] This invention also includes a remote upgrade and self-learning update module. The remote upgrade module receives update packages for the system firmware and algorithm model via a TLS encrypted communication link, supporting online upgrades and breakpoint resume functionality. During the upgrade process, the original configuration data is retained, and in case of upgrade failure, it automatically rolls back to the original version, ensuring the security and integrity of the upgrade process. The self-learning update module, based on massive historical sensing data and accuracy feedback, uses reinforcement learning algorithms to continuously adjust the dynamic weight fusion model, environmental recognition rules, and adaptive adjustment strategies, and periodically updates the sensor fault diagnosis model and accuracy threshold parameter library. The adjustment process is executed asynchronously in the background, without affecting the normal operation of the system, enabling the system's sensing accuracy and environmental adaptability to gradually improve over time, achieving performance evolution without manual intervention.

[0033] This invention also includes a multi-dimensional anti-interference enhancement module. This module employs pseudo-random code pulse code modulation technology to resist co-frequency interference for the lidar, and achieves accurate signal identification and separation through a unique code pattern design. The GPS module integrates an anti-multipath interference antenna and an adaptive signal filtering algorithm to suppress the impact of multipath reflection signals on positioning accuracy. The IMU unit is equipped with an electromagnetic shielding structure and vibration isolation device to reduce the impact of external electromagnetic interference and mechanical vibration. The vision camera adopts adaptive exposure control and anti-glare processing technology to improve image quality under complex lighting conditions. At the same time, an interference detection algorithm is introduced at the data fusion level. By analyzing data consistency and error distribution, it identifies the interfered sensor data and reduces the impact of interference on perception accuracy through multi-source data redundancy verification, enabling the system to maintain stable perception performance in complex electromagnetic environments and interference scenarios.

[0034] The following two examples further illustrate specific embodiments of the present invention: Example 1: Autonomous Driving Urban Road Scenario This embodiment is applied to urban road environment perception for autonomous passenger vehicles. It needs to realize the positioning, ranging and speed measurement of dynamic targets such as vehicles, pedestrians and non-motorized vehicles, and cope with complex situations such as tall buildings blocking the view, intersection congestion, changes in lighting and electromagnetic interference. This system ensures perception accuracy and driving safety and is adaptable to multiple scenarios such as urban expressways, main roads and residential roads.

[0035] I. Implementation Details of Core Modules Multi-source sensor unit deployment: A lidar, GPS positioning module, IMU (Inertial Measurement Unit), and vision camera are integrated on the vehicle's roof and front bumper. The lidar employs pulse time-of-flight ranging technology, emitting nanosecond-level laser pulses with a detection range of 0.5 meters to 200 meters. Its sampling frequency is adaptively adjustable from 10Hz to 20Hz, outputting target 3D coordinates and relative velocity data. The GPS positioning module supports multi-system joint positioning, receiving signals from at least six satellites. It uses real-time dynamic differential positioning technology, outputting absolute position data with timestamps, and updating at a frequency of 1Hz to 5Hz. The IMU incorporates a three-axis gyroscope and a three-axis accelerometer, with a sampling rate of 100Hz. Through zero-bias calibration and temperature compensation algorithms, it acquires angular velocity and acceleration data in real time for blind spot positioning in short-term occlusion scenarios. The vision camera uses a high-frame-rate global shutter sensor with a resolution of 1920×1080 and a frame rate of 30fps, capturing target texture features and scene semantic information. Each sensor unit is equipped with a standardized data cache module and a CAN bus interface adapter unit to ensure data transmission compatibility.

[0036] The spatiotemporal synchronization module operates as follows: A unified time reference is established based on the millisecond-level high-precision timestamp output by the GPS positioning module. Spatial coordinate transformation is completed through a pre-calibrated sensor extrinsic parameter matrix, calculated from data collected in a standard calibration field, which includes translation vectors and rotation matrices. A linear interpolation algorithm is used to correct the data transmission delay of different sensor units. The delay correction range for LiDAR and vision cameras is 0.1 to 1 millisecond, while the delay correction range for GPS and IMU is 1 to 5 milliseconds. Ultimately, all sensor data is precisely aligned to the vehicle coordinate system, with the synchronization error controlled within 5 milliseconds.

[0037] Data preprocessing module operations: LiDAR point cloud data undergoes statistical filtering to remove discrete noise points. A neighborhood of 15 points is set, with a distance threshold of 0.1 meters. After filtering effective point clouds, voxel gridding downsampling is employed, with voxel sizes of 0.05m × 0.05m × 0.05m to reduce data redundancy. GPS data uses a Kalman smoothing algorithm, with the process noise covariance and observation noise covariance dynamically adjusted based on the urban road scene to eliminate random errors caused by satellite signal obstruction. IMU data uses a zero-bias estimation algorithm to update the zero-bias value every 100 milliseconds, combined with a temperature compensation model to reduce the impact of ambient temperatures ranging from -20℃ to 60℃ on measurement accuracy. Visual images undergo a 5×5 window bilateral filtering algorithm to remove Gaussian noise, followed by histogram equalization to enhance texture contrast and improve the feature recognition of pedestrians, non-motorized vehicles, and other targets.

[0038] The dynamic weight fusion calculation module executes the following: It constructs a multimodal fusion framework combining data-level and feature-level approaches, dynamically allocating fusion weights based on the real-time quality scores of each sensor data point. The quality score is calculated comprehensively based on signal strength, data integrity, and error fluctuation range. Weights are increased when the LiDAR signal strength is above 80 points, and decreased when GPS data integrity is below 90%. The module fuses positioning, ranging, and velocity measurement results using an improved extended Kalman filter algorithm. A CNN-LSTM hybrid deep learning model is introduced, trained on over 100,000 urban road scene samples with 500 iterations and a learning rate of 0.001. This model adaptively identifies intersections, tunnels, and congestion scenarios, optimizing target trajectory and state estimation accuracy, and outputting a unified perception result after fusion.

[0039] The real-time assessment module for sensing accuracy operates by establishing a multi-dimensional evaluation index system that includes positioning error, ranging deviation, velocity stability, and data continuity. Positioning error thresholds are set at 0.1 meters for high accuracy, 0.3 meters for standard accuracy, and 0.5 meters for low accuracy; ranging deviation thresholds are set at 0.05 meters, 0.1 meters, and 0.2 meters; and velocity stability is defined by a variance of 0.1 meters. 2 / s 2 0.3m 2 / s 2 0.5m 2 / s 2 To define the criteria, data continuity was assessed using effective data rates of 95%, 90%, and 85%. A sliding window algorithm was used to analyze evaluation data from the past 10 periods to predict accuracy trends. When the accuracy approached a low-precision threshold for three consecutive periods, an adjustment strategy was triggered in advance.

[0040] Adaptive parameter adjustment module response: At low accuracy levels, such as in tunnels where GPS signals are obstructed, the LiDAR sampling frequency is increased from 10Hz to 20Hz, the GPS data update frequency is increased from 1Hz to 5Hz, the IMU compensation weight coefficient is optimized, the number of visual image feature points detected is increased to 500 per frame, and the ORB feature matching algorithm is strengthened; At standard accuracy levels, such as unobstructed urban main roads, the LiDAR sampling frequency is maintained at 15Hz and the GPS update frequency at 3Hz, and the fusion weights are dynamically fine-tuned based on real-time quality scores; At high accuracy levels, such as in open scenes on urban expressways, the LiDAR sampling frequency is reduced to 10Hz, the GPS update frequency is maintained at 3Hz, the visual camera data transmission bandwidth is reduced, and accuracy and energy consumption are balanced.

[0041] Additional modules include: a fusion weight intelligent optimization unit that dynamically calculates the weights of each sensor, enabling the stable LiDAR and vision camera to play a leading role; an environmental state perception module that analyzes parameters such as LiDAR point cloud density and GPS satellite signal strength to identify scenarios such as tunnel obstruction and dynamic interference at intersections, and matches corresponding fusion strategies; a sensor fault self-diagnosis unit that monitors the operating status of each sensor, and when intermittent anomalies are detected in the vision camera, it removes abnormal data through multi-source data cross-validation; an edge computing processing module that uses a CPU and GPU heterogeneous architecture to deploy computationally intensive tasks at the edge, keeping the perception result output latency within 20 milliseconds; a multi-dimensional anti-interference enhancement module that configures pseudo-random code pulse coding for the LiDAR and enables an anti-multipath interference antenna for the GPS module to resist urban electromagnetic interference; a remote upgrade and self-learning update module that receives algorithm update packets through a TLS encrypted link, and a self-learning module that periodically optimizes the fusion model; and a perception result visualization module that displays the target trajectory and accuracy indicators on the vehicle terminal, allowing engineers to view weight allocation and adjustment records.

[0042] Table 1: Comparison of Perception Performance of Autonomous Driving on Urban Roads

[0043] Table 1 clearly demonstrates the advantages of this invention in the application of autonomous driving in urban road scenarios. Traditional systems use fixed-weight fusion, resulting in an average positioning error of 0.45 meters, poor speed measurement stability, and a fault tolerance rate of only 65% ​​in occluded scenarios, failing to meet the safety requirements of autonomous driving. This invention, through dynamic weight fusion and adaptive parameter adjustment, reduces the average positioning error to 0.12 meters and the speed measurement stability variance to 0.08 meters. 2 / s 2 This significantly improves perception accuracy. In occluded scenarios, the fault tolerance rate reaches 94% through IMU blind spot compensation and multi-frame data correlation; in electromagnetic interference environments, multi-dimensional anti-interference design maintains an accuracy rate of 92%. At the same time, energy consumption is reduced by 28% in high-precision scenarios, achieving a dynamic balance between accuracy and energy consumption, providing reliable perception support for autonomous driving.

[0044] Example 2: UAV mapping of field scenes This embodiment is applied to field terrain mapping of medium-sized UAVs. It needs to realize 3D terrain modeling, obstacle distance measurement and UAV self-positioning. It should meet the challenges of complex mountainous terrain, tree obstruction, weak satellite signal and long-term operation energy consumption constraints. This system ensures mapping accuracy and flight safety and is adaptable to switching between multiple terrains such as mountains, plains and forests.

[0045] I. Implementation Details of Core Modules Multi-source sensor unit deployment: A lidar, GPS positioning module, IMU (Inertial Measurement Unit), and visual camera are integrated on the underside and top of the UAV fuselage. The lidar has a detection range of 1 meter to 150 meters, an adaptively adjustable sampling frequency of 8Hz to 16Hz, and outputs 3D terrain coordinates and obstacle distance data. The GPS positioning module supports dual-mode positioning with BeiDou and GPS, employs real-time dynamic differential technology, receives signals from at least 5 satellites, updates at a frequency of 1Hz to 4Hz, and outputs absolute position data with timestamps. The IMU has a sampling rate of 200Hz, a built-in temperature compensation module to adapt to outdoor temperature differences of -10℃ to 50℃, and collects angular velocity and acceleration data in real time for positioning assistance when satellite signals are blocked. The visual camera has a resolution of 2560×1440 and a frame rate of 20fps, capturing terrain texture and obstacle features to assist in semantic recognition. Each sensor unit is equipped with a standardized data cache module and an Ethernet interface adapter to ensure stable high-altitude data transmission.

[0046] The spatiotemporal synchronization module operates as follows: A unified time reference is established based on the millisecond-level timestamps of the GPS positioning module. Spatial coordinate transformation is completed through a pre-calibrated sensor extrinsic parameter matrix, which is optimized through field calibration testing. A linear interpolation algorithm is used to correct data transmission latency. The latency correction range for LiDAR and visual camera is 0.2 to 1.5 milliseconds, and for GPS and IMU, it is 1 to 6 milliseconds. All sensor data is aligned to the UAV's body coordinate system, and the synchronization error is controlled within 6 milliseconds.

[0047] Data preprocessing module operations: LiDAR point cloud data undergoes statistical filtering to remove noise points, with a neighborhood of 20 points and a distance threshold of 0.15 meters. Voxel grid downsampling is employed, with voxel sizes of 0.1m × 0.1m × 0.1m, balancing data volume and mapping accuracy. GPS data uses a Kalman smoothing algorithm to dynamically adjust noise covariance and eliminate random errors caused by satellite signal obstruction in mountainous terrain. IMU data is updated with zero bias every 50 milliseconds, and a temperature compensation model is used to reduce the impact of outdoor temperature differences on drift errors. Visual images are denoised using a 7×7 window bilateral filtering algorithm, and the CLAHE algorithm enhances local contrast, improving the feature recognition of forest obstacles.

[0048] The dynamic weight fusion calculation module performs the following: It constructs a data-level and feature-level fusion framework, assigning fusion weights based on the real-time quality scores of each sensor. Weights are increased when the GPS satellite signal strength is above 70 points, and decreased when the LiDAR point cloud density is below 50 points / square meter. The module fuses positioning and ranging results using an improved extended Kalman filter algorithm, and introduces a CNN-LSTM hybrid deep learning model. Trained on over 80,000 field scene samples with 400 iterations and a learning rate of 0.0015, it can adaptively identify mountainous, forested, and plain scenes, optimizing terrain 3D modeling accuracy and obstacle localization accuracy, and outputting unified perception results.

[0049] The real-time assessment module for sensing accuracy operates as follows: A multi-dimensional assessment index system is established, with positioning error thresholds set at 0.2 meters for high accuracy, 0.5 meters for standard accuracy, and 0.8 meters for low accuracy; ranging deviation thresholds set at 0.1 meters, 0.2 meters, and 0.3 meters; and velocity measurement stability variance threshold at 0.2 meters. 2 / s 2 0.4m 2 / s 2 0.6m 2 / s 2 The data continuity efficiency thresholds are 90%, 85%, and 80%. The sliding window algorithm analyzes the evaluation data of the past 8 periods to predict the trend of accuracy changes. When the accuracy approaches the low accuracy threshold for two consecutive periods, the adjustment strategy is triggered.

[0050] Adaptive parameter adjustment module response: In low-precision scenarios such as forest areas with satellite signal obstruction, the LiDAR sampling frequency is increased from 8Hz to 16Hz, the GPS update frequency is increased from 1Hz to 4Hz, the IMU compensation weights are optimized, the number of visual image feature points detected is increased to 400 per frame, and the feature matching algorithm is strengthened; In standard-precision scenarios such as open plains, the LiDAR sampling frequency is maintained at 12Hz and the GPS update frequency at 2Hz, and the fusion weights are dynamically fine-tuned; In high-precision scenarios such as unobstructed mountainous areas, the LiDAR sampling frequency is reduced to 8Hz, the GPS update frequency is maintained at 2Hz, the visual camera frame rate is reduced to 15fps, the drone's energy consumption is reduced, and the operation time is extended.

[0051] Additional modules include: a fusion weight intelligent optimization unit that dynamically allocates the weights of each sensor, enabling LiDAR and IMU to play a dominant role in occluded scenarios; an environmental state perception module that analyzes parameters such as LiDAR point cloud density and terrain slope to identify scenarios such as forest occlusion and complex mountainous terrain, matching corresponding fusion strategies; a sensor fault self-diagnosis unit that monitors sensor operating status, reducing the fusion weight of GPS and enhancing IMU acquisition intensity when GPS performance degradation is detected; an edge computing processing module that adopts a heterogeneous computing architecture, deploying data preprocessing and fusion tasks at the edge, controlling the output latency of perception results to within 25 milliseconds; a multi-dimensional anti-interference enhancement module that equips GPS with an anti-multipath interference antenna and enables pulse code modulation for LiDAR to resist electromagnetic interference in the field; a remote upgrade module that receives firmware update packages through an encrypted link; and a self-learning module that optimizes the fusion model based on surveying data; a perception result visualization module that displays the 3D terrain model and accuracy indicators at the ground station, allowing operators to view sensor status and weight allocation records; and a scenario-based accuracy threshold configuration module that loads UAV-specific surveying threshold parameters to ensure that surveying accuracy meets industry standards.

[0052] Table 2: Comparison of UAV Field Mapping and Sensing Performance

[0053] Table 2 data highlights the application value of this invention in UAV field mapping scenarios. Traditional systems cannot adapt to diverse terrain changes, with an average terrain modeling error of 0.65 meters, obstacle ranging accuracy of 0.35 meters, and an accuracy retention rate of only 58% in occluded scenarios, resulting in high energy consumption during long-term operations. This invention, through dynamic weight fusion and environmental adaptive adjustment, reduces the average terrain modeling error to 0.18 meters and obstacle ranging accuracy to 0.09 meters, meeting the requirements for high-precision mapping. In occluded scenarios, through IMU blind spot compensation and multi-source data fusion, the accuracy retention rate reaches 91%; multi-terrain adaptability reaches 95%, stably handling scene switching such as mountains and forests. Energy consumption during long-term operations is reduced by 32%, extending the UAV's endurance and providing an efficient and reliable perception solution for field mapping.

[0054] refer to Figure 2This figure visually demonstrates the superior perception accuracy of this invention in diverse scenarios. Traditional fixed-weight fusion systems achieve 82% accuracy on urban roads and 85% on highways, but their accuracy drops drastically to 60% in tunnels due to GPS signal obstruction, and only 70% in rainy conditions. This invention, through a dynamic weight allocation mechanism, achieves 95% accuracy on urban roads by leveraging the advantages of LiDAR and visual cameras; 97% accuracy on highways by combining GPS high-precision positioning; in tunnel scenarios, the system automatically increases the weights of the IMU and visual camera, achieving 90% accuracy through multi-frame data association and feature supplementation; in rural road scenarios, it optimizes the LiDAR scanning mode, achieving 92% accuracy; and in rainy conditions, it achieves 88% accuracy through image enhancement and point cloud filtering algorithms. The data shows that this invention maintains high-precision perception across various scenarios, significantly outperforming traditional systems.

[0055] refer to Figure 3 This figure illustrates the real-time advantages of this invention under large data volumes. Traditional centralized processing systems operate at a frequency of 25Hz for 50,000 data points / second and only 5Hz for 400,000 data points / second, failing to meet real-time requirements. This invention, through an edge computing architecture, maintains a frequency of 35Hz for 50,000 data points / second and 25Hz for 400,000 data points / second. For 200,000 data points / second, the traditional system operates at 15Hz, while this invention operates at 30Hz, doubling the frequency. Edge computing deploys data preprocessing and fusion tasks at the edge, reducing data transmission latency, while optimizing computational efficiency through model lightweighting techniques. Data shows that this invention maintains a high output frequency under varying data volumes, making it suitable for high-real-time applications such as autonomous driving, significantly outperforming traditional systems.

[0056] refer to Figure 4 This figure highlights the stability of the present invention under complex interference environments. Under interference-free conditions, both systems maintain 100% accuracy. At interference intensity 20, the traditional system maintains 85% accuracy, while the present invention maintains 98%, achieved through a multi-dimensional anti-interference algorithm, pseudo-random code pulse coding for the lidar, and an anti-multipath interference antenna for the GPS. At interference intensity 40, the traditional system maintains 70% accuracy, while the present invention maintains 95%, thanks to interference detection algorithms that identify and remove abnormal data, and multi-source data redundancy verification that ensures accuracy. At interference intensity 60, the traditional system maintains 55% accuracy, while the present invention maintains 92%, thanks to the effectiveness of IMU electromagnetic shielding and anti-glare processing for the visual camera. At interference intensity 80, the traditional system maintains only 40% accuracy, while the present invention maintains 88%, maintaining high accuracy through dynamic adjustment of the fusion strategy and enhanced anti-interference processing. The data demonstrates that the present invention maintains stable performance under different interference intensities, significantly outperforming the traditional system.

[0057] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-sensor fusion perception accuracy adaptive system, characterized in that, The following modules are included: Multi-source sensing unit: integrates laser radar, GPS positioning module, IMU inertial measurement unit and visual camera; calculate the target three-dimensional coordinates and speed through laser radar, output positioning data with timestamp through GPS positioning module, collect angular velocity and acceleration data through IMU inertial measurement unit, and capture target texture features through visual camera; Space-time synchronization module: establish a unified time reference based on timestamp, unify space coordinate conversion through sensing external parameter matrix, and correct the data transmission delay difference of different sensing units by using linear interpolation algorithm, align multi-sensor data to the same space-time coordinate system; Data preprocessing module: laser radar point cloud data is filtered by statistical filtering to remove noise points, voxel grid is reduced to sample, and core features are retained; Kalman smoothing algorithm is used for GPS data to eliminate random errors; IMU data is corrected by zero offset estimation algorithm to eliminate drift errors; visual image is denoised by bilateral filtering and histogram equalization to enhance contrast; Dynamic weight fusion calculation module: construct a multi-modal fusion framework, based on real-time quality score of sensing data, dynamically allocate fusion weight, fuse positioning, ranging and speed measurement results through improved extended Kalman filter algorithm, introduce CNN-LSTM hybrid model to adaptively adjust fusion strategy, and output unified perception results; Real-time perception accuracy evaluation module: establish a multi-dimensional evaluation index system, divide the accuracy level according to the preset scene threshold, analyze historical data combined with sliding window algorithm, and predict the accuracy trend; Adaptive parameter adjustment module: trigger corresponding strategy according to accuracy evaluation results, correct sensor sampling and update frequency, compensation weight and feature extraction when accuracy is lower than threshold; maintain core parameters and adjust fusion weight when accuracy is standard; reduce unit load and balance accuracy and energy consumption when accuracy is higher than threshold; Output and feedback control module: output fusion perception results through standardized format, real-time feedback module running state and perception accuracy level, and notify external system to take safety measures.

2. The multi-sensor fusion perception accuracy adaptive system of claim 1, wherein, Further comprising a fusion weight intelligent optimization unit, which constructs a weight calculation model based on sensing data quality, environmental adaptability and historical error, and optimizes the model expression as wherein is the fusion weight of the th sensing data, is the data quality weight coefficient, is the environmental adaptability weight coefficient, is the real-time quality score of the th sensing data, is the historical error decay coefficient of the th sensing data, is the current environmental adaptability of the th sensing data, is the redundancy contribution degree of the th sensing data, is the number of sensing data participating in fusion, and is the sensing serial number. 3.The multi-sensor fusion perception accuracy adaptive system of claim 1, wherein, It also includes an environmental state perception module, which identifies the current environment type by analyzing environmental feature parameters, sets threshold ranges for feature parameters of each environment type, establishes an environment type and sensing adaptation rule library, and performs corresponding processing strategy.

4. The multi-sensor fusion perception accuracy adaptive system of claim 1, wherein, Also included is a sampling frequency self-adaptive adjusting unit, which dynamically adjusts the sampling frequency of each sensing unit based on the sensing accuracy requirement and the environment complexity, and the adjusting model expression is , wherein is the target sampling frequency of the th sensing unit in the th environment, is the reference sampling frequency of the th sensing unit, is the frequency adjusting coefficient, is the current accuracy requirement coefficient of the th sensing unit, is the complexity coefficient of the th environment.

5. The multi-sensor fusion perception accuracy adaptive system of claim 1, wherein, It also includes a sensing fault self-diagnosis and redundancy mechanism, the fault self-diagnosis unit identifies fault types by monitoring data continuity, signal strength and error fluctuation, the redundancy mechanism automatically enables the complementary function of the remaining sensors when the sensor completely fails, reallocates fusion weight and adjusts preprocessing algorithm parameters; when the performance of the sensor is attenuated, reduce the fusion weight of the sensor, and enhance the acquisition intensity and data processing accuracy of other sensors; When the sensor appears intermittent abnormalities, start the data verification mechanism, remove abnormal data segments by cross-verification of multi-source data, and supplement missing data by interpolation algorithm.

6. The multi-sensor fusion perception accuracy adaptive system of claim 1, wherein, Also includes edge computing processing module, the module adopts CPU and GPU heterogeneous computing architecture, CPU selects multi-core processor responsible for logical control and data scheduling, GPU is configured special computing core responsible for parallel data processing, and the computation-intensive task is deployed in the edge end execution;Through model lightening technology, the deep learning fusion model is pruned and quantized, the redundant network layer is removed, and the parameter precision is adjusted to the preset range;Adopt data hierarchical cache strategy, the key data related to target sensing is stored in cache priority, and the non-key data adopts asynchronous processing mode, so that the system sensing result output delay is controlled in the millisecond level range required by the application scene.

7. The multi-sensor fusion perception accuracy adaptive system of claim 1, wherein, Also includes a scenario-based precision threshold configuration module, which presets a precision threshold parameter library for multiple application scenarios, including automatic driving high-speed scenarios, robot warehouse scenarios, unmanned aerial vehicle surveying and mapping scenarios, and indoor positioning scenarios. Each scenario corresponds to differentiated positioning error, ranging deviation and velocity stability threshold. Users can select corresponding configuration or customize threshold parameters according to actual application scenarios, and the system dynamically adjusts adaptive strategies based on configured thresholds. Meanwhile, threshold parameters support online updating and offline calibration, and threshold settings are adjusted by importing actual scene test data. 8.The multi-sensor fusion perception accuracy adaptive system of claim 1, wherein, Also includes a sensing result visualization and interaction module, which uses a Cartesian three-dimensional coordinate system to display target position, motion trajectory and sensor coverage range in real time, and presents precision indicators in combination with numerical panels and dynamic charts. It provides a graphical interactive interface, allowing users to view the working status of each sensing unit, fusion weight distribution and adaptive adjustment records. Users can manually intervene in fusion strategies and parameter configurations, and support custom fusion weights and sampling frequencies. It supports historical playback and analysis of sensing data, and traces back the sensing process and precision changes along the time axis. 9.The multi-sensor fusion perception accuracy adaptive system of claim 1, wherein, Also includes a remote upgrade and self-learning update module. The remote upgrade module receives system firmware and algorithm model update packages through a TLS encrypted communication link, supports online upgrade and breakpoint resume, retains original configuration data during upgrade, and automatically rolls back to the original version in case of upgrade failure. The self-learning update module continuously adjusts dynamic weight fusion models, environment recognition rules and adaptive adjustment strategies based on massive historical sensing data and precision feedback, and regularly updates sensing fault diagnosis models and precision threshold parameter libraries. The adjustment process is executed asynchronously in the background and does not affect the normal operation of the system.

10. The multi-sensor fusion perception accuracy adaptive system of claim 1, wherein, Also includes a multi-dimensional anti-interference enhancement module. The module uses pseudo-random code pulse coding modulation technology to resist same-frequency interference for laser radar, and realizes signal identification and separation by virtue of unique code design. The GPS module integrates an anti-multipath interference antenna and an adaptive signal filtering algorithm to suppress the influence of multipath reflection signals on positioning accuracy. The IMU unit is configured with an electromagnetic shielding structure and a vibration isolation device. The visual camera uses adaptive exposure control and anti-glare processing technology. Meanwhile, an interference detection algorithm is introduced at the data fusion level to identify interfered sensing data by analyzing data consistency and error distribution, and to reduce the influence of interference on sensing accuracy with the help of multi-source data redundancy verification.

Images