Unmanned aerial vehicle obstacle avoidance test method based on multiple sensors and virtual-real combination

Abstract

The invention discloses an unmanned aerial vehicle obstacle avoidance test method based on multi-sensor and virtual-real combination, and the method comprises the steps: firstly constructing an environment complexity grading model based on the multi-source sensor data of an unmanned aerial vehicle, dynamically selecting a low-complexity environment or a high-complexity environment according to a grading result, and predicting data return delay; evaluating the timeliness and precision of the returned data; then, on the basis of the obtained environment complexity grading result and returned data evaluation information, multiple extreme flight scenes are constructed through a virtual environment simulation platform; and finally, in combination with a virtual test result, verifying the stability and accuracy of the sensing and avoiding system in the typical-near limit state combined flight state through an actual flight test. According to the invention, on the basis of obtaining the data returned by various sensors of the unmanned aerial vehicle, the processed and evaluated data is imported into the virtual environment simulation platform for testing and actual flight to evaluate the performance, so that the safety, economy and timeliness of the unmanned aerial vehicle perception avoidance capability test can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and specifically to a UAV obstacle avoidance testing method based on multiple sensors and a combination of virtual and real technologies. Background Technology

[0002] In recent years, drones have been widely used in numerous fields such as logistics and delivery, agricultural plant protection, surveying and mapping, and film and television production. As the application scenarios for drones become increasingly complex and intensive, the types and numbers of obstacles they encounter during flight are constantly increasing, placing extremely high demands on their perception and avoidance capabilities. Reliable perception and avoidance capabilities are crucial to ensuring the safe flight of drones and preventing collisions; therefore, conducting scientific and accurate testing of drone perception and avoidance capabilities is particularly important.

[0003] Currently, existing methods for testing the perception and avoidance capabilities of unmanned aerial vehicles (UAVs) have certain limitations. Some tests rely on a single sensor, which cannot fully simulate the complex environmental information in real flight scenarios, resulting in test results that do not accurately reflect the performance of UAVs in practical applications. On the other hand, some simulation tests are disconnected from actual flight conditions and lack consideration of real environmental factors, which affects the reliability and effectiveness of the tests. Summary of the Invention

[0004] Purpose of the invention: To address the shortcomings of existing technologies, this invention proposes a drone obstacle avoidance testing method based on multiple sensors and a combination of virtual and real technologies.

[0005] Technical solution: A drone obstacle avoidance testing method based on multiple sensors and virtual-real integration, comprising the following steps:

[0006] (1) Construct an environmental complexity classification model based on UAV multi-source sensor data, and dynamically select the PCA-Lasso-Polynomial (PLP) model (low complexity environment) or Pearson distribution model (high complexity environment) according to the classification results to predict the data back transmission delay and evaluate the timeliness and accuracy of the back transmission data.

[0007] (2) Based on the obtained environmental complexity classification results and feedback data evaluation information, virtual tests are conducted. That is, various extreme flight scenarios are constructed through a virtual environment simulation platform, including static obstacles, dynamic obstacles and environmental disturbances, to conduct virtual tests of UAV perception and avoidance. This virtual test is used to screen key scenarios, pre-verify algorithm logic and preliminarily evaluate the behavior of the perception system under extreme conditions, and provide optimized test scenarios and parameters for actual flight tests.

[0008] (3) Combining the results of the virtual test, the stability and accuracy of the perception system under the typical-near-limit state combination flight state were verified through actual flight tests. Based on the results of the virtual test and the actual test, the key performance indicators were evaluated by statistical analysis methods to ensure that the UAV has efficient and reliable perception and avoidance capabilities under various flight conditions.

[0009] Further, step (1) includes the following steps:

[0010] (11) Collect sensor data related to UAV flight, obtain information on the quantity and type of various sensors including cameras, lidar, and infrared sensors, and collect data related to computing resources, including GPU computing power, to assess its impact on data processing:

[0011] X = {S1,S2,…,S} n} (1)

[0012] Where X is the dataset collected by the sensor, and S1, S2, ..., S n Data representing different sensors;

[0013] (12) Establish an environmental complexity classification model. When the following conditions are met simultaneously, it is judged as a simple situation, namely the PCA-Lasso-Polynomial (PLP) model (low complexity environment):

[0014] a1) Obstacle distance ≥ z d (z d ∈[3,6], unit m) and quantity ≤ z n (z n ∈[2,5]);

[0015] b1) Meteorological condition level ≤ z l (z l ∈[1,10], where level 1 represents the most ideal flight weather conditions, level 10 represents the worst and highest risk weather conditions, and the level value increases with the increase of weather risk.

[0016] c1) The drone operates by flying in a straight line at a constant speed;

[0017] d1) The flight mission is a routine inspection;

[0018] A situation is considered complex, i.e., a high-complexity environment, if any of the following conditions are met:

[0019] a2) Obstacle distance <z d or quantity > z n ;

[0020] b2) Meteorological conditions level > zl ;

[0021] c2) The drone performs an emergency stop / obstacle avoidance maneuver;

[0022] d2) Flight missions involve dynamic target tracking;

[0023] Environmental complexity quantification formula:

[0024]

[0025] Where Result is the quantification result of environmental complexity, simple represents a simple mode, complex represents a complex mode, and d obs It is the nearest obstacle distance, n obs It is the number of obstacles, l w A represents the weather condition level, and A represents the drone operation type. A = {a} c e s e a}, where a c e s e a These represent constant speed, sudden stop, and obstacle avoidance, respectively; T is the task type, T = {t} r , t d}, where t r t d These respectively represent routine inspections and dynamic tracking;

[0026] (13) For simple situations, the PCA-Lasso-Polynomial (PLP) model described in step (12) is used. This model calculates the delay prediction time using polynomial regression after dimensionality reduction by principal component analysis. For complex situations, the Pearson distribution model described in step (12) is used. This model calculates the delay prediction time based on the covariance matrix of multidimensional sensor data. The applicability of the two models is verified by evaluating the timeliness and accuracy indicators of the returned data in real time. That is, the timeliness indicator requires a delay time ≤ 50ms, and the accuracy indicator requires a positioning error ≤ 0.1m.

[0027] Furthermore, the specific process of predicting data backhaul delay using the PCA-Lasso-Polynomial (PLP) model in step (13) is as follows:

[0028] (1311) Calculate the covariance matrix C of the input feature data matrix X′:

[0029]

[0030] Where the superscript T indicates matrix transpose, and n represents the number of data samples;

[0031] (1312) Principal component analysis (PCA) is performed based on the covariance matrix C. The eigenvectors corresponding to the first k largest eigenvalues ​​are selected to reduce the dimensionality of the original features, resulting in the dimensionality-reduced feature matrix X. reduced :

[0032] X reduced =X′V k (4)

[0033] (1313) Using the Lasso regression model, X reduced As input features, the loss function including L1 regularization is minimized to filter key features and obtain regression coefficients δ:

[0034]

[0035] Among them, y i Let X be the actual delay time of the i-th sample. reduced,i Let λ be the dimensionality-reduced feature vector of the i-th sample, λ be the penalty parameter of the L1 regularization term, n be the number of samples, p be the number of features after dimensionality reduction, and δ be the number of features after dimensionality reduction. j The j-th element of the regression coefficient matrix;

[0036] (1314) Based on the key features selected by Lasso regression, namely δ j Features ≠ 0, construct the polynomial characteristic matrix X poly :

[0037] X poly =[1,X selected ,X selected 2 ,X selected 3 ,…,X selected d (6)

[0038] Among them, X selected d represents the subset of key features selected by step (1313), where d is the highest degree of the polynomial;

[0039] (1315) Perform multinomial regression using ordinary least squares, with X poly As input, fit the final delay prediction model and output the predicted delay time for a simple environment.

[0040]

[0041] Final Simple Environment Prediction Delay Time It can be calculated using the following formula:

[0042]

[0043] Where θ is the parameter vector obtained from multinomial regression training, and X poly Input the polynomial feature vector constructed in step (1314) into the new sample.

[0044] Furthermore, the specific process of predicting data return delay using the Pearson distribution model in step 1 is as follows:

[0045] (1321) The delay time x is modeled using the probability density function of the Pearson distribution:

[0046]

[0047] Where x represents the predicted data backhaul delay time, α is the location parameter, β is the scale parameter, k is the shape parameter, and Γ(k) is the gamma function;

[0048] (1322) Based on observed historical delay time samples x i (i = 1, 2, ..., n), the model parameters α, β, k are estimated by maximizing the log-likelihood function:

[0049]

[0050] (1323) The optimal parameter corresponding to the maximum value of L(α,β,k) Substitute the values ​​into the distribution defined in formula (9) and calculate the mode of this distribution as the final prediction delay time for complex environments.

[0051]

[0052] Furthermore, step 2 includes the following steps:

[0053] (21) Virtual Flight Environment Construction and Scene Setting: Creating multi-level virtual flight scenes based on a 3D simulation engine, including:

[0054] Static obstacle layer: Generates fixed obstacles such as buildings and mountains based on terrain features, and supports customization of obstacle size, density and distribution parameters;

[0055] Dynamic obstacle layer: Import motion models of dynamic targets such as aircraft, pedestrians, and animals, set their motion paths, speeds, and random change rules, and simulate sudden movement behavior;

[0056] Environmental disturbance layer: Wind field disturbance is simulated by a fluid dynamics model, and weather effects such as rain, fog and high temperature are generated by a particle system. Wind speed, precipitation intensity and sensor noise parameters are correlated to realize the quantitative mapping of the impact of meteorological conditions on sensor performance.

[0057] (22) Mission planning and dynamic scene interaction: Define the UAV flight mission chain (takeoff, cruise, obstacle avoidance, landing) and embed real-time interaction logic:

[0058] Delay compensation mechanism: Predict the delay time based on the simple environment. Or predict delay time in complex environments Synchronously adjust the lead time of dynamic obstacle movement trajectories in the virtual environment to ensure the timing consistency of the UAV perception-decision-execution link;

[0059] Multi-objective cooperative simulation: Simulate the cooperative avoidance behavior of other aircraft and the random movement paths of pedestrians and animals on the ground through kinematic models, and construct a two-way interactive feedback between UAVs and dynamic obstacles;

[0060] Environmental parameter coupling: Correlate wind speed, rainfall intensity and power consumption of UAV power system to dynamically affect flight stability and obstacle avoidance strategy trigger threshold;

[0061] (23) Multimodal perception and obstacle avoidance performance test:

[0062] Sensor simulation and fusion verification: Radar point cloud, camera image and lidar 3D data are generated by physics engine, and noise such as signal attenuation and optical distortion are injected to test the robustness of data fusion of multiple sensors under conditions such as rain and fog obstruction and strong light interference.

[0063] Full-scenario obstacle avoidance strategy testing: Verify the global optimality of path planning in static obstacle-dense areas, evaluate obstacle avoidance reaction time and path replanning efficiency in dynamic obstacle conflict scenarios, and statistically analyze obstacle avoidance success rate;

[0064] Extreme condition stress test: simulates flight attitude deviation under strong wind disturbance, sensor frequency reduction operation mode under low power state, and the response capability of redundant obstacle avoidance mechanism when component failure (such as monocular camera failure).

[0065] (24) Data-driven evaluation and dynamic optimization:

[0066] Multi-indicator quantitative analysis: Collect data such as sensor false alarm rate, obstacle avoidance path smoothness, and emergency avoidance success rate, and construct a comprehensive scoring model that includes perception accuracy weight and decision timeliness coefficient;

[0067] Algorithm Iteration and Optimization: Based on the test results, the feature extraction parameters of the obstacle recognition algorithm are corrected in reverse, and the adaptability of the path planning strategy in dynamic scenes is trained through reinforcement learning, so as to realize the dual-track optimization of obstacle avoidance algorithm by online learning and offline parameter tuning.

[0068] (25) System collaborative verification and reliability testing:

[0069] Sensor fusion performance verification: Compare the obstacle avoidance misjudgment rate under single sensor and multi-sensor fusion modes to verify the improvement of obstacle localization accuracy by data fusion;

[0070] End-to-end consistency verification: Through joint calibration of the virtual environment and the actual test platform, it is ensured that the error of the perception system output, obstacle avoidance algorithm decision and the action of the UAV actuator is less than the preset threshold, thus ensuring the credibility of the test results.

[0071] Furthermore, step 3, the actual flight test and comprehensive performance evaluation, includes the following steps:

[0072] (31) Test Implementation and Data Acquisition:

[0073] (311) Flight status determination: Before the flight test, determine the set of typical-near-limit state combination flight status parameters, including altitude, speed, angle of attack, and attitude parameters, and select test conditions based on UAV design specifications and mission requirements;

[0074] (312) Initialization of the perception system: complete the self-test and parameter calibration of radar, lidar and vision sensors, verify the integrity of sensor functions through system diagnosis, and establish a reference measurement coordinate system;

[0075] (313) Multimodal flight test: Perform typical-near-limit state combination test, dynamically adjust flight parameters to simulate state transition process, and monitor the operating status of the perception system and the dynamic response of the UAV in real time;

[0076] (314) Full-element data recording: synchronously collect target detection data, tracking trajectory, ranging information, attitude estimation parameters and actuator action logs to build a spatiotemporally aligned multi-source test database;

[0077] (32) Multi-dimensional performance evaluation index system: This index system adopts a composite evaluation mechanism of "triple veto + six levels of compliance". The core classification and recognition capabilities (y1~y3) are subject to strict veto. The dynamic performance index group (y4~y9) must meet the requirements of at least 5 groups of indicators. The test data is used to ensure the objectivity of the evaluation through multi-source spatiotemporal alignment technology. Finally, the full-link performance verification of the perception system reliability, decision-making real-time performance and execution accuracy is judged.

[0078] (321) Core competency veto item:

[0079] a) Obstacle classification accuracy R OC :

[0080]

[0081] Where, n OC N represents the number of times the drone correctly classifies obstacles.C The total number of times the drone was classified as an obstacle;

[0082] b) False detection rate R FD :

[0083]

[0084]

[0085] Where, n FD N represents the number of times the drone falsely reported obstacles. D Total number of drone inspections;

[0086] c) Success rate of avoidance R OAS :

[0087]

[0088] Where, n OAS N represents the number of times the drone successfully avoided an obstacle. OA The total number of avoidance attempts by the drone;

[0089] (322) Dynamic performance grading index:

[0090] a) Obstacle avoidance reaction time t R :

[0091] t R =t start -t detect (18)

[0092]

[0093] Among them, t start The moment when the drone begins its evasive maneuver, t detect The moment when the obstacle is first detected;

[0094] b) Path stability S C :

[0095] S C =σ PD (20)

[0096]

[0097] Where, σ PD The standard deviation of the flight path;

[0098] c) Heading stability S D :

[0099] S D =Δθ CC (twenty two)

[0100]

[0101] Where, Δθ CC This refers to the change in heading;

[0102] d) Path optimization degree D APO :

[0103]

[0104] Among them, L min L represents the shortest barrier-free path length, and L represents the actual obstacle avoidance path length.

[0105] e) Obstacle avoidance accuracy P OA :

[0106] P OA =min(L O (26)

[0107]

[0108] Among them, L O It is the distance to the obstacle at every moment during flight;

[0109] f) Sensor coverage area S SC :

[0110]

[0111] Among them, L C θ is the effective detection range of the sensor. C This refers to the sensor's field of view angle;

[0112] (323) Comprehensive Admission Judgment Rules:

[0113]

[0114] Where H is the evaluation score, w i , i = 1, 2, ..., 9, are the evaluation criteria and judgment parameters, and F is the final evaluation result.

[0115] The advantages of this invention compared to the prior art are:

[0116] This invention employs a hybrid approach, combining virtual and real-world simulations with multi-sensor fusion to accurately capture various environmental data. The complementary nature of simulation and real-world testing allows for comprehensive and realistic evaluation of a drone's perception and obstacle avoidance capabilities, improving the accuracy and reliability of test results. Simulated testing optimizes algorithms beforehand, reducing the number of real-world tests and costs, and lowering testing risks. Through repeated testing and optimization, the drone's obstacle avoidance performance and environmental adaptability are enhanced. Simultaneously, it provides the industry with scientific testing standards, promotes multi-domain technological integration and innovation, and contributes to the standardized development of the drone industry. Attached Figure Description

[0117] Figure 1 The flowchart is for the UAV obstacle avoidance testing method based on multiple sensors and virtual-real integration according to the present invention. Detailed Implementation

[0118] The technical solution of the present invention in specific implementation will be further described below with reference to the accompanying drawings.

[0119] Figure 1 The diagram shows the flowchart of the UAV obstacle avoidance testing method based on multiple sensors and virtual-real integration according to the present invention. The UAV obstacle avoidance testing method based on multiple sensors and virtual-real integration according to the present invention mainly includes the following steps:

[0120] Step (1), Multi-source sensor data acquisition and preprocessing

[0121] Hardware configuration:

[0122]

[0123] Through the collaborative acquisition of data from multiple sensors, the UAV can obtain real-time, multi-dimensional data on obstacles, weather, and motion status, providing an input basis for subsequent environmental modeling.

[0124] Data input example (meteorological data obtained from a weather station):

[0125]

[0126] Meteorological condition levels are assessed using the following table:

[0127]

[0128] Based on the above input parameters, the system enters the environmental complexity classification process and sets z. d =5m, z n =3, z l =3, decision rule executed:

[0129]

[0130] Conditions were verified item by item: Obstacle condition: 6.2≥5 (true) and 2≤3 (true), satisfied; Weather condition: 2≤3 (true), satisfied; Action type: Uniform linear motion (true); Task type: Routine inspection (true). Judgment result: Environmental complexity is Simple, PLP model is enabled.

[0131] Data standardization: Original feature matrix X∈R 10×4 (10-dimensional features × 4 samples):

[0132]

[0133] The standardized matrix X′ has a mean of 0 and a variance of 1 for each column.

[0134] PCA dimensionality reduction: Calculating the covariance matrix Eigenvalue decomposition yields:

[0135] λ1=2.8, λ2=1.2, λ3=0.7,…

[0136] Select the first 3 principal components (cumulative variance contribution rate 95%), projection matrix V3∈R 10×3 :

[0137]

[0138] Lasso regression training: Objective function (regularization parameter λ = 0.1):

[0139]

[0140] The regression coefficients, obtained using the coordinate descent method, are: δ = [0.32, -0.12, 0.41].

[0141] Polynomial feature extension: for the dimensionality-reduced feature X reduced,1 =[0.82,-0.15,1.48] generates quadratic terms:

[0142] X poly =[1,0.82,-0.15,1.48,0.82] 2 (-0.15) 2 1.48 2 = [1, 0.82, -0.15, 1.48, 0.672, 0.0225, 2.190]

[0143] Delayed prediction calculation: Polynomial regression parameters θ′=[2.15,0.48,-0.25,0.72,-0.18,0.05,0.38];

[0144] Predicted value:

[0145]

[0146] The measured delay y = 4.5ms, with an error of

[0147] Step (2), Refined Scenario Construction for Virtual Simulation Testing

[0148] Virtual scene parameter settings: Static obstacle layer: Terrain features: City street model, building dimensions 20m×10m×30m, density ρ=0.5 / m 2 Obstacle distribution: generated according to the Poisson process, with a minimum spacing of 5m.

[0149] Dynamic obstacle layer: Vehicle motion model: IDM car-following model, expected speed 15m / s, random change probability p=0.2 / s; Pedestrian motion: social force model, speed 1.2m / s±0.3m / s.

[0150] Environmental disturbance layer: Wind field model: based on Navier-Stokes equations, wind speed 12 m / s (direction randomly fluctuates ±30°); Rain and fog effects: particle density 10 4 / m 3 The lidar attenuation coefficient is 0.1 dB / m.

[0151] Obstacle avoidance strategy verification process: Test case: The UAV encounters a vehicle moving laterally (relative speed 8m / s) at t=10s; Perception-decision link: 1. Multi-sensor fusion positioning error ≤0.1m; 2. Path replanning time t plan =35ms; 3. Actuator response delay t act =15ms. Result: Obstacle avoidance action completion time t total =50ms, minimum obstacle avoidance distance 1.8m (meeting standard ≥0.5m and ≤3m).

[0152] Extreme condition stress test: Sensor frequency reduction mode: When the battery level drops to 10%, the LiDAR sampling rate decreases from 10Hz to 5Hz; Data fusion weight adjustment: Visual confidence increases by 20%, radar confidence decreases by 15%. Test results: Obstacle avoidance success rate remains at 93% (78% with traditional methods).

[0153] Step (3), Multi-state combination test and comprehensive scoring

[0154] Test scenario design: Typical conditions: height h = 30m, speed v = 8m / s, attitude angle θ < 5°; obstacle density ρ = 0.3 / m 2 Dynamic target number n dyn =2.

[0155] Near-limit conditions: Altitude h = 120m, velocity v = 25m / s, crosswind 15m / s; obstacle density ρ = 1.2 / m³ 2 Dynamic target number n dyn =6.

[0156] Core Competency One-Vote Veto Verification: Obstacle Classification Accuracy: Test Sample N C =200, correctly classified n OC =196; (≥90%), w1 = 1.

[0157] False detection rate: Total number of detections N D =500, false alarms n FD =1; (Meets the standard ≤ 0.2%), w2 = 1.

[0158] Avoidance success rate: Avoidance attempts N OA =150, n successes OAS =146; (≥95%) → w3 = 1.

[0159] Calculation of dynamic performance grading index:

[0160] index Measured value <![CDATA[Compliance determination (y i )]]> <![CDATA[Obstacle avoidance reaction time t R > 85ms <![CDATA[w4=1]]> <![CDATA[Path stability S C > 7.8 <![CDATA[w5=1]]> <![CDATA[Course stability S D > 0.25 rad <![CDATA[w6=1]]> <![CDATA[Path optimization degree D APO > 8.2％ <![CDATA[w7=1]]> <![CDATA[Obstacle avoidance accuracy P OA > 1.5m <![CDATA[w8=1]]> <![CDATA[Sensor coverage range S SC > 420π <![CDATA[w9=1]]>

[0161] Comprehensive admission criteria:

[0162]

[0163] Judgment result: F = Pass.

[0164] Quantitative comparison of implementation results:

[0165]

[0166]

[0167] The above description is merely a further detailed explanation of the present invention through a specific embodiment and should not be construed as limiting the invention to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for testing obstacle avoidance of unmanned aerial vehicles (UAVs) based on multi-sensor and virtual-real integration, characterized in that, The method includes the following steps: (1) Construct an environmental complexity classification model based on UAV multi-source sensor data, dynamically select the PCA-Lasso-Polynomial (PLP) model or Pearson distribution model according to the classification results to predict the data back transmission delay, and evaluate the timeliness and accuracy of the back transmission data. (2) Based on the obtained environmental complexity classification results and feedback data evaluation information, virtual tests are conducted. That is, multiple extreme flight scenarios are constructed through a virtual environment simulation platform, including static obstacles, dynamic obstacles and environmental disturbances, to conduct virtual tests of UAV perception and avoidance. This virtual test is used to screen key scenarios, pre-verify algorithm logic and preliminarily evaluate the behavior of the perception system under extreme conditions, and provide optimized test scenarios and parameters for actual flight tests. (3) Combining the results of the virtual test, the stability and accuracy of the perception system under the typical-near-limit state combination flight state were verified through actual flight tests. Based on the results of the virtual test and the actual test, the key performance indicators were evaluated by statistical analysis methods to ensure that the UAV has efficient and reliable perception and avoidance capabilities under various flight conditions.

2. The UAV obstacle avoidance testing method based on multi-sensor and virtual-real integration according to claim 1, characterized in that, Step (1) includes the following steps: (11) Collect sensor data related to UAV flight, obtain information on the quantity and type of various sensors including cameras, lidar, and infrared sensors, and collect data related to computing resources, including GPU computing power, to assess its impact on data processing: X={S1,S2,…,S n } (1) Where X is the dataset collected by the sensor, and S1, S2, ..., S n Data representing different sensors; (12) Establish an environmental complexity classification model. When the following conditions are met simultaneously, it is determined to be a simple situation or a low-complexity environment, i.e., the PCA-Lasso-Polynomial (PLP) model: a1) Obstacle distance ≥ z d (z d ∈[3,6], unit m) and quantity ≤ z n (z n ∈[2,5]); b1) Meteorological condition level ≤ z l (z l ∈[1,10], where level 1 represents the most ideal flight weather conditions, level 10 represents the worst and highest risk weather conditions, and the level value increases with the increase of weather risk; c1) The drone operates by flying in a straight line at a constant speed; d1) The flight mission is a routine inspection; A situation is classified as complex or highly complex when any of the following conditions are met, according to the Pearson distribution model: a2) Obstacle distance <z d or quantity > z n ; b2) Meteorological conditions level > z l ; c2) The drone performs an emergency stop / obstacle avoidance maneuver; d2) Flight missions involve dynamic target tracking; Environmental complexity quantification formula: Where Result is the quantification result of environmental complexity, simple represents a simple mode, complex represents a complex mode, and d obs It is the nearest obstacle distance, n obs It is the number of obstacles, l w A represents the weather condition level, and A represents the drone operation type. A = {a} c e s e a }, where a c e s e a These represent constant speed, sudden stop, and obstacle avoidance, respectively; T is the task type, T = {t} r , t d }, where t r t d These respectively represent routine inspections and dynamic tracking; (13) For simple situations, the PCA-Lasso-Polynomial (PLP) model described in step (12) is used. This model calculates the delay prediction time using polynomial regression after dimensionality reduction by principal component analysis. For complex situations, the Pearson distribution model described in step (12) is used. This model calculates the delay prediction time based on the covariance matrix of multidimensional sensor data. The applicability of the two models is verified by evaluating the timeliness and accuracy indicators of the returned data in real time. That is, the timeliness indicator requires a delay time ≤ 50ms, and the accuracy indicator requires a positioning error ≤ 0.1m.

3. The UAV obstacle avoidance testing method based on multi-sensor and virtual-real combination according to claim 2, characterized in that, The specific process of predicting data backhaul delay using the PCA-Lasso-Polynomial (PLP) model in step (13) is as follows: (1311) Calculate the covariance matrix C of the input feature data matrix X′: Where the superscript T indicates matrix transpose, and n represents the number of data samples; (1312) Principal component analysis (PCA) is performed based on the covariance matrix C. The eigenvectors corresponding to the first k largest eigenvalues ​​are selected to reduce the dimensionality of the original features, resulting in the dimensionality-reduced feature matrix X. reduced : X reduced =X′V k (4) (1313) Using the Lasso regression model, X reduced As input features, the loss function including L1 regularization is minimized to filter key features and obtain regression coefficients δ: Among them, y i Let X be the actual delay time of the i-th sample. reduced,i Let λ be the dimensionality-reduced feature vector of the i-th sample, λ be the penalty parameter of the L1 regularization term, n be the number of samples, p be the number of features after dimensionality reduction, and δ be the number of features after dimensionality reduction. j The j-th element of the regression coefficient matrix; (1314) Based on the key features selected by Lasso regression, namely δ j Features ≠ 0, construct the polynomial characteristic matrix X poly : X poly =[1,X selected ,X selected 2 ,X selected 3 ,…,X selected d ] (6) Among them, X selected d represents the subset of key features selected by step (1313), where d is the highest degree of the polynomial; (1315) Perform multinomial regression using ordinary least squares, with X poly As input, fit the final delay prediction model and output the predicted delay time for a simple environment. Final Simple Environment Prediction Delay Time It can be calculated using the following formula: Where θ is the parameter vector obtained from multinomial regression training, and X poly Input the polynomial feature vector constructed in step (1314) into the new sample.

4. The UAV obstacle avoidance testing method based on multi-sensor and virtual-real combination according to claim 2, characterized in that, The specific process of predicting data return latency using the Pearson distribution model in step 1 is as follows: (1321) The delay time x is modeled using the probability density function of the Pearson distribution: Where x represents the predicted data backhaul delay time, α is the location parameter, β is the scale parameter, k is the shape parameter, and Γ(k) is the gamma function; (1322) Based on observed historical delay time samples x i (i = 1, 2, ..., n), the model parameters α, β, k are estimated by maximizing the log-likelihood function: (1323) The optimal parameter corresponding to the maximum value of L(α,β,k) Substitute the values ​​into the distribution defined in formula (9) and calculate the mode of this distribution as the final prediction delay time for complex environments.

5. The UAV obstacle avoidance testing method based on multi-sensor and virtual-real combination according to claim 1, characterized in that, Step 2 includes the following steps: (21) Virtual Flight Environment Construction and Scene Setting: Creating multi-level virtual flight scenes based on a 3D simulation engine, including: Static obstacle layer: Generates fixed obstacles, including buildings and mountains, based on terrain features, and supports customization of obstacle size, density, and distribution parameters; Dynamic obstacle layer: Import motion models of dynamic targets such as aircraft, pedestrians, and animals, set their motion paths, speeds, and random change rules, and simulate sudden movement behavior; Environmental disturbance layer: Wind field disturbance is simulated by a fluid dynamics model, and weather effects such as rain, fog and high temperature are generated by a particle system. Wind speed, precipitation intensity and sensor noise parameters are correlated to realize the quantitative mapping of the impact of meteorological conditions on sensor performance. (22) Mission planning and dynamic scene interaction: Define the UAV flight mission chain including takeoff, cruise, obstacle avoidance, and landing, and embed real-time interaction logic: Delay compensation mechanism: Predict the delay time based on the simple environment. Or predict delay time in complex environments Synchronously adjust the lead time of dynamic obstacle movement trajectories in the virtual environment to ensure the timing consistency of the UAV perception-decision-execution link; Multi-objective cooperative simulation: Simulate the cooperative avoidance behavior of other aircraft and the random movement paths of pedestrians and animals on the ground through kinematic models, and construct a two-way interactive feedback between UAVs and dynamic obstacles; Environmental parameter coupling: Correlate wind speed, rainfall intensity and power consumption of UAV power system to dynamically affect flight stability and obstacle avoidance strategy trigger threshold; (23) Multimodal perception and obstacle avoidance performance test: Sensor simulation and fusion verification: Radar point cloud, camera image and lidar 3D data are generated by physics engine, and noise including signal attenuation and optical distortion is injected to test the robustness of data fusion of multiple sensors under rain, fog and strong light interference conditions. Full-scenario obstacle avoidance strategy testing: Verify the global optimality of path planning in static obstacle-dense areas, evaluate obstacle avoidance reaction time and path replanning efficiency in dynamic obstacle conflict scenarios, and statistically analyze obstacle avoidance success rate; Extreme condition stress test: simulates flight attitude deviation under strong wind disturbance, sensor frequency reduction operation mode under low power state, and the response capability of redundant obstacle avoidance mechanism when component failure occurs. (24) Data-driven evaluation and dynamic optimization: Multi-indicator quantitative analysis: Collect data on sensor false alarm rate, obstacle avoidance path smoothness, and emergency avoidance success rate, and construct a comprehensive scoring model that includes perception accuracy weight and decision timeliness coefficient; Algorithm Iteration and Optimization: Based on the test results, the feature extraction parameters of the obstacle recognition algorithm are corrected in reverse, and the adaptability of the path planning strategy in dynamic scenes is trained through reinforcement learning, so as to realize the dual-track optimization of obstacle avoidance algorithm by online learning and offline parameter tuning. (25) System collaborative verification and reliability testing: Sensor fusion performance verification: Compare the obstacle avoidance misjudgment rate under single sensor and multi-sensor fusion modes to verify the improvement of obstacle localization accuracy by data fusion; End-to-end consistency verification: Through joint calibration of the virtual environment and the actual test platform, it is ensured that the error of the perception system output, obstacle avoidance algorithm decision and the action of the UAV actuator is less than the preset threshold, thus ensuring the credibility of the test results.

6. The UAV obstacle avoidance testing method based on multi-sensor and virtual-real integration according to claim 1, characterized in that, Step 3, actual flight testing and comprehensive performance evaluation, includes the following steps: (31) Test Implementation and Data Acquisition: (311) Flight status determination: Before the flight test, determine the set of typical-near-limit state combination flight status parameters, including altitude, speed, angle of attack, and attitude parameters, and select test conditions based on UAV design specifications and mission requirements; (312) Initialization of the perception system: complete the self-test and parameter calibration of radar, lidar and vision sensors, verify the integrity of sensor functions through system diagnosis, and establish a reference measurement coordinate system; (313) Multimodal flight test: Perform typical-near-limit state combination test, dynamically adjust flight parameters to simulate state transition process, and monitor the operating status of the perception system and the dynamic response of the UAV in real time; (314) Full-element data recording: synchronously collect target detection data, tracking trajectory, ranging information, attitude estimation parameters and actuator action logs to build a spatiotemporally aligned multi-source test database; (32) Multi-dimensional performance evaluation index system: This index system adopts a composite evaluation mechanism of "triple veto + six levels of compliance". The core classification and recognition capabilities (y1~y3) are subject to strict veto. The dynamic performance index group (y4~y9) must meet the requirements of at least 5 groups of indicators. The test data is used to ensure the objectivity of the evaluation through multi-source spatiotemporal alignment technology. Finally, the full-link performance verification of the perception system reliability, decision-making real-time performance and execution accuracy is judged. (321) Core competency veto item: a) Obstacle classification accuracy R OC : Where, n OC N represents the number of times the drone correctly classifies obstacles. C The total number of times the drone was classified as an obstacle; b) False detection rate R FD : Where, n FD N represents the number of times the drone falsely reported obstacles. D Total number of drone inspections; c) Success rate of avoidance R OAS : Where, n OAS N represents the number of times the drone successfully avoided an obstacle. OA The total number of avoidance attempts by the drone; (322) Dynamic performance grading index: a) Obstacle avoidance reaction time t R : t R =4 start -t detect (18) Among them, t start The moment when the drone begins its evasive maneuver, t detect The moment when the obstacle is first detected; b) Path stability S C : S C =s PD (20) Where, σ PD The standard deviation of the flight path; c) Heading stability S D : S D =Δθ CC (22) Where, Δθ CC This refers to the change in heading; d) Path optimization degree D APO : Among them, L min L represents the shortest barrier-free path length, and L represents the actual obstacle avoidance path length. e) Obstacle avoidance accuracy P OA : P OA <min(L O ) (26) Among them, L O It is the distance to the obstacle at every moment during flight; f) Sensor coverage area S SC : Among them, L C θ is the effective detection range of the sensor. C This refers to the sensor's field of view angle; (323) Comprehensive Admission Judgment Rules: Where H is the evaluation score, w i , i = 1, 2, ..., 9, are the evaluation criteria and judgment parameters, and F is the final evaluation result.

Images