An integrated traffic low-altitude global intelligent perception method based on autonomous evolution

Abstract

The application discloses a kind of comprehensive traffic low altitude global intelligent perception methods based on autonomous evolution, belong to the cross field of intelligent transportation and computer vision.Method includes: collection multi-source perception data and extract structured causal variable;Dynamic causal diagram is constructed, and attribution analysis of perception error is realized by intervention learning;Cross-scene model migration and few-sample self-adaptation are realized based on knowledge graph;Adaptive inference is carried out using dynamic graph neural network of environment perception;Autonomous evolution closed loop driven by causality is realized based on multi-source feedback.The application realizes the leap of perception system from "perception-optimization" to "understanding-evolution" in low-altitude complex traffic environment, significantly improves perception accuracy, scene adaptability and system explainability, and reduces artificial operation cost.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of intelligent transportation and computer vision, and in particular to a comprehensive low-altitude all-domain intelligent perception method for transportation based on autonomous evolution. Background Technology

[0002] With the development of drones, intelligent sensing, and artificial intelligence technologies, low-altitude intelligent sensing has shown great potential in areas such as traffic infrastructure inspection and air traffic management. However, the low-altitude traffic environment is highly complex, dynamic, and heterogeneous, posing serious challenges to the accuracy, robustness, and adaptability of sensing systems.

[0003] Current mainstream technologies are mainly based on static deep learning models trained with large amounts of labeled data, but they still have significant limitations in practical applications: First, the performance of the models drops sharply when dealing with unseen or extreme scenarios (such as severe weather, complex terrain, and new facilities), the root cause of which is the lack of interpretability, and optimization relies on trial and error, which is inefficient; Second, traditional model structures are rigid and cannot dynamically adjust their inference strategies according to real-time environmental conditions (such as lighting, visibility, and sensor health), making it difficult to balance accuracy and efficiency under changing conditions; Third, perceptual knowledge between different scenarios cannot be accumulated in a structured way and transferred efficiently, and each new scenario requires a large amount of data and computing power to retrain, resulting in high system deployment and maintenance costs; Finally, existing systems lack a complete self-evolutionary closed loop, with perception, diagnosis, optimization, and deployment links being fragmented and highly dependent on human intervention and expert experience.

[0004] Therefore, there is an urgent need for an intelligent sensing method that can achieve self-diagnosis, self-optimization, and self-evolution, so as to fundamentally improve the adaptability, interpretability, and sustainable operation and maintenance capabilities of low-altitude traffic sensing systems in complex real-world environments. Summary of the Invention

[0005] The purpose of this invention is to provide a comprehensive low-altitude all-domain intelligent perception method for transportation based on autonomous evolution, in order to solve at least some of the problems existing in the prior art.

[0006] The technical solution, a comprehensive low-altitude all-domain intelligent perception method for transportation based on autonomous evolution, includes the following steps:

[0007] Collect multi-source heterogeneous sensing data in low-altitude traffic scenarios, extract structured causal variables, and obtain a set of causal variables;

[0008] Based on a set of causal variables, a causal graph model is constructed and dynamically updated. Attribution analysis of perceptual errors is performed through intervention learning and counterfactual reasoning to obtain a causal error attribution report.

[0009] Based on the causal error attribution report, matching scenarios are retrieved from the knowledge graph, the parameters of the pre-trained model are transferred, and graph-guided few-sample fine-tuning is performed to obtain a fine-tuned model adapted to the new scenario.

[0010] Based on the fine-tuned model and real-time causal variable set, a perception graph structure is dynamically constructed, and adaptive reasoning is performed through the gated graph attention mechanism of environmental perception to generate target perception results.

[0011] Based on the comparison between the target perception results and multi-source feedback data, causal-driven evolutionary decisions are triggered and executed to update the knowledge system and model library, thus completing the autonomous evolutionary closed loop.

[0012] Beneficial effects: By establishing a causal perception and autonomous evolution mechanism, this invention enables the intelligent perception system to have the ability to self-diagnose and continuously optimize in complex low-altitude traffic environments, solving the problems of performance degradation and insufficient generalization caused by the variability of the environment and the differences in the scenario in traditional methods; This method improves perception accuracy and scenario adaptability, reduces the dependence on massive labeled data and manual tuning, and realizes autonomous and efficient evolution and interpretable operation and maintenance throughout the entire life cycle of the system. Attached Figure Description

[0013] Figure 1 This is a flowchart of the overall solution of the present invention.

[0014] Figure 2 This is a flowchart of the present invention for extracting structured causal variables.

[0015] Figure 3 This is a flowchart of the process of constructing and dynamically updating the causal graph model in this invention.

[0016] Figure 4 This is a flowchart of the dynamic construction of the perception graph structure according to the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0018] Example 1

[0019] This embodiment provides an overall framework for a comprehensive low-altitude, all-domain intelligent perception method for traffic based on autonomous evolution, including:

[0020] S1: Multimodal Data Acquisition and Structured Extraction of Causal Variables

[0021] According to one aspect of this application, this embodiment begins with a systematic multi-source heterogeneous data synchronous acquisition and standardized preprocessing stage. In this embodiment, the sensing carrier (such as an inspection drone or fixed monitoring equipment) simultaneously activates its onboard visible light camera, infrared thermal imager, lidar, and miniature weather station in the low-altitude traffic operation area, using timestamps and spatial coordinates (usually obtained through a high-precision GNSS / RTK module) as a reference to achieve strict spatiotemporal alignment of the data. Specifically, the visible light camera acquires high-resolution RGB image sequences, the frame rate and resolution of which can be dynamically adjusted according to the flight speed to ensure sufficient image overlap in key areas; the infrared thermal imager simultaneously captures thermal radiation distribution images under the same field of view, and its temperature sensitivity needs to be able to distinguish subtle temperature differences on the surface of traffic facilities; the lidar emits laser beams at a specific frequency to acquire the three-dimensional point cloud coordinates and reflection intensity information of the target surface; and the airborne weather sensor records a series of continuous environmental physical quantity data such as light intensity, wind speed, wind direction, humidity, and ambient temperature in real time.

[0022] Furthermore, the acquired raw data is not used directly but enters a standardized preprocessing and feature extraction pipeline. For image data, radiometric and geometric corrections are first performed to eliminate the effects of sensor distortion and uneven illumination. For infrared images, non-uniformity correction based on a blackbody reference is also performed. LiDAR point cloud data undergoes denoising, ground point filtering, and voxel-based downsampling to control data size while preserving structural features. Preferably, quality metrics for each frame of data are calculated in real time, such as the image signal-to-noise ratio, point cloud density, and integrity; these metrics themselves become important variables for subsequent causal analysis.

[0023] After preprocessing, the core causal variable extraction stage begins. Based on further improvements in this embodiment, causal variables are explicitly divided into three categories: environmental interference variables, target-specific variables, and sensor state variables. Environmental interference variables are directly mapped from meteorological data; for example, wind speed sequences are converted into a "turbulence intensity index" representing the potential risk of motion blur, and illumination and humidity are combined to calculate a "visibility attenuation coefficient." Target-specific variables are obtained through shallow analysis of the sensing data itself; for example, "apparent motion consistency" of a specific region of interest is estimated from image sequences using optical flow, "structural stability" is calculated from changes in point clouds across consecutive frames, and "thermal inertia features" of the target are extracted from infrared images. Sensor state variables originate from real-time assessments of data quality, such as "image edge sharpness" and "effective echo ratio of point clouds." All these variables are quantified and normalized, and encapsulated along with their spatiotemporal context information into a structured causal variable feature vector set. This serves as the metadata foundation for the entire autonomous evolutionary sensing system to understand the current sensing context, providing a decision-making basis for subsequent causal inference and model adaptation.

[0024] S2: Dynamic Cause-Effect Graph Construction and Error Source Tracing Process Based on Intervention Learning

[0025] In one alternative implementation, once the structured causal variable set is ready, an online and offline combined causal modeling engine is launched to build and maintain a causal graph model that reveals how changes in environment and state lead to fluctuations in perceived performance. Initially, a prior causal graph trained offline on massive amounts of historical task data is loaded. Nodes in the graph represent various causal variables, directed edges represent hypothetical causal influence directions, and edge weights represent prior estimates of influence strength.

[0026] Specifically, online causal analysis occurs during each real-time perception task execution. The observed causal variables at the current moment are input into the causal graph, and inference is performed based on a probabilistic graphical model (such as a Bayesian network) to calculate the posterior probability of various perception errors (such as missed detection of small targets, inaccurate boundary localization, and misclassification) under the current specific environmental configuration. This step enables the prediction of potential risks. However, a static prior causal graph may not cover all novel scenarios. Therefore, according to a further improvement in this embodiment, a lightweight intervention learning simulator is embedded in the system. Intervention learning is triggered when a significant deviation is detected between the predicted error probability and the subsequent actual feedback error, or when a completely new combination of causal variables with weak explanatory power in the prior graph is encountered.

[0027] In a preferred embodiment, intervention learning takes place in a digital twin environment. First, the current problem context (combination of causal variables) is replicated. Then, a variable suspected of being a key cause (e.g., the "visibility decay coefficient") is actively intervened virtually—adjusted to different levels. Simultaneously, a high-fidelity sensor simulation model is used to generate simulated data that might be collected under the intervention conditions (e.g., a more blurred image). Subsequently, the current perception model is used to infer from this simulated data, observing the changing trends in perception outcomes. Through a series of such virtual experiments, the causal effect of the variable on perception outcomes can be quantified, thereby confirming or disproving prior causal relationships, and even discovering new causal links. The newly confirmed causal knowledge is used to update the topology and parameters of the online causal graph in real time, enabling the causal model to evolve autonomously as the scenario evolves. Ultimately, for each perception error, not only is the error itself recorded, but the updated causal graph can be queried to pinpoint the most likely root cause variable or combination of variables leading to the error, achieving deep error attribution and providing a precise direction for subsequent evolution.

[0028] S3: Knowledge Graph-Driven Scene Adaptation and Rapid Model Transfer Process

[0029] Based on the deep error tracing results, if the system determines that the performance degradation is mainly due to the migration or variation of the scene (e.g., from daytime urban bridge inspection to nighttime mountain tunnel inspection), it will activate the knowledge graph-driven migration mechanism in this step. In this embodiment, a structured semantic network of a low-altitude traffic scene knowledge graph is constructed. It does not directly store image or point cloud data, but encodes the abstract concepts of different scenes, their logical connections, and related model behavior knowledge in the form of entity-relationship-attribute.

[0030] For example, this knowledge graph might contain entities such as "mountain tunnel," "strong low light," "concrete surface," and "water seepage defects." These entities are connected by relationships such as "having environment," "common materials," and "prone anomalies." Each entity or relationship node may also be associated with several "attributes," for example, the "mountain tunnel" node might be associated with "recommended feature extraction network: attention mechanism enhanced," "historical average detection accuracy: 0.85," and "easily confused targets: shadows and cracks." When entering a new task area, the extracted causal variables and shallow target features are first used to perform rapid retrieval and graph matching within the knowledge graph to find the scene subgraph that best matches the current context.

[0031] Furthermore, the successfully matched scene subgraphs become blueprints guiding the model's rapid adaptation. The system recalls the pre-trained perceptual model (or model component) most strongly associated with the subgraph from the model library. Unlike traditional transfer learning that directly fine-tunes the entire model, this application employs a graph-guided modular transfer strategy. The knowledge graph indicates the parts of the model that should be adjusted most for the key challenges of the current scene (e.g., shallow convolutional kernels in a feature pyramid network), as well as the stable parts that can be inherited (e.g., deep feature extractors for the texture of "concrete surfaces"). Subsequently, using only a small amount of labeled data quickly collected in the new scene, the specified modules are fine-tuned, while other modules are frozen. This method greatly reduces the amount of data and computational overhead required to adapt to new scenes, achieving on-demand evolution. Finally, the effect of this transfer adaptation (e.g., the degree of accuracy improvement) serves as a feedback signal to enhance or correct the weights of corresponding nodes and relationships in the knowledge graph, thereby completing the autonomous iteration of the knowledge graph itself and forming a scene understanding and transfer system that becomes smarter with use.

[0032] S4: Topology-Adaptive Perceptual Inference Process of Dynamic Graph Neural Networks

[0033] In the core stage of real-time perception inference, this embodiment abandons the traditional static neural network architecture and introduces a graph neural network that can dynamically adjust its internal connection topology according to real-time causal variables. In a preferred embodiment, before each inference, the multimodal perception data of the current frame (such as superpixels of an image, local clustering of point clouds) is first constructed into an initial graph structure, in which each data unit is a node, and the node features are its multimodal fusion feature vectors.

[0034] Specifically, the key innovation lies in the way the graph "edges" are constructed. Instead of using fixed adjacency relationships (such as spatial K-nearest neighbors), this method uses the current environmental state (such as a high "visibility attenuation coefficient" or a large "turbulence intensity index") as a control signal, inputting it into a lightweight edge prediction network. This network dynamically calculates a connection strength score for each pair of nodes. For example, in good visibility conditions, it tends to establish connections between spatially distant but semantically similar nodes to capture a wide range of contextual information; while in dense fog or heavy rain, it emphasizes strong connections between locally close nodes to aggregate robust local features to resist noise and weakens long-range connections to avoid misleading information. This is equivalent to equipping the perception network with an attention lens with adjustable focal length.

[0035] Furthermore, this embodiment also designs a state-aware message gating mechanism during graph convolution message passing. When information is passed from one node to its neighbor, the amount of information transmitted depends not only on the similarity of node features but also on the modulation of real-time causal variables (such as current wind speed). For example, when there are severe fluctuations in wind speed indicators, the gating mechanism reduces the weight of information transmitted based on apparent motion features while increasing the reliability of information transmitted based on geometric structure features. The entire dynamic graph neural network is connected to the causal variable input in an end-to-end manner, and its dynamic topology and gating parameters can be pre-trained through meta-learning, enabling the network to learn how to optimally reorganize itself according to environmental signals. Finally, after several layers of such dynamic graph convolution operations, the updated node features are decoded into the final target detection box or segmentation mask. This design makes the perception model no longer a passive recipient of the environment but an intelligent agent that can actively adjust its thinking (computation) mode based on its understanding of the environment, thus improving robustness in variable low-altitude environments.

[0036] S5: Closed-loop feedback-driven autonomous evolutionary decision-making and model iteration process

[0037] According to a further improvement of this embodiment, this embodiment is a continuously running decision-making and execution loop. The starting point of the loop is the collection of feedback data from multiple sources: including but not limited to reports from manual spot checks and verifications of the system's perception results, verification data from other high-precision detection equipment (such as ground-based precision instruments), and comparisons of repeated observations of the same target before and after task execution. These feedback data, together with the system's own original perception results and the corresponding causal variable context, are organized into a structured performance-context record.

[0038] In a preferred embodiment, the feedback logs are first analyzed in depth. Using a dynamic cause-effect graph, the root cause of the performance fluctuation is analyzed to determine whether it stems from environmental disturbances (such as newly emerging extreme weather combinations), scenario shifts (such as encountering facility types not yet fully learned), or inherent model flaws (such as consistently poor performance in recurring scenarios). This diagnostic conclusion directly determines the branch of the evolutionary strategy.

[0039] Specifically, if the diagnosis indicates environmental interference as the primary cause, causal adversarial sample generation will be initiated. Based on the identified key causal paths, training samples with corresponding interference characteristics (such as specific patterns of motion blur and fog concentration) will be synthesized in batches within the digital twin environment. The model will then undergo targeted incremental training to enhance its immunity to such interference. If the diagnosis indicates scene migration as the primary cause, the process will proceed to step S3, initiating knowledge graph querying and rapid model migration adaptation to incorporate knowledge of the new scene into the system. If the diagnosis indicates an inherent model defect, a more in-depth model architecture search may be triggered, exploring the potential for improvement of a specific submodule (such as an edge prediction network) within the current dynamic graph neural network under limited constraints.

[0040] Furthermore, regardless of the evolutionary strategy employed, the resulting new model version will not immediately replace the online version. According to one aspect of this application, an evolutionary sandbox is established. New models must pass statistical significance tests and demonstrate that their performance metrics (e.g., mAP, IoU) and efficiency metrics (e.g., latency) are comprehensively superior to or at least not inferior to the original version in a complete test set containing historical difficult cases and synthetic extreme scenarios before being approved for deployment. Simultaneously, a complete log of this evolution, including the triggering reasons, diagnostic results, adopted strategies, data used, and proof of performance improvement, is meticulously recorded and linked to corresponding nodes in the knowledge graph and causal graph. This not only ensures the traceability and auditability of the evolutionary process but also allows the system's experience to be continuously accumulated and structured, building stronger prior knowledge to address future challenges. This truly achieves the leap from perception-evolution to understanding-creation, forming an intelligent perceptual organism with continuous learning and self-improvement capabilities.

[0041] Example 2

[0042] This embodiment, based on Embodiment 1, describes in detail the data processing flow of the integrated low-altitude full-domain intelligent perception method for transportation based on autonomous evolution, specifically including:

[0043] S1: Multimodal Data Acquisition and Structured Extraction of Causal Variables

[0044] S1-1: Acquire data from multiple sensor sources

[0045] According to one aspect of this application, the system's perception begins with a multi-source data acquisition phase that is strictly spatiotemporally synchronized. In a preferred embodiment, a sensor array mounted on a UAV or fixed station is synchronously triggered based on a unified high-precision timing signal (such as a PPS pulse). Specifically, a visible light camera acquires RGB image sequences with a resolution of at least 4K and an adjustable frame rate, and records the attitude jitter data of each frame through a built-in IMU; an infrared thermal imager coaxially mounted with the camera synchronously captures long-wave infrared spectral data, with a thermal sensitivity better than 50mK, and its spatial resolution is aligned with the visible light image at the pixel level through calibration; a solid-state lidar acquires a three-dimensional point cloud at a rate of at least 200,000 points per second, with each point containing (X,Y,Z) coordinates, reflection intensity, and timestamp information; simultaneously, an integrated micro-weather station acquires physical parameters such as wind speed (range 0-30m / s, accuracy ±0.3m / s), wind direction, ambient temperature and humidity, air pressure, and light intensity (0-200klux) at a frequency of 10Hz. All raw data streams are embedded with a unified spatiotemporal index (WGS-84 coordinates and UTC time) to form raw multimodal data packets.

[0046] S1-2: Extracting environmental variables

[0047] Furthermore, the raw meteorological data needs to be cleaned and feature-engineered to transform it into interpretable environmental variables. In this embodiment, wavelet denoising is first performed on the wind speed time series data, and its standard deviation within a time window (e.g., 1 second) is calculated, mapping it to a turbulence intensity index between 0 and 1. Simultaneously, based on the current temperature, humidity, and air pressure, the atmospheric extinction coefficient is calculated using an improved Mie scattering model, thereby deriving the theoretical visibility distance. In addition, an effective illuminance level is generated by combining illuminance, solar altitude angle (calculated based on time and geographical location), and cloud cover (preliminarily estimated through image sky region segmentation). All these derived variables, along with the original sensor readings, constitute a multidimensional environmental variable vector V. env = [Turbulence Index, Visibility Distance, Effective Illuminance, Humidity, ...], this vector will serve as a key causal factor affecting perceived quality.

[0048] S1-3: Extract target features

[0049] Specifically, this embodiment performs preliminary shallow feature analysis on the perceived data to extract the essential attributes of the target, rather than for final identification. For visible light images, a lightweight convolutional neural network (such as the shallow part of MobileNetV3) is used to extract multi-scale feature maps, and the color histogram, LBP texture features, and SIFT descriptor statistics are calculated for candidate regions to form the apparent feature vector F. appFor infrared images, temperature inversion based on radiometric calibration is first performed to obtain the absolute temperature field. Then, the highest, lowest, and average temperatures of the target area, as well as the magnitude of the temperature gradient, are calculated to form the thermal feature vector F. thermal For LiDAR point clouds, after ground segmentation and clustering, the 3D bounding box size, point density, normal vector distribution entropy, and mean reflection intensity are calculated for each point cloud cluster to form a structural feature vector F. struct These eigenvectors collectively describe the physical and apparent nature of the target.

[0050] S1-4: Perform sensor status extraction

[0051] Further improvements to this embodiment involve real-time monitoring of the sensor's health status and data quality, which are crucial causal variables affecting sensing reliability. For image data, a uniform grayscale region is selected in each frame to calculate its noise power spectrum, and edge sharpness is evaluated (through Sobel operator response amplitude statistics) to obtain an image sharpness score and noise level index. For LiDAR point clouds, the ratio of effective echo points to total emission points (echo rate) is calculated, and the uniformity of the point cloud's spatial distribution is analyzed (measured by the variance of the number of points within voxels) to obtain a point cloud integrity score. These metrics are summarized into the sensor state vector V. sensor = [Clarity score, noise figure, echo rate, integrity score, ...], used to subsequently determine whether the perception anomaly is due to environmental interference or sensor performance degradation.

[0052] S1-5: Generate a set of causal variables

[0053] Finally, all generated variables are time-aligned, standardized, and encapsulated. Preferably, a causal variable database is established, with each record containing: timestamp, spatial location, and environmental variable vector V. env The set of feature vectors extracted from each target {F app , F thermal , F struct Sensor state vector V sensor These data are published via message queues, becoming sensory signals for the entire autonomous evolutionary system. For example, a record might be characterized as: "T=12:00:00.123, Pos=(X,Y,Z), V..." env =[0.85 turbulence, 1200m visibility, illuminance level 5], Target 1 feature={...}, V sensor =[0.92 resolution, low noise, 98% echo rate]". This structured set is the data foundation for all subsequent causal analysis and autonomous decision-making.

[0054] S2: Dynamic Cause-Effect Graph Construction and Error Source Tracing Process Based on Intervention Learning

[0055] S2-1: Construct the initial cause-effect graph structure

[0056] In this embodiment, a prior causal graph based on offline learning from historical big data is loaded during initialization. Specifically, this graph is a Bayesian network whose nodes represent all types of causal variables (such as turbulence index, line-of-sight distance, target apparent motion consistency, image sharpness, etc.) and key perception performance metrics (such as detection confidence and localization error). Directed edges represent conditional probabilistic dependencies between variables. The initial structure is constructed using a combination of constraint-based PC algorithm and score-based greedy search, learned from thousands of historical task logs. Each edge is accompanied by a conditional probability table (CPT) and mutual information strength values. This prior graph provides a powerful starting point for online real-time causal inference.

[0057] S2-2: Conduct causal intervention experiments

[0058] When online perception encounters anomalies or new scenarios, intervention learning is initiated. Specifically, the current causal variable scenario C is first reproduced in the digital twin environment. Then, a variable suspected of being the "cause" (e.g., "visible distance") is selected, and a virtual "do-operation" is performed. In one optional implementation, a calibrated physical rendering engine is invoked to intervene and change the visible distance from its current value (e.g., 1200m) to a series of other values ​​(e.g., 200m, 500m, 800m), generating simulated images and point cloud data for the same standard 3D scene under the corresponding conditions. This process does not rely on additional real data acquisition but rather exhaustively studies the effects of causal variables in the simulation.

[0059] S2-3: Calculate the strength of causal effect

[0060] Furthermore, the generated simulated data batches are input into the current perception model for inference. The curves showing the change in perception performance indicators (such as mAP) with the intervention value of viewing distance are recorded. According to one aspect of this application, the average causal effect is used as a quantification indicator: ACE = E[performance|do(viewing distance=low)] - E[performance|do(viewing distance=high)]. This method can clearly distinguish whether the relationship between viewing distance and performance is a genuine causal relationship or a spurious correlation caused only by other common factors (such as "humidity"). The calculated ACE values ​​and their confidence intervals are used to update or confirm corresponding edges in the causal graph.

[0061] S2-4: Generate a causal error mapping table

[0062] Specifically, the system maintains a dynamic causal-error mapping table. Each row of this table records a specific error pattern (e.g., "a surge in false negatives for small targets at low visual distances") and is associated with the causal path leading to that pattern (e.g., "reduced visual distance -> decreased image contrast -> weakened features of small targets -> false negatives"). Furthermore, the table records the number of times the causal path has been validated, the average effect strength, and the most recently triggered evolutionary strategy. This table is a key index connecting problem diagnosis and solutions, enabling the system to address the root cause.

[0063] S2-5: Update the weights of the cause-effect graph

[0064] According to a further improvement in this embodiment, the causal graph is not a static model, but an online learning entity. Each time an intervention experiment is completed or a causal path is confirmed from real-world feedback, the conditional probability table (CPT) of the relevant nodes is updated using incremental Bayesian learning. The weights of the edges (such as mutual information values) are also dynamically adjusted based on the strength of new evidence. Edges that are repeatedly falsified have their weights decayed until they are removed. This process allows the causal graph to continuously evolve, increasingly accurately reflecting the true causal structure of the perception system in its current deployment environment, providing a reliable theoretical model for precise autonomous evolution.

[0065] According to another aspect of this application, calculating the strength of a causal effect specifically includes:

[0066] S2-3-1: Defining the Framework for Causal Effect Calculation

[0067] According to one aspect of this application, a causal effect calculation framework based on a Potential Outcome Model is first established. Specifically, each virtual intervention experiment is treated as a controlled randomized trial, where the intervention variable T (e.g., "visual distance") is the treatment variable, and the perceived performance index Y (e.g., mAP@0.5) is the outcome variable. The potential outcome Y(t) is defined as the performance value at intervention T=t. The average causal effect (ACE) is defined as: ACE(t1,t2)=E[Y(t1)]-E[Y(t2)], where t1 and t2 represent two different intervention levels. To estimate ACE from the observational data, a simulation-based counterfactual prediction method is used, assuming that in the digital twin environment, all variables except the intervention variable are effectively controlled and conditionally negligible.

[0068] S2-3-2: Implementing multi-level interventions and data simulation

[0069] Furthermore, for the selected intervention variable T, k discrete levels are uniformly selected within its reasonable range (for example, for "visual distance", t1=200m, t2=500m, t3=800m, t4=1100m, t5=1400m are selected). For each intervention level t i Execute do(T=t) in a digital twin environment i The system performs a physics-based rendering engine (such as Unreal Engine combined with a meteorological optical model) to generate N sets (e.g., N=200) of simulated multimodal perception data. Each set of data contains visible light images, infrared images, and LiDAR point clouds of the same standard 3D scene (including typical traffic facilities and targets) at different intervention levels. All simulation data are accompanied by real-world annotations, and all scene parameters except for T are kept consistent to minimize the causal effects of intervention variables.

[0070] S2-3-3: Collect performance observations under intervention

[0071] Specifically, all generated simulation data are batch-input into the current perception model to be evaluated (i.e., the online-deployed model version) for inference. For the nth set of data at each intervention level t, the performance observation Y of the model output is recorded. i,n Examples include mAP@0.5 for object detection and mIoU for instance segmentation. Subsequently, the average performance Y at each intervention level is calculated. ˉ i =(1 / N)∑ n=1 N Y i,n And performance variance σ i 2 Meanwhile, to assess the stability of the estimate, the system uses a bootstrap method to repeatedly sample and calculate Y. ˉ i The 95% confidence interval.

[0072] S2-3-4: Calculate and verify causal effects

[0073] Further improvements to this embodiment employ statistical methods based on analysis of variance (ANOVA) and post-hoc testing to quantify and verify causal effects. First, a one-way ANOVA is performed with the intervention level as a factor to test whether there are significant differences in average performance across different intervention levels (significance level α = 0.05). If the ANOVA results are significant, multiple comparisons (such as the Tukey HSD test) are then performed to calculate the t-values ​​for any two intervention levels. i With t j The average causal effect between ACE(t)i ,t j )=Y ˉ i -Y ˉ j The system will then calculate the confidence interval for each effect. A causal effect is considered significant only if |ACE| is greater than a preset practical significance threshold (e.g., mAP difference exceeds 3%) and is statistically significant (p-value < 0.05). The system will generate a causal effect matrix, recording all significant effects and their magnitudes.

[0074] S2-3-5: Storing and Updating Knowledge of Causality

[0075] Finally, all calculated causal effect estimates, confidence intervals, statistical significance indicators, and corresponding experimental metadata (such as intervention variables, scenario types, model versions, and simulation data volumes) are structured and stored in the Causal Effect Knowledge Base (CEKB). In a preferred embodiment, CEKB is organized using a graph database, where nodes represent variables, directed edges represent validated causal effects, and edge attributes include effect size, direction, confidence level, number of validations, and last update time. The system periodically checks the consistency of new and old experimental results: if significant conflicts arise in effect estimates for the same intervention (e.g., opposite effect directions and both passing statistical tests), a more refined simulation experiment (e.g., increasing the number of intervention levels or increasing the sample size N) or expert review is triggered to ensure the robustness and reliability of causal knowledge.

[0076] S3: Knowledge Graph-Driven Scene Adaptation and Rapid Model Transfer Process

[0077] S3-1: Constructing a Scene Knowledge Graph

[0078] In a preferred embodiment, an ontology-driven Low-Altitude Traffic Scene Knowledge Graph (LSKG) is constructed. Specifically, a core ontology is first defined, including "scene classes" (e.g., urban elevated roads, mountain tunnels), "facility classes" (e.g., cable-stayed bridges, guardrails), "anomaly classes" (e.g., cracks, corrosion), "environmental state classes" (e.g., strong sidelight, freezing), and "model component classes" (e.g., feature extraction modules, attention heads). Subsequently, instance data is automatically populated from historical project reports, academic literature, and labeled data using information extraction techniques (e.g., BERT-based relation triple extraction), forming triples such as (mountain tunnel, often accompanied by strong dark light) and (strong dark light, recommended use, high dynamic range feature fusion module). The graph is stored in graph databases such as Neo4j, supporting efficient multi-hop queries and inference.

[0079] S3-2: Perform scene similarity retrieval

[0080] When entering a new task region, the extracted current scene features (such as main facility types and dominant environmental variables) are transformed into a query subgraph Q. Specifically, graph embedding techniques (such as TransE) are used to map entities and relations to a low-dimensional vector space. Then, the similarity between the query subgraph Q and the embedding vectors of all candidate scene subgraphs in the knowledge graph is calculated. Simultaneously, a graph structure similarity algorithm (such as the graph kernel method) is used to comprehensively select the top K most similar reference scenes. This process can find historical experiences that can be referenced for new scenes at both semantic and structural levels.

[0081] S3-3: Transferring Pre-trained Model Parameters

[0082] Furthermore, each "scene class" or "environment state class" node in the knowledge graph is associated with one or more validated, high-performing model configuration snapshots. According to one aspect of this application, this association is not a simple pointer to a model file, but a structured skill set indicating which parts of the model (e.g., the third-stage convolutional layer of ResNet, layers 2, 4, and 6 of the Transformer encoder) are critical and transferable for that scene, along with the pre-trained parameter weights for these parts. Based on the retrieval results, the system extracts these key module parameters from the most matching scene node to initialize the model for the new scene, achieving efficient knowledge inheritance.

[0083] S3-4: Perform fine-tuning with a small number of samples

[0084] Specifically, a graph-guided fine-tuning process is initiated using a small number of samples (e.g., 50-100 images) quickly collected and labeled in the new scene. The knowledge graph not only provides initialization parameters but also guides the design of the loss function through associated confounding terms and key feature information. For example, if the graph indicates that "water accumulation and cracks" are easily confused in the current scene, the system will add discriminative loss terms for these two categories to the loss function. The fine-tuning process employs higher-order optimization methods (such as the MAML approach) to enable the model to quickly adapt to the new scene with minimal gradient update steps, while avoiding catastrophic forgetting of fundamental general knowledge.

[0085] S3-5: Update the knowledge graph

[0086] Finally, the entire migration and adaptation process and results will be fed back to enhance the knowledge graph. The system automatically creates entity nodes for the new task scenario (if they do not exist) and connects them with similar scenario nodes found through a "similar to" relationship, with the edge weights determined by the performance gain after migration. Simultaneously, the new model configuration generated by this fine-tuning, as well as the newly discovered scenario-model component relationships, will be added to the graph as new triples. For example, (new scenario X, new adaptation scheme, dynamic attention mechanism Y). This makes LSKG a continuously growing and evolving collective intelligence, with its knowledge coverage and accuracy continuously improving as the system performs more tasks.

[0087] According to another aspect of this application, fine-tuning with a small sample size specifically includes:

[0088] S3-4-1: Constructing a graph-guided fine-tuning task

[0089] According to one aspect of this application, the design of the few-shot fine-tuning task is directly driven by scene challenge pattern nodes and key feature information retrieved from the knowledge graph (LSKG). Specifically, the system parses the LSKG, extracts several challenge descriptions most relevant to the current new scene (e.g., "texture features weakened in strong low light conditions," "high confusion between small targets and background"), and transforms these challenges into specific auxiliary learning tasks. For example, for texture feature weakening, the system adds a self-supervised texture reconstruction task, requiring the model to reconstruct texture feature maps under normal lighting from low-light images; for target-background confusion, a foreground-background discrimination task is added. These auxiliary tasks share the backbone network with the main perception task (detection or segmentation), forming a multi-task learning framework that forces the model to quickly focus on the core difficulties of the new scene.

[0090] S3-4-2: Preparing a Small Sample Incremental Dataset

[0091] Furthermore, the system utilizes a small amount of labeled data collected on-site in the new scene. new (Typically 50-100 samples), and intelligent data augmentation is performed by combining LSKG information. Specifically, the system selects a targeted method from a pre-set augmentation strategy library based on typical environmental interferences and easily confused patterns indicated by the LSKG. For example, if the LSKG indicates that water reflection is a common interference in the scene, then for D... new The system applies simulated water reflection optical effects to the image. Simultaneously, it retrieves a dataset of historical scenes similar to the current scene from its model library. old However, only samples that are highly correlated with the key discriminant features in LSKG (e.g., samples containing similar material surfaces) are selected to form a curated support set D. support Finally, fine-tuning the training set Dtrain By D new Augmented D new and D support Together, the total number of samples is usually controlled within 500 to ensure the efficiency of fine-tuning.

[0092] S3-4-3: Model Initialization and Meta-Learning Optimization

[0093] In this embodiment, the fine-tuning process employs an optimization-based meta-learning algorithm, specifically an improved version of MAML (Model-Agnostic Meta-Learning). First, the model parameters are initialized using the transferred parameters θ0. Then, D... train Divided into multiple smaller tasks {T i Each task contains a small number of samples (e.g., 5-way 5-shot). The inner loop of meta-learning is implemented in each task T. i Calculate the loss L Ti (f θ Task-specific parameters are obtained through one or more steps of gradient descent. The outer loop then aggregates all tasks and updates the initial parameters. This process enables the model to learn an internal representation that can quickly adapt to new tasks, greatly improving learning efficiency with few samples.

[0094] S3-4-4: Fine-tuning of execution constraints and knowledge distillation

[0095] To prevent catastrophic forgetting when adapting to new scenarios, multiple constraint terms are introduced into the fine-tuning loss function. According to one aspect of this application, a method combining Elastic Weight Consolidation (EWC) and Knowledge Distillation is employed. First, the system calculates the Fisher information matrix F of the old model parameters θ0 to evaluate the importance of each parameter to the old task. The EWC loss term is λ∑ i F i (θ i -θ 0,i ) 2 First, it penalizes significant changes in key parameters. Second, knowledge distillation loss requires that the output (logits) of the new model remain similar to the output of the old model on augmented data, preserving the generalization knowledge of the old model. The total loss function is: L total =Lnew +λ EWC L EWC +λ KD L KD , where the weight λ EWC and λ KD The model stability requirements recorded in LSKG are dynamically adjusted.

[0096] S3-4-5: Evaluation and Selection of the Best Fine-Tuning Model

[0097] Specifically, the fine-tuning process is performed on an independent validation set D. val Monitoring is performed on the hold-out samples from the new scene. This not only tracks the accuracy (e.g., mAP) of the main task on the new scene, but also monitors the model's performance on the old scene test set D. old-test The performance retention rate (i.e., the degree of forgetting) is measured. An early stopping strategy is adopted: training stops when the validation accuracy for a new scene no longer improves for P consecutive epochs (e.g., P=10), or when the performance of an old scene degrades by more than a threshold (e.g., 5%). After training, from multiple saved model snapshots, a comprehensive score S=0.6×Acc is used. new +0.3×Retain old The optimal model is selected by adding 0.1 × Efficiency, where Efficiency is the model's inference speed score. This model will then proceed to the final pre-deployment evaluation process.

[0098] S4: Topology Adaptive Perception Inference Process Based on Dynamic Graph Neural Networks

[0099] S4-1: Constructing Dynamic Graph Nodes

[0100] In this embodiment, the multimodal sensing data of each frame is transformed into an initial representation of a graph structure. Specifically, for visible light images, a superpixel segmentation algorithm (such as SLIC) is used to divide them into multiple visually coherent regions, with each superpixel region serving as a node. For LiDAR point clouds, Euclidean clustering is first performed, with each point cloud cluster serving as a node. The initial feature vector h of each node... i (0) This involves stitching together multimodal features of the corresponding region, including low-level semantic features such as the mean RGB color, mean infrared temperature, 3D size of point cloud clusters, and mean reflectance intensity. This step transforms unstructured perceptual data into a structured set of graph nodes.

[0101] S4-2: Dynamically generate graph edge connections

[0102] According to one aspect of this application, an environment-aware edge predictor (EA-Edge Predictor) is designed. This predictor uses the features of the current node and the real-time environment variable vector V as input. env The input is the same for all nodes. Specifically, for any two nodes i and j, the predictor computes a connection score s. ij = σ( MLP( concat( h i , h j V env ) ) ), where σ is the Sigmoid function. For example, when V env When the turbulence index is high, the predictor learns to assign higher connection scores to node pairs with consistent apparent motion characteristics to facilitate the aggregation of stable features; while at low line-of-sight distances, it tends to establish strong connections between spatially adjacent nodes. By setting a dynamic threshold, the system constructs a unique adjacency matrix A for each frame that best matches the current environment. t .

[0103] S4-3: Perform attention calculation for gating graphs

[0104] In the information transmission phase, an improved Graph Attention Network (GAT) is employed. Furthermore, the strength of information transmission depends not only on node feature similarity but also on the real-time sensor state V. sensor Gating. In the l-th layer graph convolution, the coefficient α of node i aggregating information from its neighbor j. ij The calculation is as follows: α ij = softmax j ( LeakyReLU( a T [W h i (l) || W h j (l) ] ) * g(V sensor ); where g(·) is a gating function that adjusts the scale of attention weights based on sensor conditions. For example, when the image noise index is high, the gating function reduces all attention weights based on apparent texture features, forcing the network to rely more on geometric or thermal features for judgment. This mechanism allows the network to trust a more reliable data source under the current conditions.

[0105] S4-4: Perform Adaptive Inference

[0106] After several layers (e.g., L layers) of the above dynamic graph convolution operation, each node obtains a deep feature h that integrates global context information and environmental state information. i (L)Ultimately, these node features are fed into a specific task head (detection head or segmentation head). For detection tasks, the system uses node features to predict bounding boxes and categories; for segmentation tasks, the node features are upsampled back to the pixel level to generate a mask. The entire inference process, from node construction and edge prediction to gating information transmission, is implemented in an end-to-end manner, but the dynamic parameters (such as the parameters of the edge predictor and gating function) are directly modulated by real-time causal variables.

[0107] S4-5: Perform real-time structural optimization

[0108] In one optional implementation, the system also possesses millisecond-level fine-tuning capabilities under extreme scenarios. When the environmental monitoring module detects that a certain extreme condition (such as sudden dense fog) persists for multiple frames and causes a significant performance degradation, a pre-loaded emergency response routine is triggered. This routine includes a highly lightweight subgraph structure optimizer pre-trained for such extreme conditions. This optimizer can quickly calibrate the edge prediction strategy or gating function parameters of the current dynamic graph neural network in a very short time (<50ms), enabling it to rapidly adapt to abruptly changing perception conditions without requiring a complete model retraining or switching, thus ensuring the system's response speed and robustness in extreme environments.

[0109] According to another aspect of this application, the gating graph attention calculation specifically includes:

[0110] S4-3-1: Design the gating function for environment awareness

[0111] According to one aspect of this application, the gating function is designed to control the real-time sensor state variable V. sensor (Essential parameters such as noise level and point cloud integrity) are transformed into modulation signals for the attention mechanism. Specifically, a learnable gating network is designed. ,in These are the network parameters. The network outputs a multidimensional gated vector g∈R. H×C Where H is the number of attention heads, and C is the number of feature channels per head. The gating vector is normalized to the [0,1] interval using the Sigmoid activation function, and each element g h,c This represents the modulation coefficient for the h-th head and the c-th channel. For example, when V sensor When the image noise is extremely high, the gating network learns to reduce the weights of attention heads that rely on high-frequency texture features.

[0112] S4-3-2: Calculate the raw attention score between node pairs.

[0113] Furthermore, for any two connected nodes i and j in the dynamic graph (the connection is determined by the dynamic edge predictor), their original attention scores are calculated. Specifically, the feature h of each node i... i ∈R d The query vector q is obtained after linear transformation. i =W Q h i Key vector k j =W K h j Sum vector v j =W V h j W Q W K W V ∈R d′×d This is a learnable weight matrix. For each attention head h, calculate the original attention score. ,in is the dimension of the key vector. This step captures the raw feature-based correlations between nodes.

[0114] S4-3-3: Applying gating mechanisms to adjust attention scores

[0115] According to a further improvement of this embodiment, the original attention score needs to be fused with the gating vector to achieve environment-adaptive attention modulation. Specifically, for the attention head h, the gating vector g is first fused with the gating vector g. (h) Broadcast to all node pairs, then calculate the gated attention score: e ~ ij (h) =e ij (h) ·σ(MLP adapt ([g (h) ;e ij (h) ])), where σ is the Sigmoid function, MLP adapt It is a small network used to learn how to make nonlinear adjustments based on the gating signal and the original score. Then, softmax normalization is performed on all neighbors j∈N(i) of node i to obtain the final attention weights: α ij (h) =exp(e ~ ij (h) ) / ∑ k∈N(i) exp(e ~ ik (h) This design makes the attention weights not only depend on the similarity of node features, but also directly modulated by the real-time sensor state.

[0116] S4-3-4: Aggregate neighbor information and update node characteristics

[0117] Next, the calculated attention weights are used to aggregate neighbor information. For each node i and each attention head h, the weighted aggregated neighbor value vector is: z i (h) =∑ j∈N(i) α ij (h) v j (h) Then, concatenate the outputs of all attention heads: z i =∥ h=1 H z i (h) The aggregated features h are obtained through a linear projection WO and an activation function (such as GELU). i =GELU(W O z i This step integrates the features of each node with information from its neighbors, and the strength and method of information aggregation are dynamically adjusted by environmental gating, thereby suppressing unreliable information flow under adverse perception conditions (such as partial sensor failure).

[0118] S4-3-5: Residual Connectivity and Layer Normalization

[0119] Finally, to maintain training stability and information flow, residual connections and layer normalization are employed. Specifically, the final output feature of node i after passing through the gated graph attention layer is: h i new =LayerNorm(h i +h ~ i This design ensures that even if the gating mechanism applies strong suppression on certain channels, nodes can still retain their original features as the basic representation. Through multiple layers of such dynamic gating graph attention layers, node features are gradually fused with contextual information from multi-hop neighbors and filtered by the environment state, thereby achieving a high degree of cooperative adaptation between the perceptual reasoning process and the real-time physical world state.

[0120] S5: Closed-loop feedback-driven autonomous evolutionary decision-making and model iteration process

[0121] S5-1: Collect error feedback data

[0122] The self-evolutionary cycle is driven by the collection of high-quality feedback data from multiple channels. Specifically, a feedback entry point is established to receive the following types of structured data: 1) Human expert review reports: submitted in JSON format, including corrections (additions, deletions, and modifications) to the system's automatic identification results and confidence level annotations; 2) High-precision equipment cross-validation data: for example, sub-millimeter precision point clouds obtained through a ground-based 3D laser scanner, serving as the "ground truth" for evaluating the LiDAR perception accuracy of UAVs; 3) Inter-task consistency verification data: multiple observations of the same target at different times and angles, forming a self-consistent verification set after high-precision registration; 4) System self-monitoring logs: including abnormal events such as memory overflow and computing power exceeding limits during model inference. All these feedback data are spatiotemporally aligned with the original perception data and the set of causal variables to form a complete evolutionary evidence chain.

[0123] S5-2: Perform causal error analysis

[0124] Furthermore, instead of treating errors in a general way, the system initiates a deep attribution analysis based on a causal graph. The current error case (e.g., a missed crack) along with a snapshot of the complete causal variables at the time of its occurrence is input into a dynamic causal graph. By running backward reasoning through a Bayesian network (i.e., reasoning from "effect" to "cause"), the posterior probability that each causal variable is a cause of the error is calculated. The system generates a detailed "Error Attribution Report," clearly indicating the most likely combination of root causes, such as: "This missed detection has an 85% probability of being caused by a combination of 'sudden drop in visibility distance' (value: 350m) and 'the target being in a high-contrast shadow area,' with the former contributing 70%." This provides precise targets for evolution.

[0125] S5-3: Triggering Evolutionary Strategy Selection

[0126] Based on the attribution report, the most suitable strategy is selected from a pre-defined evolutionary strategy library. Specifically, the strategy library is a classifier that combines rules and machine learning: 1) If the cause is attributed to a known environmental disturbance pattern, "causal adversarial example generation and incremental training" is triggered; 2) If the cause is attributed to a new scene or a new target type, "knowledge graph-guided rapid transfer" is triggered (jumping to the S3 process); 3) If the cause is attributed to an inherent structural defect of the model under specific complex conditions (such as insufficient multi-scale target handling), "Neural Architecture Search (NAS) sandbox experiment" may be triggered to explore the possibility of improving the dynamic graph building module (S4-2); 4) If the cause is attributed to sensor performance degradation or calibration drift, device calibration warning is triggered. This decision-making process ensures the accuracy and efficiency of the evolutionary measures.

[0127] S5-4: Update the knowledge graph and model library

[0128] All new knowledge generated by evolutionary actions must be solidified. According to a further improvement in this embodiment, the system maintains a versioned model library and an associated knowledge graph. When a new model (or module) passes rigorous testing in the "evolutionary sandbox," it will be stored in the model library with a new version number. Simultaneously, the error patterns, root causes, strategies employed, and performance gains of the new model targeted in this evolution will be summarized as new triples or attributes and updated in the knowledge graph (LSKG). For example, a new triple might be added: (small target under strong turbulence, recommended model, Model_v2.5), with the attributes {validation accuracy: 0.92, improvement over baseline: 15%}. This allows the system's experience to be continuously accumulated and shared.

[0129] S5-5: Output Autonomous Evolution Log

[0130] Finally, the entire evolutionary event is fully recorded, forming an auditable log. The log employs a standardized structure, including: evolutionary event ID, trigger time, associated original task ID, description of the error phenomenon, attribution analysis results (with screenshots of the cause-effect graph reasoning path), selected evolutionary strategy, amount of data generated / used during the evolution process, a detailed performance comparison table from the sandbox test, the final decision (deployment / rollback), and a unique identifier for the new model. These logs are not only used for system state backtracking and performance evaluation, but they are also valuable data assets for analyzing system evolution patterns and discovering deeper optimization directions, thereby driving the system towards higher levels of autonomous intelligence.

[0131] According to another aspect of this application, conducting causal error analysis specifically includes:

[0132] S5-2-1: Constructing a snapshot of causal variables for error cases

[0133] According to one aspect of this application, when a specific case of perceived error is captured through feedback channels (such as manual review or device cross-validation), a complete snapshot of causal variables is first constructed. Specifically, based on the timestamp t of the error occurrence... e and spatial location loc e The system retrieves all causal variable values ​​corresponding to that location at that time from the generated causal variable database. This includes: the environmental variable vector V. env (t e ,loc e Sensor state vector V sensor (t e ), and the shallow feature vectors {F} of the relevant targets. app ,F thermal ,F structSimultaneously, detailed error information is recorded, such as error type (missed detection, false detection, classification error, localization bias), true target category, model predicted category and confidence level, and error severity score. All this information is encapsulated into a structured error case object (EE), which serves as input for causal analysis.

[0134] S5-2-2: Using Cause-and-Effect Graphs for Reverse Reasoning

[0135] Furthermore, the "error type" and "error severity" from error case E are input as evidence into the dynamic causal graph to perform reverse reasoning. Specifically, an exact reasoning algorithm (such as the joint tree algorithm) is used to calculate the posterior probability distribution P(X) of all other causal variable nodes given the evidence. i |Evidence). The focus is on analyzing those posterior probabilities relative to their prior probabilities P(X). i ) Variable X that has significantly increased i These variables are likely suspect factors causing the error; generate a list L of suspect factors sorted by the magnitude of the posterior probability increase. suspec t=[(X1,ΔP1),(X2,ΔP2),…].

[0136] S5-2-3: Perform a counterfactual query to verify the root cause.

[0137] To pinpoint the root cause from the suspected factors, a series of counterfactual queries are performed. According to a further improvement of this embodiment, the counterfactual query takes the form: "If the value of variable X at that time was x..." ′ (Instead of the actual observed value x), will the error E still occur? The system uses a structural equation model (SEM) with a causal graph for intervention calculations. For example, for a missed detection case, if "visible distance = 200m" is in the list of suspected factors, then the query is: P(missed detection | do(visible distance = 1000m), other variables are actual observed values). By calculating the ratio of the counterfactual probability to the actual probability (probability ratio), the contribution of that variable to the error can be quantified. The system will perform such queries for the top K suspected factors (e.g., the top 3) and calculate their respective contribution scores.

[0138] S5-2-4: Generate a multi-dimensional error attribution report

[0139] Specifically, by combining the results of reverse reasoning and counterfactual queries, a structured "Error Attribution Analysis Report" is automatically generated. This report includes the following core components: 1) Case Summary: Time, location, and error description; 2) Root Cause Analysis Results: Listing the most likely combination of root cause variables, with each variable accompanied by a contribution score (percentage) and confidence level (based on posterior probability); 3) Causal Path Visualization: Displaying the causal chain from the root cause variable to the error node in a highlighted path diagram, clearly illustrating the logical process of error generation; 4) Contextual Comparison: Showing the actual values ​​of key causal variables at the time of the error and their typical value range in normal successful cases, highlighting the anomaly; 5) Action Recommendations: Recommending specific evolutionary strategies based on the root cause type (e.g., if the root cause is a "novel occlusion pattern," then it is recommended to initiate knowledge graph updates and sample collection). The report is output in both HTML and JSON formats for easy manual review and automated system processing.

[0140] S5-2-5: Update the error pattern library and cause-effect graph

[0141] Finally, all the results of this error analysis will be used to update the system's long-term memory. First, error case E and its complete attribution report will be added to a versioned historical error pattern library. This library supports similarity retrieval based on causal features, allowing for rapid matching of historical patterns when a new error occurs, thus accelerating diagnosis. Second, new causal relationships validated in this analysis (e.g., confirming that "a specific type of freezing" leads to "infrared thermal signature failure") will be used to update the dynamic causal graph. If new variables or relationships are involved, a causal graph structure learning process will be initiated to incrementally incorporate new knowledge. Through this continuous, evidence-based learning, the causal graph will increasingly accurately and completely characterize the complex causal relationships in the low-altitude traffic perception system, providing increasingly reliable theoretical guidance for autonomous evolution.

[0142] This invention addresses two core pain points in low-altitude traffic perception: model failure due to environmental variability and repetitive training caused by scene differences. By introducing causal reasoning and dynamic knowledge graphs, it can accurately diagnose the root causes of performance bottlenecks and achieve parsable, reusable, and transferable knowledge across scenes. This transforms the traditional black-box approach into a white-box evolution, improving the system's interpretability and scene generalization capabilities.

[0143] In terms of engineering practicality, this invention reduces the operation and maintenance costs and data dependence throughout the system's entire lifecycle. Through an environment-adaptive dynamic graph neural network and a few-sample meta-learning mechanism, the system can quickly adapt with only a small amount of new scene data. At the same time, it can autonomously adjust its inference strategy based on real-time sensing status, maintaining robust performance even under harsh conditions, thus freeing the system from massive annotation and manual tuning.

[0144] Ultimately, this invention constructs a complete autonomous closed loop of "perception-diagnosis-evolution," driving the evolution of low-altitude intelligent perception systems from static tools to continuously learning intelligent agents. This not only improves the reliability and efficiency of tasks such as inspection and monitoring in complex environments, but also lays a key technological foundation for building the next generation of intelligent operation and maintenance systems for transportation infrastructure with self-improvement capabilities.

[0145] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A comprehensive low-altitude, all-domain intelligent perception method for transportation based on autonomous evolution, characterized in that, include: Collect multi-source heterogeneous sensing data in low-altitude traffic scenarios, extract structured causal variables, and obtain a set of causal variables; Based on a set of causal variables, a causal graph model is constructed and dynamically updated. Attribution analysis of perceptual errors is performed through intervention learning and counterfactual reasoning to obtain a causal error attribution report. Based on the causal error attribution report, matching scenarios are retrieved from the knowledge graph, the parameters of the pre-trained model are transferred, and graph-guided few-sample fine-tuning is performed to obtain a fine-tuned model adapted to the new scenario. Based on a fine-tuned model and a real-time causal variable set, a perceptual graph structure is dynamically constructed. Adaptive inference is then performed using a gated graph attention mechanism for environment perception to generate target perception results. The dynamic construction of the perceptual graph structure includes: converting the multimodal perceptual data of each frame into a graph structure, where superpixel regions or point cloud clusters serve as nodes, and the initial features of each node are concatenated vectors of its multimodal low-level semantic features; dynamically calculating the connection scores between nodes using an edge predictor based on the feature concatenation vectors of real-time environmental variables to generate a dynamic adjacency matrix for the current frame; constructing a dynamic graph structure for adaptive inference based on the dynamic adjacency matrix; and performing adaptive inference using a gated graph attention mechanism for environment perception. The process includes: generating multi-dimensional gating vectors based on real-time sensor state vectors through a learnable gating network; calculating the original attention scores for connected node pairs in a dynamic graph structure based on the dot product of the query vector and the key vector; non-linearly fusing the original attention scores with the gating vectors to generate gating-adjusted attention scores; performing softmax normalization on the gating-adjusted attention scores to obtain the final attention weights, which are used to weight and aggregate the value vectors of neighboring nodes; concatenating and projecting the outputs of all attention heads and adding them to the residuals of the original node features, followed by layer normalization to obtain updated node features; and feeding the updated node features into the task head to generate target perception results. Based on the comparison between the target perception results and multi-source feedback data, causal-driven evolutionary decisions are triggered and executed to update the knowledge system and model library, thus completing the autonomous evolutionary closed loop.

2. The method according to claim 1, characterized in that, Extracting structured causal variables includes: Simultaneously acquire raw data streams from visible light cameras, infrared thermal imagers, lidar, and meteorological sensors to obtain spatiotemporally aligned multi-source data streams; Time series analysis and feature engineering are performed on meteorological sensor data from multi-source data streams to obtain environmental variable vectors; Shallow feature analysis is performed on visible light images, infrared images, and lidar point cloud data from multi-source data streams to obtain a set of target feature vectors, including appearance feature vectors, thermal feature vectors, and structural feature vectors. Quality assessment is performed on image data and point cloud data from multi-source data streams to obtain sensor state vectors; The environmental variable vector, the target feature vector set, and the sensor state vector are time-aligned and encapsulated to generate a causal variable set.

3. The method according to claim 2, characterized in that, Time series analysis and feature engineering of meteorological sensor data in multi-source data streams, including: Wavelet denoising was performed on the wind speed sequence in the meteorological sensor data, and its standard deviation was calculated to obtain the turbulence intensity index. Based on temperature, humidity and air pressure parameters from meteorological sensor data, the atmospheric extinction coefficient is calculated using an improved Mie scattering model to obtain the theoretical visibility distance; The effective illuminance level is calculated based on the parameters of light intensity, solar altitude angle and cloud coverage from meteorological sensor data. The turbulence intensity index, theoretical visibility distance, and effective illuminance level are standardized to form an environmental variable vector.

4. The method according to claim 1, characterized in that, Building and dynamically updating causal graph models includes: Load a prior causal graph learned from historical data. The causal graph adopts the form of a Bayesian network, and the nodes represent causal variables and perception performance indicators. Based on a set of causal variables, virtual intervention operations are performed on key causal variables in a digital twin environment to generate corresponding simulated perception data; Input the simulated perception data into the current perception model, calculate the average performance under different intervention levels, and obtain the performance observation values ​​for each intervention level; Based on performance observations, analysis of variance and post-hoc testing methods were used to verify the statistical significance of performance differences among different intervention levels and obtain the average causal effect. Based on the average causal effect, the structure and parameters of the causal graph are updated using the incremental Bayesian learning method to form an updated causal graph model.

5. The method according to claim 4, characterized in that, Performing virtual interventions on key causal variables includes: Reproduce the current causal variable context in a digital twin environment; For the selected intervention variable, multiple discrete levels are selected within its reasonable range of values; For each intervention level, the physical rendering engine is invoked to generate simulation images and point cloud data for the same standard 3D scene under the corresponding conditions, forming simulated perception data.

6. The method according to claim 1, characterized in that, Retrieving matching scenarios from a knowledge graph includes: Transform the current scene features into a query subgraph; In a knowledge graph, the most similar reference scene subgraph is retrieved and queried by graph embedding and graph structure similarity calculation. Based on the reference scene subgraph, obtain the associated pre-trained model configuration snapshot to obtain transferable model parameters.

7. The method according to claim 6, characterized in that, Spectrum-guided small-sample fine-tuning includes: Based on the scene challenge patterns retrieved from the knowledge graph, auxiliary learning tasks are constructed to form a multi-task learning framework with the main perception task; By utilizing a small amount of labeled data from new scenarios, combined with targeted augmented data and a support set carefully selected from historical data, a few-sample incremental training set is constructed. An optimization-based meta-learning algorithm is adopted, which initializes the model parameters with transferable parameters and performs inner and outer loop optimization on a small sample incremental training set to obtain the optimized model parameters. By introducing elastic weight consolidation loss and knowledge distillation loss into the fine-tuning loss function, the variation of important parameters is constrained and generalized knowledge is preserved, resulting in the final fine-tuned model.

8. The method according to claim 1, characterized in that, Causally driven evolutionary decisions that trigger and execute include: Collect multi-source structured feedback data from manual review, cross-validation with high-precision equipment, inter-task consistency verification, and system self-monitoring; The target perception results are compared with the feedback data to obtain error cases; Input the error cases and their corresponding causal variables into the causal graph model, perform reverse reasoning and counterfactual queries, and generate an error attribution report; Based on the error attribution report, select and trigger the corresponding evolutionary strategy from the preset strategy library; By executing an evolutionary strategy, updated models and knowledge are obtained and simultaneously updated to the versioned model library and knowledge graph, forming a new self-evolutionary closed loop.

Images