Autonomous driving perception fusion strategy preloading method and system based on environment prediction

Abstract

The application discloses an automatic driving perception fusion strategy preloading method and system based on environment prediction, which comprises the following steps: acquiring multi-source perception data, acquiring the source confidence of each perception data, performing source confidence weighted fusion on the multi-source perception data to obtain overall confidence, extracting a plurality of key environmental factors from the multi-source perception data, pre-processing and aggregating the plurality of key environmental factors to obtain an environment prediction vector, performing Top-K vector retrieval based on a strategy index library to obtain a candidate strategy set comprising a plurality of candidate strategies, obtaining the matching degree of each candidate strategy based on the overall confidence and the environment prediction vector, obtaining a decision through dynamic decision gating, and obtaining a final fusion strategy parameter based on the matching degree of each candidate strategy. The application can eliminate the strategy adaptation empty window period caused by environmental mutation, and realizes zero perception lag of strategy switching through feedforward preloading.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and specifically to a method and system for preloading autonomous driving perception fusion strategies based on environmental prediction. Background Technology

[0002] Currently, to improve the robustness of autonomous driving systems in different environments, mainstream solutions generally adopt dynamic adaptive fusion strategies. This strategy aims to adjust fusion behavior according to real-time environmental changes, and its implementation mainly follows two paths: 1. Dynamic weighting based on real-time data quality assessment. This method continuously calculates the real-time quality indicators of each sensor data stream during the fusion process (e.g., assessing blur or noise levels in camera images, and performing density or reflection intensity consistency analysis on LiDAR point clouds), and assigns dynamically changing fusion weights to each sensor based on these real-time assessment results. When the data quality of a sensor deteriorates due to environmental interference (such as raindrops adhering to the lens), its weight is automatically reduced. 2. Strategy switching based on preset scene rules. This method relies on the pre-definition of a limited number of typical environments (e.g., sunny days, rainy days, nighttime). By identifying the current scene category (usually based on simple sensor data threshold judgments or time and geographical location information), it calls the corresponding fixed set of fusion parameters from a preset strategy index library (e.g., under rainy weather rules, millimeter-wave radar is pre-set to have a higher weight).

[0003] However, the aforementioned existing dynamic fusion technologies share a common, fundamental limitation stemming from their design logic: they are essentially reactive or passive adaptive models. Whether relying on posterior evaluation of data quality or on the identification of a past scenario, the system's policy adjustment instructions are always triggered and generated only after the environmental change has occurred. This forces the system to sequentially undergo a complete delay chain: "perceiving environmental changes," "calculating evaluation / identification," and "deciding and executing adjustments." When encountering drastic or rapid changes in environmental conditions (e.g., suddenly entering a rainstorm area or tunnel from an open road), this delay chain creates a significant window of opportunity for perception policy adaptation. During this window, the system continues to use fusion strategies optimized for previous environments to process entirely new sensor data, easily leading to transient false detections, missed detections, or inaccurate trajectory tracking, posing a clear safety risk. The core flaw of existing technologies lies in their adaptive mechanisms' complete reliance on current or historical local vehicle information, lacking the ability to proactively configure the perception system using forward-looking environmental prediction information. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method and system for preloading autonomous driving perception fusion strategies based on environmental prediction, which eliminates the policy adaptation gap caused by sudden environmental changes and achieves zero perception lag in policy switching through feedforward preloading.

[0005] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0006] According to a first aspect of this application, a method for preloading autonomous driving perception fusion strategies based on environmental prediction is provided, comprising: Acquire multi-source sensing data; Obtain the source confidence score of each sensing data point, and perform source confidence score weighted fusion on the multi-source sensing data to obtain the overall confidence score; Several key environmental factors are extracted from the multi-source sensing data, and the key environmental factors are preprocessed and aggregated to obtain an environmental prediction vector. Top-K vector retrieval is performed based on the strategy index library to obtain a set of candidate strategies including several candidate strategies. Based on the overall confidence and the environment prediction vector, the matching degree of each candidate strategy is obtained. Decisions are obtained through dynamic decision gating, and the final fusion strategy parameters are obtained based on the matching degree of each candidate strategy.

[0007] In some embodiments of this application, based on the aforementioned scheme, during the source confidence weighted fusion of the multi-source sensing data, the reciprocal of the sensor measurement variance corresponding to each sensing parameter is used as the source confidence weight of the sensing parameter.

[0008] In some embodiments of this application, based on the foregoing scheme, the source confidence weight is further adjusted, specifically as follows: Based on the environmental state information and sensor state information corresponding to each sensing parameter, a weight adjustment function is constructed to adjust the source confidence weight of each sensing parameter. The calculation formula is as follows:

[0009] in, Indicates the first Weight adjustment function for each sensing parameter, Indicates the first Source confidence weights for each sensing parameter Indicates weather-related parameters. This indicates the parameter affecting visibility. Indicates the first The state influence parameter of the first sensing parameter is used to characterize the acquisition of the first sensing parameter. The degree of abnormality in the working state of the sensor for each sensing parameter. The source confidence weights are adjusted based on the state. No. The weight adjustment function for each sensing parameter is:

[0010] in, , , To obtain the first The influence coefficients corresponding to the sensor's sensing parameters are used to characterize the degree of influence of weather factors, visibility factors, and sensor condition factors on sensor reliability; and ; The source confidence weights after state adjustment are adjusted based on the prediction error to obtain the error-adjusted source confidence weights, which are then used as the final source confidence weights.

[0011] In some embodiments of this application, based on the foregoing scheme, the policy index library is defined as a set of policy entries:

[0012] in, This indicates the total number of policy entries in the policy index, the [number]th [entry]. Each strategy entry can be represented as:

[0013] in: For the first A unique identifier for each strategy entry; For the first Each strategy entry corresponds to an applicable environment reference vector, and the environment reference vector and the environment prediction vector... They have the same dimensions and semantics; For the first The environmental tolerance range parameter corresponding to each strategy entry is used to characterize the allowable deviation range of the strategy in each environmental dimension. The environmental tolerance range parameter is expressed in the form of a dimension-wise tolerance vector or covariance. For the first Each strategy entry corresponds to a fusion strategy parameter package, which consists of directly executable strategy parameters and supports interpolation or adjudication processing. For the first Each policy entry has an activation threshold parameter, which includes at least one of minimum confidence, minimum matching degree, and disabling conditions. For the first The version and effect statistics of each strategy item, wherein the version and effect statistics include at least one of success rate, average error, applicable operating design domain and update time.

[0014] In some embodiments of this application, based on the foregoing scheme, the step of performing Top-K vector retrieval based on the strategy index library to obtain a candidate strategy set including several candidate strategies includes: Upon receiving the environmental prediction vector Then, through the vector index Perform candidate strategy recall to obtain a candidate strategy set. :

[0015] in, This indicates the preset candidate recall number, used to control the number of candidates retrieved from the vector index. The number of candidate strategies initially retrieved; Represents the set of candidate policies, based on the environmental prediction vector. Retrieved from the policy index library Each strategy entry is as follows One candidate strategy, A set of candidate strategies, which is an aggregation of individual candidate strategies.

[0016] In some embodiments of this application, based on the foregoing scheme, obtaining the matching degree of each of the candidate strategies based on the overall confidence level and the environmental prediction vector includes: Obtain the prediction vectors of each candidate strategy and environment. Matching distance between :

[0017] in, Represents the environmental prediction vector In the Predicted values ​​across environmental dimensions; Indicates the first The environment reference vector of each candidate strategy In the Reference values ​​for each environmental dimension; Indicates the first The candidate strategies in the first Tolerance range parameters in each environmental dimension; This represents the dimension of the environment vector, i.e., the number of environment features involved in the matching calculation; To prevent tiny positive numbers with a denominator of zero; The matching distance is represented by an exponential mapping. Mapped to matching similarity The matching similarity score ranges from 0 to 1, and is calculated using the following formula:

[0018] in, To match the decay control parameters of the similarity mapping; The overall confidence score and the matching similarity score of each candidate strategy are weighted and fused to obtain the matching score of each candidate strategy.

[0019] In some embodiments of this application, based on the foregoing scheme, the step of obtaining decisions through dynamic decision gating and obtaining the final fusion strategy parameters based on the matching degree of each candidate strategy includes: Define a key matching quantity, which includes the candidate strategy with the highest matching degree in the candidate strategy set. The second highest candidate strategy Matching discrimination ; When satisfied This triggers a high-confidence mode, outputting the final fusion strategy parameters as the strategy parameter package corresponding to the candidate strategy with the highest matching degree. This is the confidence threshold. As a matching threshold, When matching the discrimination threshold, Indicates the overall confidence level; When satisfied For interpolable continuous parameters in the candidate strategies, interpolation weights are obtained, and the continuous parameters are fused based on these weights to obtain a first hybrid continuous parameter set. For non-interpolable continuous parameters in the candidate strategies, a weighted decision is applied to obtain a second hybrid continuous parameter set. The first and second hybrid continuous parameter sets are then aggregated to obtain the final fused strategy parameters. The threshold for triggering fuzzy synthesis is determined by the matching degree. The low confidence threshold; When satisfied This triggers the security meta-policy mode, outputting the final fusion policy parameters as a predefined set of conservative fusion policy parameters, where... Match thresholds for unknown environments.

[0020] In some embodiments of this application, based on the foregoing scheme, after obtaining the final fusion strategy parameters, the method further includes: At the preload trigger moment, feedforward preloading and smooth transition control are executed. The method for obtaining the preload trigger moment is as follows: The remaining time to reach the predicted environmental boundary is calculated using the following formula:

[0021] in, To determine the distance to the predicted environmental boundary, For the vehicle's speed, This is the preset minimum speed limit; Based on the remaining time to reach the predicted environment boundary The time consumed by the final fusion strategy parameters The preload trigger time is calculated using the following formula:

[0022] in, This is a preset safety margin time.

[0023] In some embodiments of this application, based on the foregoing scheme, after obtaining the final fusion strategy parameters, the method further includes: A double-buffering mechanism is used to perform policy switching. Buffer-A is used to maintain the continuous output of the last acquired final fusion policy parameters, denoted as... Buffer-B: Used to carry the continuous output of the currently acquired final fusion strategy parameters, denoted as... ; Within the switching time window, the previously acquired final fusion strategy parameters and the currently acquired final fusion strategy parameter outputs are weighted and fused. The calculation formula is as follows:

[0024] in, This represents the time weighting factor during the handover process; The time weighting factor It can be represented as:

[0025] in, Indicates the start time of the switch; Strategy switching window duration; Represents the amplitude limiting function; when hour, Execute Buffer-A; when hour, The value gradually increases from 0 to 1, indicating a smooth transition from Buffer-A to Buffer-B; when hour, Execute Buffer-B.

[0026] According to a second aspect of this application, an autonomous driving perception fusion strategy preloading system based on environment prediction is provided, comprising: The data acquisition module is used to acquire multi-source sensing data; The overall confidence level acquisition module is used to acquire the source confidence level of each sensing data, and to perform source confidence level weighted fusion on the multi-source sensing data to obtain the overall confidence level; An environmental prediction vector acquisition module is used to extract several key environmental factors from the multi-source sensing data, preprocess and aggregate the key environmental factors to obtain an environmental prediction vector. The matching degree acquisition module is used to perform Top-K vector retrieval based on the strategy index library to obtain a set of candidate strategies including several candidate strategies, and to obtain the matching degree of each candidate strategy based on the overall confidence and the environment prediction vector. The final fusion strategy parameter acquisition module is used to obtain decisions through dynamic decision gating and to obtain the final fusion strategy parameters based on the matching degree of each candidate strategy.

[0027] The beneficial effects of this application are as follows: (1) The autonomous driving perception fusion strategy preloading method and system based on environment prediction provided in this application achieves a zero-lag experience in perception strategy switching: through feedforward prediction and preloading mechanism, the strategy preparation time is completely masked during the driving process before the vehicle reaches the environmental change point. When the vehicle actually enters the new environment, the optimal strategy is already in a hot standby state and can be activated instantly, fundamentally eliminating the safety risks caused by the perception-evaluation-switching delay in the existing technology, and providing key safety redundancy for high-level autonomous driving.

[0028] (2) The autonomous driving perception fusion strategy preloading method and system based on environmental prediction provided in this application improves the robustness and intelligence of decision-making under uncertain conditions: it abandons the rigid decision-making logic of either / or. By introducing prediction confidence assessment and strategy synthesis mechanisms, it can gracefully handle the actual situation of incomplete information and ambiguous predictions, and make the most reasonable and safest compromise decision. This biomimetic flexible intelligence makes it exhibit a much better adaptability than traditional rules when facing the complex and ever-changing environment of the real world.

[0029] (3) The preloading method and system for autonomous driving perception fusion strategy based on environmental prediction provided in this application ensures the extreme smoothness and stability of perception output: the unique double buffer and smooth transition design is an important technical detail to ensure the driving comfort and control stability of high-level autonomous driving. It ensures that even if the underlying fusion algorithm is fundamentally switched, the environmental perception results transmitted to the planning and control module are continuous and gradual, avoiding sudden braking or steering jerking caused by abrupt changes in perception results, and improving the overall quality.

[0030] (4) The preloading method and system for autonomous driving perception fusion strategy based on environmental prediction provided in this application effectively solves the long tail problem of traditional systems due to scene solidification and knowledge staticity, laying a solid foundation for realizing full-domain autonomous driving.

[0031] (5) The preloading method and system for autonomous driving perception fusion strategy based on environmental prediction provided in this application offer a highly modular and engineering-friendly implementation path: the entire scheme adopts a loosely coupled modular design, and the modules communicate with each other through clearly defined structured data interfaces. The preloading, double buffering and other mechanisms fully consider the computational resource constraints and real-time requirements of the vehicle-mounted embedded system. This makes it relatively easy to integrate this embodiment with existing autonomous driving software stacks (such as architectures based on ROS or AUTOSAR), and it has good technical feasibility and industrial application prospects.

[0032] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and do not limit this application. Attached Figure Description

[0033] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and are intended to explain the invention, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart of a preloading method for autonomous driving perception fusion strategy based on environmental prediction according to the present invention. Figure 2 This is a schematic diagram of a preloading system for autonomous driving perception fusion strategy based on environmental prediction according to the present invention. Figure 3 This is a schematic diagram of an electronic device according to the present invention. Detailed Implementation

[0034] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0035] It should be understood that the terms "comprising" and other similar expressions in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, or apparatus that includes a series of steps or units and is not limited to the listed steps or units. Furthermore, "first" and "second" are used to distinguish different objects and are not intended to describe a specific order.

[0036] According to the first aspect of this application, Figure 1 As shown, this embodiment provides a method for preloading autonomous driving perception fusion strategies based on environmental prediction, including: Step S1: Acquire multi-source sensing data.

[0037] In some embodiments of this example, the multi-source perception data includes data from different sources during vehicle operation, mainly including high-precision map data, on-board sensor data, and V2X communication data.

[0038] In this embodiment, high-precision map data provides static environmental information, such as road topology, traffic signs, and traffic lights.

[0039] Vehicle sensor data: including real-time dynamic data from LiDAR, cameras, millimeter-wave radar, ultrasonic sensors, etc.

[0040] V2X communication data: Information exchanged between vehicles and other vehicles and infrastructure (such as streetlights, traffic signs, etc.), mainly including real-time road conditions, vehicle speed, traffic flow, etc.

[0041] In some embodiments of this example, preprocessing of the multi-source sensing data is also included. All multi-source sensing data first undergoes spatiotemporal alignment and preprocessing operations. The goal of this process is to ensure that multi-source sensing data from different sensing data sources can be fused within the same temporal and spatial reference frame. Specifically, this includes: Spatiotemporal alignment: Timestamp synchronization algorithms are used to ensure that sensing data from different sensors (such as LiDAR and cameras) or different sensing data sources (such as V2X and map sensing data) are aligned within the same time window. Timestamp synchronization algorithms ensure temporal consistency among sensing data sources by marking the sensing data streams.

[0042] Coordinate Transformation: A coordinate transformation algorithm is used to transform the data sensed by each sensor from its own coordinate system to the global reference coordinate system. This algorithm unifies the data into the world coordinate system, ensuring consistency among the sensed data.

[0043] After the sensing data is preprocessed, all sensing data will be denoised using Kalman filtering or mean filtering algorithms to remove noise and ensure the quality of the sensing data.

[0044] Step S2: Obtain the source confidence of each sensing data, and perform source confidence weighted fusion on the multi-source sensing data to obtain the overall confidence.

[0045] In some embodiments of this example, each sensed data is assigned a confidence value based on its working environment, data quality, and reliability. The confidence level is used to quantify the credibility of the sensed data in the current environment. The calculation of source confidence depends on the type of sensor acquiring the sensed data, its operating conditions, and a predefined performance model.

[0046] In some embodiments, the lidar signal attenuates significantly in complex environments (such as fog, heavy rain, etc.), resulting in lower confidence levels. The source confidence level of the lidar is calculated by adjusting for both measurement accuracy and environmental conditions. ,like:

[0047] in, and These are correction coefficients related to environmental conditions. The weather factor represents the impact of severe weather on the lidar, while the obstruction factor represents objects in the environment that may affect the lidar's line of sight.

[0048] Camera: Under good lighting conditions, the camera's signal quality is high, resulting in high confidence. However, in low-light or bright-light environments, the camera's performance degrades, leading to lower confidence. Image analysis (such as exposure time and dynamic range) is used to calculate the camera's source confidence. ,like:

[0049] Millimeter-wave radar: Millimeter-wave radar performs stably in complex weather conditions such as rain and fog, therefore its source confidence is relatively high. The source confidence of millimeter-wave radar is calculated by evaluating its reflected signal strength and interference noise. .

[0050] In some embodiments of this example, to facilitate fusion, the source confidence of all perceived data will be normalized to the range of [0, 1]. Specifically, the normalization process is performed using the following formula:

[0051] in, For the first Source confidence of individual sensing data and These represent the minimum and maximum source confidence scores for all perceived data, respectively.

[0052] In some embodiments of this example, after calculating the source confidence scores for each piece of sensing data, the source confidence scores of each piece of sensing data are fused to obtain a global confidence score. To improve the adaptability of the fusion results to complex environments, this embodiment adopts a weighted average fusion strategy and dynamically adjusts the fusion weights of each piece of sensing data by combining environmental state information and sensor operating status information for acquiring the sensing data.

[0053] Let the number of sensing data participating in the fusion be . , No. The normalized source confidence level corresponding to each perceptual data point is: Its source confidence weight is Then the global confidence level It can be represented as:

[0054]

[0055] in, Indicates the first The source confidence value of each sensed data after normalization processing. Indicates the first The source confidence weights of the perceived data. The determination is based on the measurement uncertainty of the sensed data. Specifically, the reciprocal of the sensor measurement variance corresponding to each sensed parameter is used as its source confidence weight, i.e.:

[0056] in, Indicates obtaining the first The measurement variance of a sensor with a given sensing parameter is used to characterize the uncertainty of the sensor's output. The smaller the measurement variance, the higher the measurement stability of the sensor, and the greater the source confidence weight assigned accordingly; conversely, the larger the measurement variance, the smaller the source confidence weight assigned accordingly.

[0057] In some embodiments of this example, considering that different environmental conditions and the sensor's own operating state can affect the reliability of the sensor output results, in some scenarios, relying solely on fixed source confidence weights for confidence fusion is insufficient to accurately reflect the actual effectiveness of each sensor in the current environment. Therefore, this embodiment further adaptively adjusts the source confidence weights of each sensing parameter so that the source confidence weights can be dynamically updated according to changes in the environment and sensor state. Specifically, for the first... For each sensing parameter, obtain the corresponding environmental state information and sensor state information, and construct a weight adjustment function for the sensor based on the environmental state information and sensor state information. The weight adjustment function is used to obtain the first... The weight correction coefficients for each sensor in the current scene are used to correct the source confidence weights, thus obtaining the adjusted weights. Specifically, the calculation formula is as follows:

[0058] in, Indicates weather-related parameters. This indicates the parameter affecting visibility. Indicates the parameters that affect the state.

[0059] Specifically, regarding the first For each sensor, the following parameters are selected as input variables for the weight adjustment function: Environmental status information: Weather impact parameters Used to characterize the impact of weather factors such as rainfall, snowfall, fog, haze, and dust on sensor performance; visibility influence parameter. : Used to characterize the visibility distance or air penetration conditions in the current scene.

[0060] Sensor status information: Status-affecting parameters Used to characterize the acquisition of the first The degree of abnormality in the working state of a sensor with a given sensing parameter.

[0061] Among them, state influence parameters It can be determined by at least one of the following: signal-to-noise ratio degradation, frame loss rate, echo anomaly rate, temperature anomaly, and data integrity anomaly.

[0062] To facilitate standardized calculations, the aforementioned weather impact parameters are first... Visibility Influencing Parameters and state influence parameters Normalization is performed to ensure that the numerical values ​​fall within a preset range. In one embodiment, the preset range is... The larger the value, the stronger the corresponding adverse effect. Based on the above parameters, the first... The weight adjustment function for each sensing parameter is established as follows:

[0063] in, , , For the first The influence coefficients corresponding to each sensor are used to characterize weather factors, visibility factors, and state influence parameters, respectively; and .

[0064] As can be seen from the above expression, when the impact of weather intensifies, visibility deteriorates, or the degree of abnormality in sensor operation increases, the function... The output value of the function decreases, thereby reducing the adjusted weight of the corresponding sensor; conversely, when environmental conditions are good and the sensor is working properly, the function... The output value is close to 1, and the source confidence weight of the corresponding sensor remains basically unchanged.

[0065] The influence coefficient , , The following method can be used to determine the following: under different weather conditions, different visibility conditions, and different sensor operating states, collect test sample data corresponding to each sensor, and statistically analyze the detection accuracy, false detection rate, false negative rate, confidence fluctuation, or recognition stability of each sensor under different conditions; then, based on the statistical results, fit the degree of influence of weather parameters, visibility parameters, and state influence parameters on sensor reliability, and then determine the influence coefficient.

[0066] Real-time weight updates: Weight adjustments not only rely on environmental and sensor status information but also incorporate real-time sensor feedback to optimize weights. The real-time feedback learning mechanism ensures continuous improvement of the source confidence weight adjustment strategy for sensing parameters during operation. The specific process is as follows: Feedback learning mechanism: After each perception decision, the performance of the sensor is evaluated based on the subsequent decision results (e.g., the accuracy of obstacle detection, the stability of vehicle control, etc.).

[0067] Decision feedback: During vehicle operation, the system continuously monitors the discrepancies between the data provided by sensors and the actual vehicle control feedback. For example, if the LiDAR fails to correctly identify obstacles in complex environments, the error will be analyzed, and the LiDAR's weight will be reduced.

[0068] Error Correction: The weights of the sensors are corrected using a weighted error feedback algorithm. In each feedback cycle, the weight of the corresponding sensor is adjusted based on the magnitude of the error. For example, when the error of the lidar is large, its weight is reduced; while when the millimeter-wave radar performs well, its weight is appropriately increased.

[0069] Real-time updated weight formula:

[0070] in, It is the learning rate, which represents the sensitivity of weight updates. It is the first The prediction error of each perception parameter is considered. If a large error is detected (such as an obstacle not being identified), the weight of that perception parameter will be negatively adjusted.

[0071] Online optimization and continuous adjustment: This mechanism continuously updates the sensor weights using online optimization algorithms (such as incremental learning or adaptive filtering). After each learning cycle, the updated weights are applied to the next round of the sensing task.

[0072] Step S3: Extract several key environmental factors from the multi-source sensing data, preprocess and aggregate the key environmental factors to obtain an environmental prediction vector.

[0073] In some embodiments of this example, after completing the alignment of multi-source sensing data and the source confidence weighted fusion, the next step is to extract key environmental factors from the multi-source sensing data. These key environmental factors are directly related to the physical performance of the sensors and the sensing task. Common key environmental factors include: LiDAR point cloud attenuation rate: This represents the degree of attenuation of the LiDAR signal under different environmental conditions, affecting its detection range and accuracy. The effectiveness and attenuation rate of the LiDAR point cloud are calculated using point cloud processing algorithms (such as Gaussian filtering). The calculation formula is as follows:

[0074] This value represents the degree of attenuation of the lidar, serving as a key environmental factor. .

[0075] Camera Dynamic Range Requirement Index: This index reflects the dynamic range requirements of a camera under different lighting conditions, affecting its ability to recognize high-contrast scenes. The dynamic range requirement of the camera for the current environment is calculated through image brightness analysis, using the following formula:

[0076] This value serves as a key environmental factor. .

[0077] Road sign recognition index: The ability of a camera to recognize road signs under different environmental conditions. The recognizability of road signs is evaluated using image recognition algorithms (such as convolutional neural networks), and the calculation formula is as follows:

[0078] This value serves as a key environmental factor. .

[0079] In this embodiment, each extracted key environmental factor has a different numerical range and scale. Next, the key environmental factors need to be standardized and normalized to ensure that all key environmental factors can be processed uniformly. Then, the key environmental factors are combined into an environmental prediction vector in a certain order. The environmental prediction vector contains all the key predictive information about the environmental state. The specific formula is as follows:

[0080] in, These are the standardized or normalized key environmental factors. Each key environmental factor is closely related to the perception tasks of autonomous driving (such as obstacle detection and road sign recognition), providing an important basis for subsequent decision-making and control.

[0081] Step S4: Perform Top-K vector retrieval based on the policy index library to obtain a candidate policy set including several candidate policies. Then, fine-rank the candidate policy set and calculate the prediction vector between each candidate policy and the environment. and overall confidence level The matching distance, matching similarity, and matching degree are used to determine the final fusion strategy parameters. Based on these parameters, a dynamic decision gating system is employed to obtain the final fusion strategy parameters. This is provided for the dynamic execution module to call.

[0082] The strategy matching and decision engine in this embodiment is used to process the environmental prediction vector output by the multi-source environmental prediction module. and its overall confidence level Then, it completes the entire decision-making process of "strategy retrieval - candidate ranking - confidence gating decision - strategy synthesis / backup output". Unlike the traditional "rule-based table lookup", this engine adopts the approach of "vectorized retrieval + two-stage decision + parameterizable strategy synthesis" to achieve fast optimal matching for deterministic environments, flexible mixing for uncertain environments, and safe backup for unknown environments.

[0083] In this embodiment, the policy index library is defined as a set of policy entries:

[0084] in, This indicates the total number of policy entries in the policy index, the [number]th [entry]. Each strategy entry can be represented as:

[0085] in: For the first A unique identifier for each strategy entry; For the first Each strategy entry corresponds to an applicable environment reference vector, and the environment reference vector and the environment prediction vector... They have the same dimensions and semantics; For the first The environmental tolerance range parameter corresponding to each strategy entry is used to characterize the allowable deviation range of the strategy in each environmental dimension. The environmental tolerance range parameter is expressed in the form of a dimension-wise tolerance vector or covariance. For the first Each strategy entry corresponds to a fusion strategy parameter package, which consists of directly executable strategy parameters and supports interpolation or adjudication processing. For the first Each policy entry has an activation threshold parameter, which includes at least one of minimum confidence, minimum matching degree, and disabling conditions. For the first The version and effect statistics of each strategy item, wherein the version and effect statistics include at least one of success rate, average error, applicable operating design domain and update time.

[0086] In this embodiment, the Top-K vector retrieval based on the strategy index library yields a candidate strategy set including several candidate strategies, including: Upon receiving the environmental prediction vector Then, through the vector index Perform candidate strategy recall to obtain a candidate strategy set. Specifically, it can be expressed as:

[0087] in, This indicates the preset candidate recall number, used to control the number of candidates retrieved from the vector index. The number of candidate strategies initially retrieved; This represents the set of candidate strategies, i.e., the prediction vector based on the environment. Retrieved from the policy index library Each strategy entry is as follows One candidate strategy, The set of candidate policies, which is an aggregation of candidate policies, can be represented as:

[0088] Each candidate strategy in the candidate strategy set is a candidate strategy to be ranked that has a high probability of matching the current environment. Subsequently, fine ranking calculation is performed on each candidate strategy in the candidate strategy set to reduce the computational overhead caused by calculating each item in the full strategy index.

[0089] For the first in the candidate policy set The nth candidate strategy is calculated. Candidate strategies and environmental prediction vectors Matching distance between and matching similarity To reflect the importance of different environmental dimensions and the tolerance range of each candidate strategy across these dimensions, this embodiment uses a weighted distance with tolerance, calculated as follows:

[0090] in, Represents the environmental prediction vector In the Predicted values ​​across environmental dimensions; Indicates the first The environment reference vector of each candidate strategy In the Reference values ​​for each environmental dimension; Indicates the first The candidate strategies in the first Tolerance range parameters in each environmental dimension; This represents the dimension of the environment vector, i.e., the number of environment features involved in the matching calculation; To prevent tiny positive numbers with a denominator of zero.

[0091] Based on the matching distance This is further mapped to matching similarity. The matching similarity score ranges from 0 to 1. Specifically, it adopts the following exponential mapping form:

[0092] in, This is a decay control parameter for matching similarity mapping, used to adjust the degree of influence of changes in matching distance on the matching similarity results. When the distance is small, even a slight increase in the matching distance will cause the matching similarity to drop rapidly; when... When the value is large, the change in matching similarity with matching distance is more gradual.

[0093] Based on this, we further introduce the overall confidence level. Forming a matching degree This is done to balance the degree of environmental matching with the reliability of the current prediction results. The degree of matching can be expressed as:

[0094] in, For the preset weighting coefficients, and This is used to adjust the relative contributions of the candidate strategy environment matching degree and the overall confidence score to the matching degree. Based on the matching degree... For candidate policy set The candidate strategies are sorted to obtain the fine-ranking result, which is represented as a set of candidate strategy matching degrees and used for subsequent dynamic decision gating.

[0095] This embodiment employs a dual-gated decision logic to output policy parameters for three modes, based on the candidate policy set and fusion confidence level, to achieve intelligent decision-making for deterministic, fuzzy, and unknown environments. Specifically: Define key match quantity:

[0096] in, This represents the candidate strategy with the highest matching degree in the candidate strategy set. This represents the candidate strategy with the second-highest matching degree in the candidate strategy set. For matching discrimination.

[0097] (1) High-confidence exact matching mode, with a global confidence level of the input. The set of matching degrees with candidate strategies is considered to be true when the following conditions are met:

[0098] in, This represents the confidence threshold. As a matching threshold, When matching the discrimination threshold, the high-confidence mode is triggered, and the output is:

[0099] in, This represents the policy parameter package corresponding to the candidate policy with the highest matching degree. This represents the final fusion strategy parameters. This mode corresponds to the typical scenario of "reliable prediction and unique matching," and its advantage is strong decision determinism.

[0100] (2) Fuzzy region continuous synthesis mode: when the environment falls within the intersection of multiple policies, that is:

[0101] in, The threshold for triggering fuzzy synthesis is determined by the matching degree. The low confidence threshold If the match is below the discrimination threshold, it indicates that the current environment does not offer a unique match, and policy synthesis is required. Policy synthesis can be performed in three ways: Continuous parameter weighted fusion: Weighting the interpolable continuous parameters in the candidate strategies.

[0102] in, This represents the number of candidate strategies participating in the synthesis. For the first The matching degree of each candidate strategy To control the temperature parameters for weighted smoothing, For the first Normalized weights of candidate strategies in continuous parameter weighting. Based on the weights for continuous parameters. To merge:

[0103] in, Indicates the first Interpolable continuous parameters for each candidate strategy. This represents the set of hybrid continuous parameters obtained through weighted fusion.

[0104] Weighted decision-making for discrete parameters: For discrete parameters that cannot be interpolated (such as algorithm branches, switch selections, etc.), a weighted decision-making (voting) method is used.

[0105] in, Indicates the first Discrete parameters among candidate strategies. This indicates that the decision was made through a weighted decision or a vote. The resulting set of mixed discrete parameters. It is based on weight The decision result is determined by selecting the one with the highest frequency or the highest weight.

[0106] Final hybridization strategy output: a continuous set of parameters obtained by weighted fusion. Discrete parameter set obtained by weighted decision By combining these elements, a final hybrid output strategy is formed, specifically expressed as follows:

[0107] in, This represents a hybrid output policy synthesized from multiple candidate policies. This represents the policy parameter package that is ultimately loaded into the execution module. The hybrid output policy includes a continuously adjustable parameter portion and a discrete decision parameter portion, thereby enabling a smooth transition and stable output within the boundary region of multiple candidate policies.

[0108] The technical advantage of this mode is that when the current environment is close to multiple candidate strategies, it does not directly switch to a single strategy, but generates a hybrid output strategy through parameter fusion and adjudication. This allows the strategy parameters to be continuously adjusted as the environment changes, thereby avoiding policy mutations and improving the smoothness, stability and environmental adaptability of the control output.

[0109] (3) Unknown Environment Safety Meta-Strategy Model: When the environment is identified as uncertain or with low confidence, i.e.:

[0110] in, This is a low-confidence threshold used to trigger the fuzzy interval strategy synthesis mode. To match a threshold for an unknown environment, trigger the security meta-policy mode and output:

[0111] in, A predefined set of conservative fusion strategy parameters is used to tighten the correlation threshold, improve output confidence, enhance temporal smoothness, and incorporate the current environment prediction vector. As unknown samples, they are written into the experience pool for subsequent updates.

[0112] In some implementations of this embodiment, a feedforward preloading and smooth transition mechanism is also included. To avoid policy loading lag or sudden output changes during environment switching, the policy engine outputs the final policy parameters. Then, feedforward preloading and smooth transition control are further implemented. By pre-calculating the policy trigger time and combining it with a double-buffered smooth switching mechanism, the final fusion policy parameters acquired at the moment can be initialized and smoothly take over the output before reaching the target environment boundary, thereby improving the real-time performance, continuity and stability during the environment switching process.

[0113] Preload trigger point calculations and output them in the strategy engine. Then, the preloading module calculates the preloading trigger time based on the "remaining time to reach the prediction environment boundary" and the "policy initialization time". Let the distance to the prediction environment boundary be... The It can be estimated jointly from map information and positioning results; let the vehicle speed be... ; Set the strategy parameter package The corresponding initialization time is The initialization time can be obtained through historical strategy execution records, offline statistics, or online estimation. The remaining time to reach the predicted environment boundary can then be expressed as:

[0114] in, This is a preset minimum speed limit to prevent calculation errors caused by an excessively small denominator when the vehicle is at low speed or stationary. Furthermore, the preload trigger time can be expressed as:

[0115] in, A preset safety margin time is provided to reserve buffer time for policy initialization, communication delays, scheduling jitter, and execution uncertainties.

[0116] When the clock reaches the trigger time At that time, start the policy parameter package. The corresponding preloading process for the currently acquired final fusion strategy parameters ensures that the currently acquired final fusion strategy parameters complete the necessary initialization, cache establishment, and parameter distribution before entering the target environment.

[0117] In some implementations of this embodiment, a double-buffered seamless switching mechanism is also included. To prevent control jumps or perception output jitters when switching between the previously acquired final fusion strategy parameters and the currently acquired final fusion strategy parameters, a double-buffered mechanism is used to perform the strategy switching.

[0118] in: Buffer-A: Used to maintain the continuous output of the last acquired final fusion strategy parameters, denoted as... ; Buffer-B: Used to hold the output generated after initializing the final fusion strategy parameters, denoted as... .

[0119] Switching time windows Within this process, the previously obtained final fusion strategy parameter output and the currently obtained final fusion strategy parameter output are weighted and fused to obtain the final output:

[0120] in, The time weighting factor represents the time during the handover process and is used to characterize the moment. The current final fusion strategy parameters are output. The proportion in the final output, and the output of the final fusion strategy parameters obtained last time. The proportion is As time progresses from the start of the switch, The time weighting factor is gradually increased, thus smoothly transitioning the output from the previously acquired final fusion strategy parameters to the currently acquired final fusion strategy parameters. In this embodiment, the time weighting factor... It can be represented as:

[0121] in, Indicates the start time of the switch; The strategy switching window duration is the preset time required to smoothly transition from the last obtained final fusion strategy parameters to the currently obtained final fusion strategy parameters. This represents a limiting function, used to restrict the calculation results to an interval. Inside.

[0122] Therefore, when hour, Only the final fusion strategy parameters obtained in the previous iteration are output; when hour, It gradually increases from 0 to 1, in a smooth switching phase; when hour, Complete the switch and output only the final fusion strategy parameters obtained at the moment.

[0123] By employing the aforementioned double-buffered weighted transition method, output jumps during strategy switching can be avoided, improving the continuity and stability of input and output on the control or sensing side. Furthermore, the... It can be preset according to the response characteristics of the execution module, control cycle, calculation delay or stability requirements, or it can be adaptively adjusted according to the running status.

[0124] In some implementations of this embodiment, an effect monitoring and online update mechanism is also included. To improve the long-term adaptability to complex environments, this mechanism is further implemented. After the strategy is executed, the prediction results of the current environment, the actual executed strategy, and the execution effect are correlated and recorded. Based on the operational feedback, the strategy index library is continuously corrected and supplemented, thereby forming a closed-loop optimization process of "prediction-decision-execution-evaluation-update".

[0125] 1. Experience Unit Records Record experience quadruples:

[0126] in, This represents the environmental prediction vector used when making strategic decisions. This indicates the policy parameter package that is currently being loaded and executed; This indicates the actual perceived or control effect after the strategy takes effect; This represents the true environment vector obtained by processing the actual observation feedback information using the posterior environment coding structure.

[0127] In this embodiment, the This includes at least one of the following: false detection rate, false negative rate, tracking drift, delay, stability index, control error, or boundary safety index. It can be constructed from the posterior environment coding structure based on the real observation data collected during the execution phase, and is used to characterize the actual environmental state during policy execution.

[0128] 2. Online content updates Actions that can be taken for online updates include: The final fusion strategy parameters obtained in the unknown area are sampled and stored in the database. When the environmental prediction vector is discovered... If the unknown environment mode is frequently triggered and the actual execution effect does not meet the preset requirements, a new strategy sampling point will be generated. This information is then written to the policy index library to expand the existing policy index library's coverage of this type of environment region. Among these, This represents the environment vector corresponding to the newly sampled environment point. This represents the currently acquired final fusion strategy parameter package that matches the environment vector.

[0129] Strategy entry statistics field update: Based on the strategy execution results, statistical fields of the corresponding entries in the strategy index are analyzed. The data is updated for use in subsequent candidate strategy ranking, matching degree, and gating decisions. In one implementation, the statistical field... This includes at least one of the following evaluation metrics: success rate, average error, historical hit count, applicable operating domain, update time, or stability evaluation.

[0130] 3. Closed-loop optimization effect Through the aforementioned online recording and updating mechanism, the mapping relationship between "environment-strategy-effect" can be continuously accumulated, and the strategy index can be dynamically corrected and expanded using actual execution feedback. This gradually improves the accuracy of strategy retrieval, the rationality of gating decisions, and the environmental adaptability of strategy output, thus forming a closed-loop optimization process of prediction, decision-making, execution, evaluation, and updating.

[0131] In summary, this application proposes a pre-loading method for autonomous driving perception fusion strategies based on environmental prediction. First, it eliminates the policy adaptation gap caused by sudden environmental changes by achieving zero perception lag in policy switching through feedforward pre-loading. Second, it addresses the rigidity of traditional fixed rules or single matching modes in the face of prediction uncertainty and complex unknown environments by introducing an intelligent and flexible dynamic decision-making mechanism. Finally, this application aims to build an ecosystem with continuous self-optimization capabilities, capable of autonomously learning from actual operational data to continuously improve the prediction accuracy and policy matching accuracy of future environments. This fundamentally enhances the perception reliability, driving safety, and system intelligence of autonomous vehicles in all weather and all scenarios. According to a second aspect of this application, this embodiment provides an autonomous driving perception fusion strategy preloading system based on environmental prediction, comprising: The data acquisition module is used to acquire multi-source sensing data; The overall confidence level acquisition module is used to acquire the source confidence level of each sensing data, and to perform source confidence level weighted fusion on the multi-source sensing data to obtain the overall confidence level; An environmental prediction vector acquisition module is used to extract several key environmental factors from the multi-source sensing data, preprocess and aggregate the key environmental factors to obtain an environmental prediction vector. The matching degree acquisition module is used to perform Top-K vector retrieval based on the strategy index library to obtain a set of candidate strategies including several candidate strategies, and to obtain the matching degree of each candidate strategy based on the overall confidence and the environment prediction vector. The final fusion strategy parameter acquisition module is used to obtain decisions through dynamic decision gating and to obtain the final fusion strategy parameters based on the matching degree of each candidate strategy.

[0132] Specifically, this embodiment corresponds one-to-one with the above method embodiments. The functions of each module have been described in detail in the corresponding method embodiments, so they will not be repeated here.

[0133] According to a third aspect of this application, this embodiment provides a computer-readable storage medium having a computer program stored thereon, the computer program including executable instructions that, when executed by a processor, implement the method described above.

[0134] The present invention can implement all or part of the processes in the above methods, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0135] According to a fourth aspect of this application, an electronic device is provided, comprising: One or more processors; Memory is used to store executable instructions for the processor, which, when executed by one or more processors, cause one or more processors to implement the methods described above.

[0136] Electronic devices are manifested in the form of general-purpose computing devices. Components of an electronic device may include, but are not limited to: at least one processor, at least one memory, and a bus connecting different system components (including memory and processor).

[0137] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of a computer system, connecting all parts of the computer system through various interfaces and lines.

[0138] Memory can be used to store computer programs and / or modules. The processor implements various functions of the computer system by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system and at least one application program required for a function (e.g., sound playback, image playback, etc.); the data storage area can store data created based on the use of the mobile phone (e.g., audio data, video data, etc.). Furthermore, memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMedia Cards (SMC), Secure Digital (SD) cards, Flash Cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0139] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, servers, or computer program products. Therefore, the present invention takes the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention takes the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and memory) containing computer-usable program code.

[0140] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), servers, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0141] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0142] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0143] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0144] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0145] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for preloading perception fusion strategies for autonomous driving based on environmental prediction, characterized in that, include: Acquire multi-source sensing data; Obtain the source confidence score of each sensing data point, and perform source confidence score weighted fusion on the multi-source sensing data to obtain the overall confidence score; Several key environmental factors are extracted from the multi-source sensing data, and the key environmental factors are preprocessed and aggregated to obtain an environmental prediction vector. Top-K vector retrieval is performed based on the strategy index library to obtain a set of candidate strategies including several candidate strategies. The matching degree of each candidate strategy is obtained based on the overall confidence and the environment prediction vector. Decisions are obtained through dynamic decision gating, and the final fusion strategy parameters are obtained based on the matching degree of each candidate strategy.

2. The method according to claim 1, characterized in that: In the process of performing source confidence weighted fusion on the multi-source sensing data, the reciprocal of the sensor measurement variance corresponding to each sensing parameter is used as the source confidence weight of the sensing parameter.

3. The method according to claim 2, characterized in that, This also includes adjusting the source confidence weights, specifically: Based on the environmental state information and sensor state information corresponding to each sensing parameter, a weight adjustment function is constructed to adjust the source confidence weight of each sensing parameter. The calculation formula is as follows: in, Indicates the first Weight adjustment function for each sensing parameter, Indicates the first Source confidence weights for each sensing parameter Indicates weather-related parameters. This indicates the parameter affecting visibility. Indicates the first The state influence parameter of the first sensing parameter is used to characterize the acquisition of the first sensing parameter. The degree of abnormality in the working state of the sensor for each sensing parameter. The source confidence weights are adjusted based on the state. No. The weight adjustment function for each sensing parameter is: in, , , To obtain the first The influence coefficients corresponding to the sensor's sensing parameters are used to characterize the degree of influence of weather factors, visibility factors, and sensor condition factors on sensor reliability; and ; The source confidence weights after state adjustment are adjusted based on the prediction error to obtain the error-adjusted source confidence weights, which are then used as the final source confidence weights.

4. The method according to claim 1, characterized in that: The policy index library is defined as a collection of policy entries: in, This indicates the total number of policy entries in the policy index, the [number]th [entry]. Each strategy entry can be represented as: in: For the first A unique identifier for each strategy entry; For the first Each strategy entry corresponds to an applicable environment reference vector, and the environment reference vector and the environment prediction vector... They have the same dimensions and semantics; For the first The environmental tolerance range parameter corresponding to each strategy entry is used to characterize the allowable deviation range of the strategy in each environmental dimension. The environmental tolerance range parameter is expressed in the form of a dimension-wise tolerance vector or covariance. For the first Each strategy entry corresponds to a fusion strategy parameter package, which consists of directly executable strategy parameters and supports interpolation or adjudication processing. For the first Each policy entry has an activation threshold parameter, which includes at least one of minimum confidence, minimum matching degree, and disabling conditions. For the first The version and effect statistics of each strategy item, wherein the version and effect statistics include at least one of success rate, average error, applicable operating design domain and update time.

5. The method according to claim 1, characterized in that, The Top-K vector retrieval based on the strategy index library yields a candidate strategy set including several candidate strategies, including: Upon receiving the environmental prediction vector Then, through the vector index Perform candidate strategy recall to obtain a candidate strategy set. : in, This indicates the preset candidate recall number, used to control the number of candidates retrieved from the vector index. The number of candidate strategies initially retrieved; Represents the set of candidate policies, based on the environmental prediction vector. Retrieved from the policy index library Each strategy entry is as follows One candidate strategy, A set of candidate strategies, which is an aggregation of individual candidate strategies.

6. The method according to claim 1, characterized in that, The process of obtaining the matching degree of each candidate strategy based on the overall confidence level and the environment prediction vector includes: Obtain the prediction vectors of each candidate strategy and environment. Matching distance between : in, Represents the environmental prediction vector In the Predicted values ​​across environmental dimensions; Indicates the first The environment reference vector of each candidate strategy In the Reference values ​​for each environmental dimension; Indicates the first The candidate strategies in the first Tolerance range parameters in each environmental dimension; This represents the dimension of the environment vector, i.e., the number of environment features involved in the matching calculation; To prevent tiny positive numbers with a denominator of zero; The matching distance is represented by an exponential mapping. Mapped to matching similarity The matching similarity score ranges from 0 to 1, and is calculated using the following formula: in, To match the decay control parameters of the similarity mapping; The overall confidence score and the matching similarity score of each candidate strategy are weighted and fused to obtain the matching score of each candidate strategy.

7. The method according to claim 1, characterized in that, The process of obtaining decisions through dynamic decision gating and obtaining final fusion strategy parameters based on the matching degree of each candidate strategy includes: Define a key matching quantity, which includes the candidate strategy with the highest matching degree in the candidate strategy set. The second highest candidate strategy Matching discrimination ; When satisfied This triggers a high-confidence mode, outputting the final fusion strategy parameters as the strategy parameter package corresponding to the candidate strategy with the highest matching degree. As the confidence threshold, As a matching threshold, When matching the discrimination threshold, Indicates the overall confidence level; When satisfied For interpolable continuous parameters in the candidate strategies, interpolation weights are obtained, and the continuous parameters are fused based on these weights to obtain a first hybrid continuous parameter set. For non-interpolable continuous parameters in the candidate strategies, a weighted decision is applied to obtain a second hybrid continuous parameter set. The first and second hybrid continuous parameter sets are then aggregated to obtain the final fused strategy parameters. The threshold for triggering fuzzy synthesis is determined by the matching degree. The low confidence threshold; When satisfied This triggers the security meta-policy mode, outputting the final fusion policy parameters as a predefined set of conservative fusion policy parameters, where... Match thresholds for unknown environments.

8. The method according to claim 1, characterized in that, After obtaining the final fusion strategy parameters, the following is also included: At the preload trigger moment, feedforward preloading and smooth transition control are executed. The method for obtaining the preload trigger moment is as follows: The remaining time to reach the predicted environmental boundary is calculated using the following formula: in, To determine the distance to the predicted environmental boundary, For the vehicle's speed, This is the preset minimum speed limit; Based on the remaining time to reach the predicted environment boundary The time consumed by the final fusion strategy parameters The preload trigger time is calculated using the following formula: in, This is a preset safety margin time.

9. The method according to claim 1, characterized in that, After obtaining the final fusion strategy parameters, the following is also included: A double-buffering mechanism is used to perform policy switching. Buffer-A is used to maintain the continuous output of the last acquired final fusion policy parameters, denoted as... Buffer-B: Used to carry the continuous output of the currently acquired final fusion strategy parameters, denoted as... ; Within the switching time window, the previously acquired final fusion strategy parameters and the currently acquired final fusion strategy parameter outputs are weighted and fused. The calculation formula is as follows: in, This represents the time weighting factor during the handover process; The time weighting factor It can be represented as: in, Indicates the start time of the switch; Strategy switching window duration; Represents the amplitude limiting function; when hour, Execute Buffer-A; when hour, The value gradually increases from 0 to 1, indicating a smooth transition from Buffer-A to Buffer-B; when hour, Execute Buffer-B.

10. A preloading system for autonomous driving perception fusion strategies based on environmental prediction, characterized in that, include: The data acquisition module is used to acquire multi-source sensing data; The overall confidence level acquisition module is used to acquire the source confidence level of each sensing data, and to perform source confidence level weighted fusion on the multi-source sensing data to obtain the overall confidence level; An environmental prediction vector acquisition module is used to extract several key environmental factors from the multi-source sensing data, preprocess and aggregate the key environmental factors to obtain an environmental prediction vector. The matching degree acquisition module is used to perform Top-K vector retrieval based on the strategy index library to obtain a set of candidate strategies including several candidate strategies, and to obtain the matching degree of each candidate strategy based on the overall confidence and the environment prediction vector. The final fusion strategy parameter acquisition module is used to obtain decisions through dynamic decision gating and to obtain the final fusion strategy parameters based on the matching degree of each candidate strategy.

Images