A multi-expert fusion cross-expert comparable reliability representation and selective non-fusion decision method

Abstract

The application discloses a kind of fusion multi-expert cross-expert comparable reliability representation and selective non-fusion decision method, and the field belongs to multimodal perception and intelligent decision-making field, comprising: obtaining the original summary index of camera expert, laser radar expert and fusion expert;Cross-expert comparable calibration is carried out to each original summary index, including global percentile mapping, conditional percentile mapping and weighted fusion, to obtain comparable reliability core index in unified semantic space;Based on the relative leading relationship between the comparable reliability core index and expert, conservative candidate selection is carried out, to generate candidate expert and stability signal;According to the environmental quality information, the environmental conditions are determined, and in the degradation state, the fusion candidate is executed based on the reliability gain to review the fusion execution margin, and the optimal single expert is rolled back when not passing the review;After rollback, the reliability difference between the fusion expert and the single expert is selectively restored to the fusion output.The application improves the consistency and robustness of multimodal perception decision.

Description

Technical Field

[0001] This invention belongs to the field of multimodal perception and intelligent decision-making, and particularly relates to a cross-expert comparable reliability representation and selective non-fusion decision-making method that integrates multiple experts. Background Technology

[0002] With the development of autonomous driving, mobile robots, and other multi-sensor intelligent systems, multimodal perception methods based on multi-source information such as cameras and LiDAR have become a key technical route for improving environmental perception accuracy and system robustness. In existing multimodal perception systems, a camera-based visual perception branch, a LiDAR-based point cloud perception branch, and a cross-modal fusion perception branch combining the two are typically constructed simultaneously. These branches utilize visual information, geometric information, and cross-modal joint information to complete tasks such as target detection, scene understanding, or state determination. Because different perception branches have inherent differences in information sources, feature representation methods, training mechanisms, and decision paths, each branch often exhibits different applicability and reliability under different degradation conditions (such as image blurring, sparse point clouds, and partial occlusion). Therefore, how to rationally select among multiple perception branches has become an important issue in multimodal perception decision-making. In existing technologies, a common approach is to directly compare the confidence scores, detection statistics, or quality assessment indicators of each branch's output and select the final branch or detection result accordingly. Another approach is to dynamically fuse or select the outputs of different branches by designing heuristic rules or learning-based routing modules.

[0003] However, existing multimodal perception decision-making technologies still have the following shortcomings: First, the original indicators such as confidence scores and summary statistics output by different perception branches usually come from different feature spaces, training objectives, and decision paths. Their numerical distribution range, scale, and statistical semantics all differ significantly. Directly comparing these indicators horizontally lacks a rigorous mathematical foundation and semantic consistency premise. That is, the same value does not represent equivalent reliability meanings between different branches, leading to the general failure of cross-expert comparability. Second, under the premise of insufficient cross-expert comparability, when the system directly selects experts based on the original scores or insufficiently aligned summary indicators, there may be situations where branches that are more reliable in terms of actual task utility lose in the comparison due to their conservative score expression, low numerical scale, or semantic deviation. This advantage reversal or selection distortion phenomenon will seriously reduce the accuracy and stability of the final perception decision. Especially under complex conditions with multiple quality perturbations, the risk of misselection will be further amplified. Furthermore, in multimodal perception scenarios, existing methods generally exhibit a default tendency to retain fusion branches. That is, if the fusion result shows a slight lead over the single-modal result, the fusion output is directly retained. However, in practical applications, the advantage of the fusion branch is not always sufficient or stable. Under degradation conditions, the fusion result sometimes only shows a weak lead. Such a weak advantage is insufficient to constitute a sufficient basis for stably retaining the fusion output. If the system accepts the fusion result without review, it can easily lead to excessive retention of the fusion result, impairing the robustness and interpretability of the overall decision. In addition, existing multi-expert decision-making methods often suffer from unclear coupling of responsibilities in their architectural design. Functions such as candidate generation, risk review, and rollback processing are mixed in the same level, increasing the difficulty of method debugging and ablation analysis, and hindering the effective verification of each decision-making stage. At the same time, existing methods generally lack external validation anchors that directly align with the real-world task utility at the sample level. This results in system performance optimization often being limited to improving the internal scoring space, making it difficult to guarantee stable consistency between internal decision ranking and actual detection accuracy. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a cross-expert comparable reliability representation and selective non-fusion decision-making method that integrates multiple experts, comprising: Obtain the raw processing results output by multiple experts for the data to be processed; Cross-expert comparability calibration is performed on the original processing results to obtain the core indicators of comparability reliability for each expert with unified comparative semantics under the current data. Based on the comparable reliability core indicators and the relative relationships among experts, candidate experts are determined from among the experts. Obtain environmental condition information of the current data, and execute a selective decision-making process based on the candidate experts and the environmental condition information; When the selective decision-making process determines that a specific expert among the candidate experts needs to be rejected, it falls back to other experts as the final output. After rollback, the specific expert is selectively restored as the final output based on the reliability difference between the specific expert and the other experts.

[0005] Optionally, cross-expert comparability calibration is performed on each of the original processing results, including: For each type of original processing indicator corresponding to each expert, obtain the set of values ​​of the indicator in all historical outputs of the expert, and calculate the ascending position of the current original processing indicator value in the set of values. Based on the ratio of the ascending position to the total sample size of the set of values, determine the global percentile mapping result, so that the original indicators with different values ​​are transformed into a unified 0,1 interval to represent their relative position within the expert. Extract the environmental quality information of the current data. The environmental quality information includes at least one or more of the following: image blur information, image brightness information, point cloud density information, and point cloud coverage information. Determine the condition group to which the current data belongs based on the image blur information in the environmental quality information. Within the same condition group, recalculate the ratio of the ascending order of the current original processing index value to the group sample size to determine the condition percentile mapping result. A fusion coefficient is configured for each type of original processing index. The global percentile mapping result and the conditional percentile mapping result are weighted and summed. The fusion coefficient is used to control the weight ratio of the global mapping and the conditional mapping respectively to obtain a hybrid calibration result. The hybrid calibration result constitutes a comparable reliability representation under the unified semantic space.

[0006] Optionally, obtaining comparable reliability core indicators with unified comparative semantics for each of the aforementioned experts under the current data also includes: Multiple statistics are obtained after cross-expert comparability calibration. The multiple statistics include at least one of the following: mean score calibration value, high-frame average score calibration value, and high-confidence prediction quantity calibration value. An aggregation weight is assigned to each of the statistics. The products of all the statistics and their corresponding aggregation weights are summed to obtain the uncorrected basic reliability value. Extract the image quality information and point cloud quality information of the current data. The image quality information includes at least image brightness features, and the point cloud quality information includes at least point cloud quantity and point cloud coverage features. Normalize each quality information to obtain a normalized quality value. Configure a correction weight for each quality information. Construct an environmental quality correction term based on the normalized quality value and the correction weight. The basic reliability value is superimposed with the environmental quality correction item, and the superposition result is truncated to a preset value range to keep the corrected value bounded, thereby obtaining the comparable reliability core index.

[0007] Optionally, candidate experts are determined from among the experts based on the comparable reliability core indicators and the relative relationships between experts, including: The comparable reliability core indicators of each expert can be directly used as the basic reliability score of that expert, or the comparable reliability core indicators can be monotonically transformed and used as the basic reliability score. For each target expert, the differences between the comparable reliability core indicators of the target expert and the comparable reliability core indicators of all other experts are calculated. All the differences are summed to obtain the total lead margin of the target expert, or the maximum value among all the differences is taken as the total lead margin of the target expert. The total lead margin is then input into the Sigmoid function for nonlinear mapping to obtain the competitive support of the target expert. The steepness parameter of the Sigmoid function is used to control the sensitivity of the lead margin to the support. The basic reliability score of each expert is weighted and combined with the competitive support score, or a product combination is performed, to construct the candidate evaluation value of each expert. The candidate evaluation values ​​of all experts are compared, and the expert with the highest evaluation value is selected as the candidate expert.

[0008] Optionally, determining candidate experts from among the experts based on the comparable reliability core indicators and the relative relationships between experts further includes: The multiple experts include at least a first single expert based on a camera, a second single expert based on lidar, and a fusion expert. During the candidate evaluation process, the expert with better comparable core reliability indicators is determined from the first single expert and the second single expert as the optimal single expert. Calculate the difference between the comparable reliability core index of the fusion expert and the comparable reliability core index of the optimal single expert, and use the difference as the reliability gain of the fusion expert. Configure a preset gain threshold, compare the reliability gain with the preset gain threshold, if the reliability gain is lower than the preset gain threshold, apply an inhibition coefficient to the competitive support of the fusion expert to reduce it, if the reliability gain is not lower than the preset gain threshold, keep the competitive support of the fusion expert unchanged, the value of the inhibition coefficient is positively correlated with the degree to which the reliability gain is lower than the preset gain threshold or is a fixed attenuation value; Based on the basic reliability score and suppressed competitive support of the fusion experts, a candidate evaluation value for the fusion experts is constructed, making it difficult for the fusion experts to be selected as candidate experts when they only have a weak advantage.

[0009] Optionally, environmental condition information of the current data is obtained, and a selective decision-making process is executed based on the candidate experts and the environmental condition information, including: The image blur information, image brightness information, and point cloud density information of the current data are extracted. After normalizing each information, an environmental degradation weight is assigned to each of them. The product of all information and its corresponding environmental degradation weight is weighted and summed to obtain the environmental degradation intensity index. Configure a degradation judgment threshold, compare the environmental degradation intensity index with the degradation judgment threshold, if the environmental degradation intensity index is lower than the degradation judgment threshold, then determine that the current environmental conditions are in a non-degraded state, directly use the candidate expert as the final output, and do not perform any rejection operation; If the environmental degradation intensity index is not lower than the degradation judgment threshold, the current environmental conditions are determined to be in a degraded state. Further, it is determined whether the candidate expert is a fusion expert. If the candidate expert is a single expert, the single expert is directly used as the final output. If the candidate expert is a fusion expert, the execution conditions of the fusion expert are reviewed.

[0010] Optionally, triggering an execution condition review of the fusion expert includes: The expert with better comparable reliability core indicators between the first single expert and the second single expert is identified as the optimal single expert, and the comparable reliability core indicators of the optimal single expert are obtained. Calculate the difference between the comparable reliability core index of the fusion expert and the comparable reliability core index of the optimal single expert, and multiply the difference by a preset fusion coefficient to obtain the fusion execution margin. The fusion coefficient is a value between 0 and 1, which is used to adjust the magnitude of the conversion of reliability difference into execution margin. Configure a preset review threshold, compare the fusion execution margin with the preset review threshold, and if the fusion execution margin is greater than or equal to the preset review threshold, then determine that the fusion expert meets the execution conditions and retain the fusion expert as the final output; If the fusion execution margin is less than the preset review threshold, the fusion expert is determined not to meet the execution conditions, the fusion expert is rejected, and the process reverts to the optimal single expert as the final output.

[0011] Optionally, after rollback, based on the reliability difference between the specific expert and the other experts, the specific expert is selectively restored as the final output, including: After the fusion expert is rejected due to failure to meet the execution conditions and the process reverts to the optimal single expert, the candidate expert identification information output by the candidate selection stage is obtained to confirm whether the candidate expert output by the candidate selection stage is the fusion expert. It is confirmed that the fusion expert has been rejected and a rollback operation has been performed, and the final output has been switched to the optimal single expert; Calculate the difference between the comparable reliability core index of the fusion expert and the comparable reliability core index of the optimal single expert, and use the difference as the advantage value of the fusion expert relative to the optimal single expert; Configure a preset protection threshold, compare the advantage value with the preset protection threshold, and if the advantage value is greater than or equal to the preset protection threshold, it is determined that the fusion protection condition is met, the rollback operation is canceled, and the final output is restored from the optimal single expert to the fusion expert. If the advantage value is less than the preset protection threshold, it is determined that the fusion protection condition is not met, and the optimal single expert is retained as the final output.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: First, by performing cross-expert comparability calibration, the original scores of each expert are converted into comparable reliability core indicators under a unified semantic space, effectively solving the problem of cross-expert comparability failure and providing a mathematical basis for horizontal comparison of the outputs of different experts. Second, a conservative reliability-aware candidate selection mechanism is constructed based on the comparable reliability core indicators, reducing the risk of advantage reversal and selection distortion caused by inconsistencies in score semantics, ensuring that more reliable experts can be correctly identified. Third, under degradation conditions, a fusion execution margin review based on reliability gain is performed on fusion candidates, retaining them only when the fusion expert has a sufficient advantage over the best single expert, effectively avoiding the over-retention problem of weak-advantage fusion. Finally, through a fusion protection and compensation mechanism, after rejection, the fusion output is selectively restored based on the reliability difference between the fusion expert and the single expert, correcting over-rejection in a very small number of boundary samples, balancing the conservatism of the system and the decision accuracy of key samples without changing the overall review strategy. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the cross-expert comparability calibration process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the conservative reliability-aware candidate selection mechanism according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the selective non-fusion decision-making process according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the fusion protection mechanism in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides a cross-expert comparable reliability representation and selective non-fusion decision-making method that integrates multiple experts, including: like Figure 1 As shown in the figure, this embodiment proposes a cross-expert comparable reliability representation and selective non-fusion decision-making method for camera experts, LiDAR experts, and fusion experts. The method includes the following steps: acquisition of original summary indicators, cross-expert comparability calibration, construction of core comparable reliability indicators, selection of conservative reliability-aware candidates, determination of environmental conditions, review of selective non-fusion, and protective backfilling for fusion.

[0019] In this embodiment, multimodal input data of the scene to be processed is first acquired. This multimodal input data includes at least image data acquired by a camera and point cloud data acquired by a LiDAR. The acquired multimodal input data is then processed by camera experts, LiDAR experts, and fusion experts, respectively, to obtain the original summary metrics output by each expert. The original summary metrics include, but are not limited to, mean score, average score of the top few high-scoring bounding boxes, number of high-confidence predictions, number of prediction boxes, or other statistics that can reflect the expert's output status. These original summary metrics are directly derived from each expert's own feature space and decision path and have not yet undergone cross-expert semantic alignment processing.

[0020] After obtaining the original abstract metrics from each expert, this embodiment performs cross-expert comparability calibration on each of the original abstract metrics. The core purpose of this calibration process is to map the original metrics output by different experts, which were originally not directly comparable due to their different sources, to a unified semantic space, thereby obtaining a reliable representation with comparability. Specifically, cross-expert comparability calibration includes three levels: global percentile calibration, conditional percentile calibration, and hybrid calibration.

[0021] At the global percentile calibration level, for each expert and each type of original summary indicator, the set of values ​​for that indicator across all of the expert's historical outputs is obtained. The ascending order position of the current original summary indicator value within this set is calculated. Based on the ratio of this ascending order position to the total sample size of the set, the global percentile mapping result is determined. Through this global percentile mapping, original indicators with different values ​​are transformed into a unified interval. This result characterizes the relative position of the current indicator value within the expert's overall distribution, thus ensuring consistent ranking semantics for indicators within each expert.

[0022] At the conditional percentile calibration level, this embodiment further performs conditional grouping processing on the original summary indicators based on the environmental quality context. Specifically, environmental quality information of the current sample is extracted, which includes at least one or more of image blur information, image brightness information, point cloud density information, and point cloud coverage information. The conditional group to which the current sample belongs is determined based on the image blur information in the environmental quality information. Within the same conditional group, the ratio of the ascending order of the current original indicator value to the group sample size is recalculated, thereby determining the conditional percentile mapping result. This conditional percentile mapping can reflect the influence of different environmental conditions on the indicator distribution, giving the calibration result context-adaptive capabilities.

[0023] After completing the global percentile mapping and conditional percentile mapping, this embodiment performs a weighted fusion process on the two. Specifically, a corresponding fusion coefficient is configured for each type of original summary indicator. The global percentile mapping result and the conditional percentile mapping result are weighted and summed. The fusion coefficient is used to control the weight ratio of the global mapping and the conditional mapping, thereby obtaining a hybrid calibration result. This hybrid calibration result constitutes a cross-expert comparable reliability representation in a unified semantic space, making the outputs of different experts, which were originally not directly comparable, numerically comparable relative semantics.

[0024] After obtaining the comparable reliability representations for each expert, this embodiment further constructs core comparable reliability indicators for each expert under the current sample. The construction process of these core comparable reliability indicators is as follows: First, multiple statistics calibrated for cross-expert comparability are obtained. These multiple statistics include at least several of the following: mean score calibration value, high-frame average score calibration value, and high-confidence prediction quantity calibration value. An aggregation weight is assigned to each statistic, and the products of all the statistics and their corresponding aggregation weights are summed to obtain the uncorrected basic reliability value.

[0025] Furthermore, this embodiment introduces an environmental quality correction term to constrain the adjustment of the basic reliability value. Specifically, image quality information and point cloud quality information of the current sample are extracted. The image quality information includes at least image brightness features, and the point cloud quality information includes at least point cloud quantity and point cloud coverage features. Normalized quality values ​​are obtained by normalizing each quality information, and correction weights are assigned to each type of quality information. An environmental quality correction term is constructed based on the normalized quality values ​​and the correction weights. This environmental quality correction term is adjusted through a weighting factor. Then, the basic reliability value and the environmental quality correction term are superimposed, and the superposition result is truncated to a preset numerical range to ensure that the corrected value remains bounded. This correction process satisfies boundedness and monotonicity constraints to ensure that the adjustment to the original comparable reliability representation is a mild and interpretable correction, thereby forming a unified reliability metric for cross-expert comparison, i.e., the core comparable reliability index.

[0026] After obtaining comparable core reliability indicators from three experts, this embodiment enters the conservative reliability perception candidate selection stage. The core task of this stage is to generate candidate experts and their stability signals based on the comparable core reliability indicators and the relative leading relationships among the experts. This stage is only responsible for candidate generation and does not perform rejection or backoff operations.

[0027] Specifically, the candidate selection mechanism first constructs a basic reliability score based on the comparable core reliability indicators of each expert. In one feasible implementation, the comparable core reliability indicators of each expert are directly used as the basic reliability score for that expert, or the comparable core reliability indicators are monotonically transformed and then used as the basic reliability score.

[0028] Then, the lead margin is calculated based on the relative differences among the experts. For each target expert, the difference between the target expert's comparable reliability core index and the comparable reliability core indices of all other experts is calculated. All these differences are summed or the maximum value is taken to obtain the total lead margin of the target expert. The total lead margin is then input into a Sigmoid function for non-linear mapping to obtain the competitive support of the target expert. The steepness parameter of the Sigmoid function is used to control the sensitivity of the lead margin to support conversion.

[0029] Based on this, a candidate evaluation function is constructed using the comparable reliability core indicators and competitive support. Specifically, the basic reliability score and competitive support of each expert are weighted and combined or multiplied to construct the candidate evaluation value for each expert. The candidate evaluation values ​​of all experts are compared, and the expert with the highest evaluation value is selected as the candidate expert. Simultaneously, the stability signal corresponding to the candidate expert is output. This stability signal is used for subsequent selective non-fusion review. The stability signal can be taken as the difference between the candidate score and the second-best score.

[0030] For fusion experts, this embodiment introduces a reliability gain constraint term relative to the optimal single expert during the candidate selection process, subjecting fusion experts to stricter selection constraints compared to single experts. Specifically, the expert with better comparable reliability core indicators from camera experts and LiDAR experts is selected as the optimal single expert. The difference between the comparable reliability core indicators of the fusion expert and the comparable reliability core indicators of the optimal single expert is calculated, and this difference is used as the reliability gain of the fusion expert. An asymmetric support mechanism is constructed based on this reliability gain: a preset gain threshold is configured, and the reliability gain is compared with the preset gain threshold. Only when the reliability gain of the fusion expert relative to the optimal single expert exceeds the preset threshold is the fusion expert allowed to obtain candidate support equivalent to that of the single expert. When the reliability gain does not reach the preset threshold, the candidate support of the fusion expert is suppressed, so that the fusion expert will not be selected as a candidate expert if it only has a weak advantage.

[0031] After candidate selection, this embodiment determines the environmental conditions of the current sample based on pre-extracted environmental quality information. Specifically, environmental quality information reflecting the degree of image degradation and the point cloud quality state is extracted. This environmental quality information includes at least image blur information, image brightness information, and point cloud density information. After normalizing each piece of information, an environmental degradation weight is assigned to each. The product of all information and its corresponding environmental degradation weight is weighted and summed to obtain an environmental degradation intensity index. A degradation judgment threshold is configured, and the environmental degradation intensity index is compared with the degradation judgment threshold to classify the current sample into a non-degraded state or a degraded state. When the environmental degradation intensity index is lower than the degradation judgment threshold, the environmental conditions of the current sample are determined to be non-degraded; when the environmental degradation intensity index is not lower than the degradation judgment threshold, the environmental conditions of the current sample are determined to be degraded.

[0032] After determining the environmental conditions, this embodiment executes a selective non-fusion decision-making process based on the candidate experts. When the environmental conditions are determined to be in a non-degenerate state, the candidate experts output in the candidate selection stage are directly used as the final output, without performing additional review.

[0033] When the environmental condition assessment result is a degraded state, selective non-fusion review and subsequent fusion protection mechanism are initiated. Specifically, the candidate experts output in the candidate selection stage are used as the initial decision objects. First, the type of the candidate expert is determined: when the candidate expert is a single expert (i.e., a camera expert or a lidar expert), the single expert is directly used as the output result without additional review; when the candidate expert is a fusion expert, the execution condition review for the fusion candidate is initiated.

[0034] The execution condition review for fusion candidates is as follows: First, the expert with the better comparable reliability core index among the camera expert and LiDAR expert is identified as the optimal single expert, and the comparable reliability core index of the optimal single expert is obtained. The difference between the comparable reliability core index of the fusion expert and the comparable reliability core index of the optimal single expert is calculated, and the fusion execution margin is determined based on this difference. Specifically, the difference can be directly used as the fusion execution margin, or the difference can be multiplied by a preset fusion coefficient to obtain the fusion execution margin. Then, the fusion execution margin is compared with a preset review threshold. When the fusion execution margin is not lower than the preset review threshold, the fusion candidate is determined to meet the execution conditions, and the fusion candidate is retained as the final output result; when the fusion execution margin is lower than the preset review threshold, the fusion candidate is determined to not meet the execution conditions, the fusion candidate is rejected, and the process reverts to the optimal single expert output.

[0035] After a fusion candidate is rejected for failing the selective non-fusion review and has fallen back to the optimal single-expert output, this embodiment further executes a fusion protection mechanism. This fusion protection mechanism is used to correct for excessive rejection of the fusion result that may occur in a very small number of boundary samples.

[0036] Specifically, after the rollback operation is completed, it is further determined whether the fusion protection conditions are met. The fusion protection conditions include at least the following three conditions: First, the candidate experts output during the candidate selection phase are fusion experts; second, the fusion expert was rejected and a rollback has been performed during the selective non-fusion review phase; third, the comparable reliability core index of the fusion expert has an advantage of at least a preset protection threshold relative to the optimal single expert, that is, the difference between the comparable reliability core index of the fusion expert and the comparable reliability core index of the optimal single expert is greater than or equal to the preset protection threshold. When all three conditions are met simultaneously, it is determined that the fusion protection conditions are met, the rollback result is revoked, and the fusion output is restored; when the fusion protection conditions are not met, the output of the optimal single expert is retained as the final result.

[0037] After the above-mentioned processing steps, for each sample to be processed, this embodiment finally outputs the final selected expert, the final decision-making mode, and the comparable reliability core indicators, candidate scores, and environmental condition judgment results corresponding to the final decision. The decision-making mode includes one of the following: direct output under non-degradation conditions, fusion retention under degradation conditions, regression to a single expert under degradation conditions, and restoration of fusion output through a protection mechanism.

[0038] Figure 1 This is a schematic diagram of the overall process of the cross-expert comparable reliability representation and selective non-fusion decision-making method proposed in this invention. The process includes the main steps of obtaining the original summary indicators, cross-expert comparable calibration, candidate selection, and selective non-fusion decision-making. Figure 2 This is a schematic diagram of the cross-expert comparability calibration process in this invention, showing the global percentile mapping, conditional percentile mapping and their fusion processing of the original summary indicators of different experts, thereby constructing a comparable reliability representation under a unified semantic space; Figure 3 This is a schematic diagram of the conservative reliability perception candidate selection mechanism in the present invention, which shows the process of uniformly evaluating single experts and fusion experts and generating candidate experts based on comparable reliability core indicators and the competitive support relationship among experts. Figure 4 This is a schematic diagram of the selective non-fusion decision-making process in this invention, illustrating the process of conditionalizing candidate experts under different environmental conditions, including direct output under non-degenerate conditions and fusion execution margin review and rollback mechanism under degenerate conditions. Figure 5This is a schematic diagram of the fusion protection mechanism in this invention. It shows the process of determining whether to restore the fusion output after the fusion candidate is rejected and the system reverts to the single expert output by judging whether the comparable reliability advantage of the fusion expert relative to the optimal single expert exceeds a preset threshold.

[0039] In summary, this embodiment constructs a comparable reliability representation in a unified semantic space through cross-expert comparability calibration, and on this basis, completes candidate selection, conditional review, and fusion protection supplementation, thus forming a complete multi-expert decision-making closed loop. This method can effectively reduce the risk of misselection caused by inconsistencies in score semantics and avoid the default retention of fusion results under weak dominance conditions, while taking into account the robustness, interpretability, and engineering feasibility of the system.

[0040] Example 2 This embodiment provides a cross-expert comparable reliability representation and selective non-fusion decision-making method that integrates multiple experts, including: like Figure 1-5 As shown in the figure, this embodiment proposes a cross-expert comparable reliability representation and selective non-fusion decision-making method for camera experts, LiDAR experts, and fusion experts. The method as a whole includes six steps: original summary index acquisition, cross-expert comparability calibration, conservative reliability-aware candidate selection, environmental condition determination, selective non-fusion review, and fusion protective backfilling.

[0041] In this embodiment, the sample set is denoted as: ; The expert group consists of: ; Where C stands for camera expert, L for lidar expert, and F for fusion expert.

[0042] For any sample i ∈ I, the three experts output their corresponding original summary metrics. Let the vector of the original summary metrics of the e-th expert on sample i be: ; Where K represents the dimension of the summary metric. The summary metric may include, but is not limited to, the mean score, the average score of the top few high-scoring boxes, the number of high-confidence predictions, the number of prediction boxes, or other statistics that can reflect the state of expert output.

[0043] Step 1: Obtaining raw summary metrics; For each sample i to be processed, it is input into the camera expert, LiDAR expert, and fusion expert respectively to obtain the original summary index vectors corresponding to the three experts. , and Simultaneously, auxiliary information related to environmental quality is extracted to construct an environmental quality vector: ; The environmental quality information includes at least one or more of the following: image blur, image brightness, point cloud density, and point cloud coverage.

[0044] Step 2: Cross-expert comparability calibration; Since the original abstract metrics from different experts come from different feature spaces and statistical distributions, they cannot be directly compared horizontally. Therefore, cross-expert comparability calibration is required first.

[0045] 1. Global percentile calibration; For the k-th summary index of the e-th expert, let its set of values ​​across all samples be: ; remember In the set The ascending rank in the sequence is: ; Its global percentile representation is then defined as: ; thus, ∈[0,1], used to characterize the global relative position of sample i in the distribution of the k-th index of expert e.

[0046] 2. Conditional percentile calibration; To reflect the impact of different environmental conditions on the distribution of indicators, this embodiment further performs conditional percentile calibration in the context of environmental quality.

[0047] Suppose that the samples are divided into B context groups based on image blur, and let the group to which sample i belongs be: ; Within the b(i)th group, for the kth summary index of the eth expert, let the sample size of this group be... The conditional percentile representation is then defined as: ; in, ( ) indicates the ascending rank within the corresponding condition group.

[0048] 3. Hybrid calibration; To balance global stability and conditional adaptability, this embodiment performs a weighted fusion of the global percentile results and the conditional percentile results to obtain a hybrid calibration result: ; in, ∈[0,1] is the fusion coefficient corresponding to the k-th index.

[0049] This yields the calibrated index vector: ; Step 3: Constructing comparable core reliability indicators; After obtaining the mixed calibration results of various summary indicators, this embodiment further constructs a core index of comparable reliability to uniformly measure the reliability of different experts on the current sample.

[0050] First, the calibrated metrics are weighted and aggregated to obtain the uncorrected baseline reliability value: ; in, ≥0, and satisfying: ; Furthermore, environmental quality correction items are introduced. Let the normalized value of the m-th environmental quality quantity on sample i be denoted as . Then the bounded correction function can be defined as: ; Therefore, the environmental quality correction term corresponding to the e-th expert is constructed as follows: ; in, To adjust the weight parameters.

[0051] Finally, the core indicator of the comparability reliability of the e-th expert on sample i is defined as: ; in, This means truncating the value to the interval [0,1].

[0052] thus, This is the core comparable reliability indicator in this embodiment, used for subsequent candidate selection, selective non-fusion review, and fusion protection mechanism.

[0053] Step 4: Conservative reliability-aware candidate selection; After obtaining comparable reliability metrics from three experts, this embodiment proceeds to the candidate selection phase. This phase is only responsible for generating candidate experts and stability signals, and does not perform rejection or backoff.

[0054] 1. Basic reliability score; For sample i and expert e, the basic reliability score is defined as: ; 2. Lead margin and competitive support For expert e, its lead relative to other experts is defined as: ; Further define competitive support as: ; in, ; For the Sigmoid function, γ>0 is a parameter that controls the sensitivity of the lead magnitude.

[0055] 3. Single expert candidate scoring; For camera experts and LiDAR experts, the candidate score is defined as follows: ; 4. Integrate expert asymmetric constraints; For fusion experts, this embodiment introduces a reliability gain constraint relative to the optimal single expert. Let the optimal single expert on sample i be: ; Its corresponding core indicators are: ; Define the reliability gain of the fusion expert relative to the optimal single expert as: ; The gain support term for fusion experts is further defined as follows: ; Where β>0 is the gain sensitivity parameter. This is the fusion gain threshold.

[0056] Therefore, the candidate score for the fusion expert is defined as follows: With this design, fusion experts only receive candidate support strength comparable to that of a single expert when they have a sufficient advantage over the best single expert.

[0057] 5. Candidate experts and stability signal output; Finally, the candidate experts on sample i are defined as: ; Meanwhile, the stability signal is defined as the difference between the candidate score and the suboptimal score, i.e.: ; Candidate Experts With stability signal This information will be output to subsequent decision-making stages.

[0058] Step 5: Determine environmental conditions; This embodiment further determines the environmental conditions of the current sample based on environmental quality information. Let the normalized image blur, image brightness, and point cloud density be respectively... , and Then, the intensity of environmental degradation can be defined as: ; Where ω1, ω2, ω3 ≥ 0, and satisfy: ; when If the condition is met, sample i is determined to be in a non-degenerate state; otherwise, it is determined to be in a degenerate state.

[0059] Step Six: Selective Non-Integration Review; For sample i, if its environmental conditions are non-degraded, then the candidate experts are directly adopted. As the final output.

[0060] If the environmental conditions are in a degraded state, it will enter the selective non-fusion review stage.

[0061] 1. Single expert candidate processing; like: ; Then the single expert will be used directly as the output result.

[0062] 2. Fusion candidate processing; like: ; First, determine the optimal single expert: ; Based on this, the fusion execution margin is defined as: ; Where α∈[0,1] is the fusion coefficient.

[0063] The fusion execution margin and review threshold are used. Comparison: when: ; At that time, the fusion candidates are retained as the final output; when: ; At that time, reject the fusion candidate and fall back to the optimal single expert. As the output.

[0064] Step Seven: Integration Protection Mechanism To prevent the fusion results in a very small number of boundary samples from being overly rejected, this embodiment further sets up a fusion protection mechanism.

[0065] When the following conditions are met simultaneously: (1) The candidate experts output during the candidate selection phase are fusion experts, i.e. (2) The fusion candidate was rejected during the selective non-fusion review stage and has been reverted to the best single expert; (3) The fusion of experts still has an advantage over the optimal single expert, which is no less than the preset protection threshold. Then the rollback result will be canceled and the merged output will be restored.

[0066] Otherwise, retain the optimal single-expert output as the final result.

[0067] Step 8: Constructing sample-level real utility anchors; To ensure that the aforementioned comparable representations, candidate selection, and conditional review are all centered around a unified external objective, this embodiment further constructs sample-level real utility anchors.

[0068] Let the number of true positive examples, false positive examples, and false negative examples of expert e on sample i be respectively... , and Then its sample-level detection utility is defined as: ; in, , and Geometric matching is obtained by matching samples of the same category with the actual annotations and prediction results. Geometric matching can be completed based on the center distance threshold of the top view plane.

[0069] Based on this definition of utility, the optimal expert for sample i is: ; This external utility anchor can be used for parameter selection, validation set tuning, or offline evaluation, thereby ensuring that the comparability calibration and decision-making process in this invention are not limited to the consistency of the internal scoring space, but can be aligned with the utility of the real task.

[0070] Step Nine: Overall output of this embodiment; After the above steps, for each sample i to be processed, the final output of this embodiment is: Final selection of experts ; (2) The final decision mode includes one of the following: direct output under non-degenerate conditions, fusion retention under degenerate conditions, regression to single expert under degenerate conditions, and restoration of fusion output by protection mechanism; (3) Comparable reliability core indicators, candidate scores and environmental condition judgment results corresponding to the final decision.

[0071] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0072] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0073] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A cross-expert comparable reliability representation and selective non-fusion decision-making method integrating multiple experts, characterized in that, include: Obtain the raw processing results output by multiple experts for the data to be processed; Cross-expert comparability calibration is performed on the original processing results to obtain the core indicators of comparability reliability for each expert with unified comparative semantics under the current data. Based on the comparable reliability core indicators and the relative relationships among experts, candidate experts are determined from among the experts. Obtain environmental condition information of the current data, and execute a selective decision-making process based on the candidate experts and the environmental condition information; When the selective decision-making process determines that a specific expert among the candidate experts needs to be rejected, it falls back to other experts as the final output. After rollback, the specific expert is selectively restored as the final output based on the reliability difference between the specific expert and the other experts.

2. The method according to claim 1, characterized in that, Cross-expert comparability calibration is performed on the original processing results, including: For each type of original processing indicator corresponding to each expert, obtain the set of values ​​of the indicator in all historical outputs of the expert, and calculate the ascending position of the current original processing indicator value in the set of values. Based on the ratio of the ascending position to the total sample size of the set of values, determine the global percentile mapping result, so that the original indicators with different values ​​are transformed into a unified 0,1 interval to represent their relative position within the expert. Extract the environmental quality information of the current data. The environmental quality information includes at least one or more of the following: image blur information, image brightness information, point cloud density information, and point cloud coverage information. Determine the condition group to which the current data belongs based on the image blur information in the environmental quality information. Within the same condition group, recalculate the ratio of the ascending order of the current original processing index value to the group sample size to determine the condition percentile mapping result. A fusion coefficient is configured for each type of original processing index. The global percentile mapping result and the conditional percentile mapping result are weighted and summed. The fusion coefficient is used to control the weight ratio of the global mapping and the conditional mapping respectively to obtain a hybrid calibration result. The hybrid calibration result constitutes a comparable reliability representation under the unified semantic space.

3. The method according to claim 1, characterized in that, The core indicators of comparable reliability obtained from the experts mentioned above, which have a unified comparative semantic under the current data, also include: Multiple statistics are obtained after cross-expert comparability calibration. The multiple statistics include at least one of the following: mean score calibration value, high-frame average score calibration value, and high-confidence prediction quantity calibration value. An aggregation weight is assigned to each of the statistics. The products of all the statistics and their corresponding aggregation weights are summed to obtain the uncorrected basic reliability value. Extract the image quality information and point cloud quality information of the current data. The image quality information includes at least image brightness features, and the point cloud quality information includes at least point cloud quantity and point cloud coverage features. Normalize each quality information to obtain a normalized quality value. Configure a correction weight for each quality information. Construct an environmental quality correction term based on the normalized quality value and the correction weight. The basic reliability value is superimposed with the environmental quality correction item, and the superposition result is truncated to a preset value range to keep the corrected value bounded, thereby obtaining the comparable reliability core index.

4. The method according to claim 1, characterized in that, Based on the comparable reliability core indicators and the relative relationships among experts, candidate experts are determined from among the experts, including: The comparable reliability core indicators of each expert can be directly used as the basic reliability score of that expert, or the comparable reliability core indicators can be monotonically transformed and used as the basic reliability score. For each target expert, the differences between the comparable reliability core indicators of the target expert and the comparable reliability core indicators of all other experts are calculated. All the differences are summed to obtain the total lead margin of the target expert, or the maximum value among all the differences is taken as the total lead margin of the target expert. The total lead margin is then input into the Sigmoid function for nonlinear mapping to obtain the competitive support of the target expert. The steepness parameter of the Sigmoid function is used to control the sensitivity of the lead margin to the support. The basic reliability score of each expert is weighted and combined with the competitive support score, or a product combination is performed, to construct the candidate evaluation value of each expert. The candidate evaluation values ​​of all experts are compared, and the expert with the highest evaluation value is selected as the candidate expert.

5. The method according to claim 4, characterized in that, Based on the comparable reliability core indicators and the relative relationships among experts, candidate experts are determined from among the experts, including: The multiple experts include at least a first single expert based on a camera, a second single expert based on lidar, and a fusion expert. During the candidate evaluation process, the expert with better comparable core reliability indicators is determined from the first single expert and the second single expert as the optimal single expert. Calculate the difference between the comparable reliability core index of the fusion expert and the comparable reliability core index of the optimal single expert, and use the difference as the reliability gain of the fusion expert. Configure a preset gain threshold, compare the reliability gain with the preset gain threshold, if the reliability gain is lower than the preset gain threshold, apply an inhibition coefficient to the competitive support of the fusion expert to reduce it, if the reliability gain is not lower than the preset gain threshold, keep the competitive support of the fusion expert unchanged, the value of the inhibition coefficient is positively correlated with the degree to which the reliability gain is lower than the preset gain threshold or is a fixed attenuation value; Based on the basic reliability score and suppressed competitive support of the fusion experts, a candidate evaluation value for the fusion experts is constructed, making it difficult for the fusion experts to be selected as candidate experts when they only have a weak advantage.

6. The method according to claim 5, characterized in that, Obtain environmental condition information of the current data, and execute a selective decision-making process based on the candidate experts and the environmental condition information, including: The image blur information, image brightness information, and point cloud density information of the current data are extracted. After normalizing each information, an environmental degradation weight is assigned to each of them. The product of all information and its corresponding environmental degradation weight is weighted and summed to obtain the environmental degradation intensity index. Configure a degradation judgment threshold, compare the environmental degradation intensity index with the degradation judgment threshold, if the environmental degradation intensity index is lower than the degradation judgment threshold, then determine that the current environmental conditions are in a non-degraded state, directly use the candidate expert as the final output, and do not perform any rejection operation; If the environmental degradation intensity index is not lower than the degradation judgment threshold, the current environmental conditions are determined to be in a degraded state. Further, it is determined whether the candidate expert is a fusion expert. If the candidate expert is a single expert, the single expert is directly used as the final output. If the candidate expert is a fusion expert, the execution conditions of the fusion expert are reviewed.

7. The method according to claim 6, characterized in that, Triggering an execution condition review of the fusion expert includes: The expert with better comparable reliability core indicators between the first single expert and the second single expert is identified as the optimal single expert, and the comparable reliability core indicators of the optimal single expert are obtained. Calculate the difference between the comparable reliability core index of the fusion expert and the comparable reliability core index of the optimal single expert, and multiply the difference by a preset fusion coefficient to obtain the fusion execution margin. The fusion coefficient is a value between 0 and 1, which is used to adjust the magnitude of the conversion of reliability difference into execution margin. Configure a preset review threshold, compare the fusion execution margin with the preset review threshold, and if the fusion execution margin is greater than or equal to the preset review threshold, determine that the fusion expert meets the execution conditions and retain the fusion expert as the final output; If the fusion execution margin is less than the preset review threshold, the fusion expert is determined not to meet the execution conditions, the fusion expert is rejected, and the process reverts to the optimal single expert as the final output.

8. The method according to claim 1, characterized in that, After rollback, based on the reliability difference between the specific expert and the other experts, the specific expert is selectively restored as the final output, including: After the fusion expert is rejected due to failure to meet the execution conditions and the process reverts to the optimal single expert, the candidate expert identification information output by the candidate selection stage is obtained to confirm whether the candidate expert output by the candidate selection stage is the fusion expert. It is confirmed that the fusion expert has been rejected and a rollback operation has been performed, and the final output has been switched to the optimal single expert; Calculate the difference between the comparable reliability core index of the fusion expert and the comparable reliability core index of the optimal single expert, and use the difference as the advantage value of the fusion expert relative to the optimal single expert; Configure a preset protection threshold, compare the advantage value with the preset protection threshold, and if the advantage value is greater than or equal to the preset protection threshold, it is determined that the fusion protection condition is met, the rollback operation is canceled, and the final output is restored from the optimal single expert to the fusion expert. If the advantage value is less than the preset protection threshold, it is determined that the fusion protection condition is not met, and the optimal single expert is retained as the final output.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images