Automatic driving target detection reliability enhancement method and system based on evidence conflict analysis

Abstract

The invention discloses an automatic driving target detection reliability enhancement method and system based on evidence conflict analysis, and belongs to the technical field of automatic driving decision making. The problem of subsequent decision errors caused by the fact that a traditional automatic driving target detection reliability enhancement method does not consider high-risk output such as high-confidence false detection in the prior art is solved. According to the method, an evidence theory framework is introduced into the field of target detection, and semantic contradictions in the model are revealed by calculating conflict degree indexes between detection results; a conflict analysis and safety quality inspection module is constructed and serves as an external post-processing unit to be embedded into any depth detection model; and combining a conflict degree index with a risk threshold value, establishing a grading early warning system, identifying and marking a potential high-risk detection result in advance, automatically triggering a safety prompt when a high-conflict-degree sample is found, and providing a protection signal for an upper-layer decision-making module. The overall robustness and safety guarantee capability of the sensing system are improved, and the method can be applied to automatic driving decision making.

Description

TECHNICAL FIELD

[0001] The present application relates to an automatic driving target detection reliability enhancement method and system, in particular to an automatic driving target detection reliability enhancement method and system based on evidence conflict analysis, and belongs to the technical field of automatic driving decision. BACKGROUND

[0002] The rapid development of automatic driving technology makes the perception system the most critical component in the intelligent decision chain of the whole vehicle. The main task of the perception module is to detect and identify targets such as pedestrians, vehicles, traffic signs, etc. in the environment, and to provide reliable environmental information for path planning and control systems. However, the current mainstream target detection model generally relies on deep neural networks (DNN), which has achieved excellent accuracy on public datasets, but its reliability and safety in real complex environments still face serious challenges; especially in safety-critical scenarios, the detector often has the problem of high-confidence false positives (HCFP), i.e. the model outputs false results with very high confidence. This error is extremely deceptive because the downstream decision module often considers confidence as a reliability indicator, thus performing unsafe operations under the driving of false information.

[0003] In recent years, a large number of studies have been devoted to improving the reliability of deep models in uncertain environments, among which uncertainty quantification methods based on probability modeling have become a research hotspot. Traditional uncertainty evaluation relies on the confidence or predictive entropy of the model output to measure the stability of the prediction, but these indicators only reflect the "centralization degree" of the output distribution and cannot reveal the source of internal contradictions. Experiments have shown that under complex conditions such as occlusion, light changes or foreign interference, the confidence Softmax often remains high, thus masking potential recognition errors, which indicates that simple probability distribution-based metrics cannot meet the high safety requirements of the automatic driving scenario. To address this problem, Bayesian deep learning provides another approach, which introduces random variables in the network weights or output layer. The model can theoretically estimate the uncertainty of the prediction and reflect the prediction reliability through variance or confidence interval. Representative methods include Monte Carlo Dropout (MC Dropout) and Deep Ensemble. However, this type of method generally requires multiple forward propagations or model retraining, which has high computational overhead and poor real-time performance, making it unsuitable for deployment in vehicle-mounted or embedded systems. In addition, Bayesian deep learning-based methods focus on the confidence interval of the overall prediction rather than the internal semantic conflict of the recognition model, so there is still a significant deficiency in recognizing high-confidence false positives.

[0004] In recent years, evidence theory (also known as Dempster-Shafer theory) has been gradually introduced into the field of deep learning, used to measure the "belief strength" and "evidence conflict" of the model beyond probability; Evidence theory is different from the traditional probability framework, it can quantify the uncertainty of model prediction through the support and conflict of multiple hypotheses; This framework can more intuitively explain the internal decision-making mechanism of the model when facing complex inputs, so it performs well in anomaly detection, adversarial sample identification and open set identification tasks; Currently, methods that apply evidence theory to deep learning models for reliability quantification have received widespread attention, such frameworks assign independent evidence to each class and calculate the support and conflict between evidences, achieving fine-grained measurement of model prediction uncertainty; However, existing work mainly focuses on tasks such as image classification, and has not fully explored the extension and adaptation of this theory in more complex and challenging object detectors, especially in solving high-confidence false detection problems, there is still a research gap.

[0005] The object detection task is more complex than the classification task, not only needs to judge the target class, but also needs to locate the position of the target in space, the modern architecture represented by DINO, DETR and other Transformer-based detectors significantly improves the detection accuracy and inference speed through one-to-one set prediction mechanism; But the detector still relies on a single confidence threshold to filter the detection results, lacking an effective internal uncertainty calibration mechanism; When the detector faces unseen targets or blurred feature inputs, it often outputs high-confidence but incorrect prediction results, directly threatening the safety of the autonomous driving system; Therefore, introducing a risk review mechanism independent of the model structure in the detection stage, based on conflict degree real-time evaluation of model output, has become a key requirement to ensure system safety.

[0006] From an engineering perspective, the robustness of the autonomous driving perception system not only lies in the detection accuracy, but also in the system's awareness of its own reliability. The design principle of high-reliability systems requires that the perception module should have self-monitoring and risk warning mechanisms, so that it can alert the downstream system when it detects abnormalities; However, existing detectors almost lack this ability, their internal reasoning process is highly black-boxed and cannot quantify the inconsistency between internal evidence; The introduction of a detection mechanism based on evidence conflict analysis not only makes up for the limitations of traditional confidence evaluation, but also provides interpretable risk signals for subsequent decision-making modules, providing key support for building a more reliable 'perception-decision' information channel.

[0007] Therefore, the prior art has the following deficiencies: (1) the current mainstream detector generally uses Softmax confidence as the basis for reliability, but this index only reflects the concentration degree of the output probability and cannot reveal the potential conflict between different classes inside the model. When the model produces high responses to multiple similar targets (such as "pedestrians" and "riders") at the same time, the confidence may still remain high, causing the system to ignore the potential risks by mistaking the result as reliable. This distorted expression of "uncertainty" makes it difficult to identify high-confidence false detections in a timely manner, which is an implicit threat to the safety of autonomous driving; (2) common uncertainty estimation methods (such as Monte Carlo Dropout and deep model integration) can theoretically quantify the prediction confidence distribution, but they rely on multiple forward propagations or model retraining, which has extremely high computational complexity. For real-time systems such as autonomous driving that require millisecond-level responses, this method significantly increases the inference delay and computational energy consumption, making it difficult to be used on vehicle embedded platforms, thus limiting its practical application value in high-reliability real-time perception scenarios: (3) existing researches mostly focus on general uncertainty estimation, ignoring the special risk type of high-confidence false detection, which is "confident but wrong". Traditional detectors lack quantitative means for semantic conflicts inside the model and do not establish an independent safety review channel for high-confidence false predictions, causing the system to still output seemingly reliable but essentially incorrect results when facing rare scenarios or ambiguous feature inputs, thereby introducing serious safety hazards to the downstream decision planning module.

[0008] In summary, although existing target detection research has made significant progress in accuracy and generalization, there is still a significant gap in self-evaluation of model reliability. There is currently a lack of an uncertainty estimation method that can balance explainability, real-time performance, and universality to identify high-confidence false detections and other high-risk outputs. Therefore, it is necessary to propose an automatic driving target detection reliability enhancement method and system based on evidence conflict analysis, which quantifies the internal evidence conflict of the model to establish a lightweight "reliability quality inspection layer" and provide an efficient and explainable risk detection solution for autonomous driving perception systems. SUMMARY

[0009] A brief summary of the application is given below in order to provide a basic understanding of some aspects of the application. It should be understood that this summary is not an exhaustive overview of the application. It is not intended to identify key or important parts of the application nor is it intended to limit the scope of the application. Its purpose is merely to present some concepts in a simplified form as a prelude to the more detailed description that follows.

[0010] In view of this, in order to solve the problem that the traditional autonomous driving target detection reliability enhancement methods in the prior art do not consider the subsequent decision-making errors caused by high-risk outputs such as high-confidence false detections, the present invention provides an autonomous driving target detection reliability enhancement method and system based on evidence conflict analysis.

[0011] Technical Solution 1 is as follows: A method for enhancing the reliability of autonomous driving target detection based on evidence conflict analysis, comprising the following steps:

[0012] S1. Input the preprocessed image data into the target detection and original evidence extraction module to obtain the feature map. Based on the cross-attention mechanism, call the target detector to perform forward inference and extract the original evidence.

[0013] S2. The original evidence is transformed into an evidence vector through the conflict quantification module and mapped to the parameters of the Dirichlet distribution to calculate the degree of evidence conflict.

[0014] S3. Based on the degree of evidence conflict, the conflict quantification process is integrated into the inference pipeline of the target detector through the risk assessment and result integration module. Combined with the dual standard decision rule, the detection results filtered through the confidence threshold are finally risk-classified and an early warning signal is output.

[0015] S4. The warning signal is directly input into the decision planning system of autonomous driving through the downstream application and decision support module. The driving state marked as high risk target is transformed into a conservative state. The false detection is adjusted by the fusion weight to obtain the final fusion confidence and reject the decision result of high risk and high confidence false detection.

[0016] Furthermore, in step S1, the model forward inference submodule in the target detection and original evidence extraction module processes the input data and executes the core detection algorithm. The specific steps are as follows:

[0017] The input data is a preprocessed RGB image, represented as a three-dimensional tensor. Where H, W, and 3 are the height, width, and depth of the RGB image, respectively. The deep visual features of the preprocessed RGB image are extracted through the backbone network of the convolutional neural network to obtain a feature map Frame containing rich spatial and semantic information.

[0018] The feature map Fimg, which contains rich spatial and semantic information, is represented as follows:

[0019]

[0020] in, ,width , The height and width of the feature map are represented by D, and the number of feature channels is represented by D.

[0021] The feature map Fimg is input into the Transformer encoder for self-attention computation, generating the enhanced feature sequence. The decoder uses a set of N learnable objects to query data through multiple iterations. Image features output by the encoder The system interacts with the target, gradually locating and identifying it. Each decoder layer uses a cross-attention mechanism for model inference, ensuring that each object query contains precise information about the specific target.

[0022] The cross-attention mechanism is represented as:

[0023]

[0024] Where Attention represents attention, Q represents the query vector matrix, K represents the key vector matrix, and V represents the value vector matrix. This represents the transpose of matrix K. The dimension of the key vector is represented by , and softmax is the function that normalizes the attention score.

[0025] After the model inference is completed, the original evidence is extracted using the original evidence location and extraction submodule. The specific steps are as follows:

[0026] For each of the N objects, retrieve the classification prediction head from the (L-1)th layer decoder. Extract its corresponding Logits vector, i.e., the original evidence. ,in, ;

[0027] Original evidence Represented as:

[0028]

[0029] in, For the first The state vector of an object query after optimization by the (L-1)th layer decoder. For the classification prediction head of the L-1 layer decoder, the original evidence It is a C-dimensional real vector. C represents the total number of predefined target categories. Each element in the Logits vector represents an element determined by the model. The corresponding target belongs to category c of the original evidence. =1, 2, ..., C.

[0030] Furthermore, in step S2, the original evidence is converted through the evidence conversion submodule. This is transformed into a non-negative quantity of evidence that conforms to the form of evidence theory, i.e., an evidence vector. ;

[0031] Vector of Evidence Represented as:

[0032]

[0033] in, For the original evidence The reference value for the centered translation is s, which is a positive hyperparameter scaling factor. It is a non-linear activation function;

[0034] The evidence vector, after being transformed by the evidence transformation submodule, is generated through the Dirichlet parameter generation submodule. Parameters mapped to the Dirichlet distribution In the context of evidence theory, the parameters of the Dirichlet distribution That is, the evidence count for each category;

[0035] Parameters of the Dirichlet distribution Represented as:

[0036]

[0037] Among them, by ,get , Evidence representing category c corresponding to element i Let i represent the evidence vector corresponding to category c of element i. Here, the parameters of the Dirichlet distribution are... As a complete distribution of evidence, we obtain the first... Total amount of evidence for each object query , , Indicates excluding categories The sum of evidence for all other categories, Indicates excluding categories The sum of evidence corresponding to all other categories;

[0038] Using the generated Dirichlet parameters The degree of conflict of evidence is calculated through the evidence conflict calculation submodule. ;

[0039] Conflict of Evidence Represented as:

[0040]

[0041] in, Indicates the first Querying one object. The total number of predefined target categories, and For category indexes, satisfying j,k∈ and , and They represent the first A query for an object in the category and categories The Dirichlet distribution parameters on the surface.

[0042] Furthermore, in S3, the optimal timing for intervention in the target detection process is defined by the comprehensive risk review submodule. The optimal timing is designed after the target detector outputs the original prediction results of all N candidate queries, but before performing the final confidence threshold filtering.

[0043] The risk score integration submodule integrates the degree of evidence conflict. The original detection results are bound to the data structure; that is, for each candidate element i, its original detection results are bound to the data structure. Includes bounding box Confidence level and prediction categories , the degree of conflict of evidence As a new key attribute, it is added to the original detection result. This results in an enhanced detection outcome. ;

[0044] Enhanced detection results Represented as:

[0045] = { , , , }

[0046] According to the dual-standard decision-making rule, the high-risk target early warning submodule performs a final risk classification on all detection results filtered through the confidence threshold. Detection results that simultaneously meet the following two conditions are judged as high-risk, high-confidence false detections: 1) The confidence level of the current detection result. Higher than the preset decision threshold ;2) Degree of conflict of evidence in the current test results Risk alarm threshold higher than preset That is, early warning logic ,in, This represents the logical AND operation, and is used for alert logic. If true, the current target triggers a high-risk warning and outputs a warning signal; otherwise, no high-risk warning is issued.

[0047] Furthermore, in step S4, the warning signal is directly input into the autonomous driving decision-making and planning system through the safety policy adjustment submodule, and the driving state of the system is defined as... When the perceived result One of them is marked as high risk When the target is reached, the decision-making system immediately switches to a conservative state. ;

[0048] The cross-validation process among multiple sensors is initiated through the sensor redundancy verification submodule. When the main sensor outputs a high degree of evidence conflict... When detecting a target, its 3D spatial location is treated as a high-priority verification request and distributed to redundant sensors. The redundant sensors will perform a matching search within their corresponding spatial ROI. If other sensors fail to find any physical entity within the current spatial ROI, or if the attributes of the found entity are significantly inconsistent with the main sensor's detection result, the confidence value for classifying the main sensor's detection result as a false detection is increased; that is, by fusing weights... False positive adjustment is performed to obtain the final fusion confidence level. Reject high-risk, high-confidence false positive decision results;

[0049] Final fusion confidence Represented as:

[0050] = * .

[0051] Technical Solution 2 is as follows: an autonomous driving target detection reliability enhancement system based on evidence conflict analysis, used to execute the autonomous driving target detection reliability enhancement method based on evidence conflict analysis described in Technical Solution 1, including a target detection and original evidence extraction module, a conflict quantification module, a risk assessment and result integration module, and a downstream application and decision support module connected in sequence.

[0052] The target detection and original evidence extraction module includes a model forward inference submodule and an original evidence localization and extraction submodule connected in sequence.

[0053] The conflict quantification module includes an evidence transformation submodule, a Dirichlet parameter generation submodule, and an evidence conflict degree calculation submodule connected in sequence.

[0054] The risk assessment and result integration module includes a comprehensive risk review submodule, a risk score integration submodule, and a high-risk target early warning submodule connected in sequence.

[0055] The downstream application and decision support module includes a security policy adjustment submodule and a sensor redundancy verification submodule connected in sequence.

[0056] The beneficial effects of this invention are as follows: This invention aims to address several shortcomings in the reliability assessment of existing autonomous driving target detection systems by proposing a high-confidence false detection identification method based on evidence conflict analysis. Existing detection models often over-rely on the confidence level (Softmax) as a single reliability indicator, failing to truly reflect the semantic contradictions and uncertainties within the model, thus producing deceptive "confident but erroneous" outputs in critical scenarios. Therefore, this invention introduces an evidence theory framework, treating the output of deep detection models as multi-source evidence for fusion and conflict analysis. By quantifying the inherent contradictions between different categories, a novel risk metric, "conflict degree," independent of confidence level, is defined for use in... This invention identifies potentially high-risk prediction results. The core idea is to leverage the characteristics of evidence conflict to reveal logical inconsistencies within the model, thereby enabling self-checking of the model's output reliability without modifying the detector structure or retraining. By embedding a lightweight conflict degree calculation and risk assessment module into the detection process, the system can detect abnormal predictions in real time during the inference stage, issuing early risk warnings and providing more reliable perceptual input for subsequent decision-making and control modules. This invention is simple in structure, computationally efficient, and widely applicable to various deep target detection models. It has significant application and promotion value in fields with stringent reliability requirements, such as autonomous driving, drone monitoring, and industrial vision inspection.

[0057] This invention introduces an evidence theory framework into the field of target detection. By calculating the conflict degree index between detection results, it reveals semantic contradictions within the model, thereby identifying high-confidence false detections. This mechanism can quantify the inherent inconsistencies in prediction results during the model inference stage, providing a more reliable basis for safety assessment of the perception system. This invention constructs a conflict analysis and safety quality inspection module independent of the detector structure. It can be embedded as an external post-processing unit into any deep detection model. This module can run without model retraining, has extremely low computational overhead, and possesses real-time performance and versatility, making it suitable for deployment in high-time-sensitive scenarios such as autonomous driving and industrial inspection. This invention combines the conflict degree index with risk thresholds to establish a hierarchical early warning system, which can identify and label potentially high-risk detection results in advance. When the system detects high-conflict-degree samples, it automatically triggers a safety alert, providing a protective signal to the upper-level decision-making module, thereby significantly improving the overall robustness and safety assurance capabilities of the perception system.

[0058] Compared to traditional methods that rely on confidence softmax or prediction entropy, this invention quantifies the internal conflict relationships of the model through an evidence theory framework, which can more comprehensively reflect the sources of uncertainty in the detection results. The evidence conflict index shows a significant advantage in identifying high-confidence false detections, effectively distinguishing between true positive detections and high-risk false detections, thereby significantly improving the system's perception reliability and security for complex scenarios and abnormal inputs. The conflict analysis module of this invention is lightweight and independent, independent of model structure or training process, and can be directly embedded into various autonomous driving detection systems. Its computational complexity is far lower than Monte Carlo Dropout or deep ensemble methods, making it suitable for scenarios with extremely high real-time requirements such as autonomous driving, UAV vision, and industrial quality inspection. This makes the invention easier to implement in engineering applications and has significant promotion and integration value. Attached Figure Description

[0059] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0060] Figure 1 This is a flowchart illustrating a method for enhancing the reliability of autonomous driving target detection based on evidence conflict analysis.

[0061] Figure 2 This is a schematic diagram of an embodiment of a method for enhancing the reliability of target detection in autonomous driving based on evidence conflict analysis;

[0062] Figure 3 This is a schematic diagram of the structure of an autonomous driving target detection reliability enhancement system based on evidence conflict analysis.

[0063] Figure descriptions: 1. Target detection and original evidence extraction module; 2. Conflict quantification module; 3. Risk assessment and result integration module; 4. Downstream application and decision support module. Detailed Implementation

[0064] To make the technical solutions and advantages of the embodiments of the present invention clearer, the exemplary embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0065] Example 1: Reference Figures 1-2 This embodiment describes a method for enhancing the reliability of autonomous driving target detection based on evidence conflict analysis, specifically including the following steps:

[0066] S1. Input the preprocessed image data into the target detection and original evidence extraction module to obtain the feature map. Based on the cross-attention mechanism, call the target detector to perform forward inference and extract the original evidence.

[0067] S2. The original evidence is transformed into an evidence vector through the conflict quantification module and mapped to the parameters of the Dirichlet distribution to calculate the degree of evidence conflict.

[0068] S3. Based on the degree of evidence conflict, the conflict quantification process is integrated into the inference pipeline of the target detector through the risk assessment and result integration module. Combined with the dual standard decision rule, the detection results filtered through the confidence threshold are finally risk-classified and an early warning signal is output.

[0069] S4. The warning signal is directly input into the decision planning system of autonomous driving through the downstream application and decision support module. The driving state marked as high risk target is transformed into a conservative state. The false detection is adjusted by the fusion weight to obtain the final fusion confidence and reject the decision result of high risk and high confidence false detection.

[0070] Specifically, this invention, as a lightweight post-processing module, performs in-depth uncertainty analysis on the raw output of advanced target detectors. By accurately quantifying the risk of high-confidence false detections (HCFP), it provides a confidence-independent "safety warning signal" for safety-critical systems such as autonomous driving.

[0071] High Confidence False Detection (HCFP): This refers to the phenomenon where a target detection model outputs erroneous detection results with extremely high confidence during the recognition process. This type of false detection is the most dangerous because its high confidence often misleads downstream decision-making or control systems, leading to serious security risks or system failures.

[0072] Evidence Theory: A mathematical framework for dealing with uncertain information, which can fuse and analyze conflicts of evidence from different sources, thereby quantifying the credibility of the model's judgment on events. It is widely used in the reliability analysis of artificial intelligence systems.

[0073] Conflict: In evidence theory, this is an indicator used to describe the degree of contradiction between different pieces of evidence. The higher the conflict value, the stronger the semantic inconsistency or reasoning contradiction within the model. It can be used as an early warning signal for high-risk detection results.

[0074] Model calibration refers to the ability to measure the consistency between a model's prediction confidence and its actual prediction accuracy. A perfectly calibrated model should have an output confidence score (e.g., 80%) that accurately corresponds to its actual probability of success (i.e., 80% probability of being correct). The lack of model calibration is the root cause of neural networks being "overconfident" or "underconfident" and is also a key step in evaluating the credibility of their output.

[0075] When this invention is implemented in the following three traffic situations, the following effects can be achieved respectively: 1) At urban intersections, the target detection model simultaneously identifies the two categories of "pedestrian" and "cyclist" with high confidence. The conflict analysis module detects that the two are highly overlapping in spatial location and semantics during the inference stage, calculates a significant conflict value, and the system marks the detection result as high risk and sends deceleration and review instructions to the upper-level control module to avoid unsafe behavior caused by misidentification.

[0076] 2) On highways with insufficient lighting, the model mistakenly identified a suspended road sign as a "traffic light," with a confidence level exceeding 0.8. The conflict analysis module found that this prediction strongly contradicted other visual features, and the conflict level reached the upper limit of the risk threshold. The system automatically triggered a safety warning, temporarily classifying the target as an untrusted area and prompting the perception module to re-detect it, significantly reducing the risk of misjudgment at night.

[0077] 3) In the construction section, some vehicles have degraded visual features due to obstruction or dust. The detector misidentifies the construction vehicle as a "bus" with high confidence. The conflict analysis module finds obvious conflicts between the predictive evidence through multi-layer semantic feature comparison, automatically issues a high-risk signal, and the system enters a conservative driving mode and updates the perception cache to ensure safe and reliable operation under low-quality input conditions.

[0078] Furthermore, in step S1, the model forward inference submodule in the target detection and original evidence extraction module processes the input data and executes the core detection algorithm. The specific steps are as follows:

[0079] The input data is a preprocessed (e.g., normalized, resized) RGB image, represented as a three-dimensional tensor. Where H, W, and 3 are the height, width, and depth of the RGB image, respectively. The deep visual features of the preprocessed (such as normalization and resizing) RGB image are extracted through the backbone network of the convolutional neural network (CNN) (such as ResNet or Swin Transformer) to obtain a feature map Frame containing rich spatial and semantic information.

[0080] The feature map Fimg, which contains rich spatial and semantic information, is represented as follows:

[0081]

[0082] in, , , D represents the height, width, and depth of the feature map Fimg, respectively.

[0083] The feature map Fimg is input into the Transformer encoder for self-attention computation to capture global contextual dependencies and generate an enhanced feature sequence. The encoder-decoder architecture is the core of the model adapted by this invention, namely the object detector. The decoder uses a set of N learnable object queries through multi-level iteration. Image features output by the encoder The system interacts with the target, gradually locating and identifying it. Each decoder layer uses a cross-attention mechanism for model inference, ensuring that each object query contains precise information about the specific target.

[0084] The cross-attention mechanism is represented as:

[0085]

[0086] Where Attention represents attention, Q represents query, K represents key, and V represents value. Query by object K was obtained. Through feature sequences We get that softmax represents the normalization function. This represents the transpose of matrix K. Indicates the dimension of the key vector;

[0087] After the model inference is completed, the original evidence is extracted using the original evidence location and extraction submodule. The specific steps are as follows:

[0088] Each element in the classification prediction head (a feedforward neural network FFN) of the L-1 layer decoder will be queried for all N objects. Extract its corresponding Logits vector, i.e., the original evidence. ,in, ;

[0089] Original evidence Represented as:

[0090]

[0091] in, For the first The state vector of an object query after optimization by the (L-1)th layer decoder. For the classification prediction head of the L-1 layer decoder, the original evidence It is a C-dimensional real vector. C represents the total number of predefined target categories. Each element in the Logits vector represents an element determined by the model. The corresponding target belongs to category c of the original evidence. =1, 2, ..., C, original evidence It will serve as the core input for the next module.

[0092] Specifically, the object detection and raw evidence extraction module is responsible for receiving preprocessed image data, calling a pre-trained, Transformer-based advanced object detector to perform complete forward inference, and accurately extracting raw information that reflects potential contradictions between categories from a specific intermediate layer within the model, providing high-quality input for subsequent conflict quantification.

[0093] The cross-attention mechanism allows each object query to focus on the most relevant region in the image and continuously update its own state based on this. After iterative optimization by the L-layer decoder, each object query contains precise information about the specific target.

[0094] The original evidence localization and extraction submodule aims to find and extract the most ideal "original evidence" from the above reasoning process. The key insight of this invention is that the final decoding layer (layer L) of the model often undergoes internal "sharpening" processing in order to pursue the highest classification accuracy, tending to form a "winner-takes-all" situation, thereby masking the potential conflicts between different categories of evidence. In contrast, the intermediate decoding layers, especially the penultimate layer (layer L-1), output query states that contain sufficiently rich semantic information for classification and have not yet undergone final decision solidification, which can serve as an excellent source for extracting the "internal contradictions" of the model.

[0095] Furthermore, in step S2, the original evidence is converted through the evidence conversion submodule. This is transformed into a non-negative quantity of evidence that conforms to the form of evidence theory, i.e., an evidence vector. To ensure the effectiveness and robustness of the transformation, the evidence transformation submodule is designed with a lightweight processing flow that includes three steps: translation, scaling and activation.

[0096] Vector of Evidence Represented as:

[0097]

[0098] in, For the original evidence The baseline value for centering and shifting is usually the mean value. To eliminate the influence of the overall offset of the Logits vector, s is a positive hyperparameter scaling factor used to adjust the overall strength of the evidence or the "concentration of belief." According to ablation experiments, when s=10, the model performs optimally on the task of identifying high-confidence false detections. It is a non-linear activation function;

[0099] The evidence vector, after being transformed by the evidence transformation submodule, is generated through the Dirichlet parameter generation submodule. Parameters mapped to the Multinomial-Dirichlet distribution In the context of evidence theory, the parameters of the Dirichlet distribution That is, the evidence count for each category;

[0100] Parameters of the Dirichlet distribution Represented as:

[0101]

[0102] Among them, by ,get , Evidence representing category c corresponding to element i This represents the evidence vector for category c corresponding to element i. The increment operation ensures that all parameters are updated. All values ​​are positive, which is a necessary condition for the definition of the Dirichlet distribution. In this case, the parameters of the Dirichlet distribution... As a complete distribution of evidence, we obtain the first... Total amount of evidence for each object query (Also known as Dirichlet strength). , Indicates excluding categories The sum of evidence for all other categories, Indicates excluding categories The sum of evidence corresponding to all other categories;

[0103] Using the generated Dirichlet parameters The core risk indicator of this invention—the degree of conflict of evidence—is calculated through the evidence conflict calculation submodule. This metric originates from the classic measure in Dempster-Shafer theory used to measure inconsistency between two or more pieces of evidence. It applies to evidence generated by the same observation (i.e., the same element i) but pointing to different categories c. When multiple categories of evidence When both are very large, it means that the model has received strong but contradictory signals about the identity of the target, and the degree of conflict should be very high at this time;

[0104] Conflict of Evidence Represented as:

[0105]

[0106] in, Indicates the first Querying one object. The total number of predefined target categories, and For category indexing, degree of evidence conflict The evidence conflict between any two different categories j and k is precisely quantified, and all conflict pairs are summed to satisfy j,k∈ and , and They represent the first A query for an object in the category and categories The Dirichlet distribution parameters on the surface act as a normalization factor, affecting the degree of conflict of evidence. The value of is a scalar, independent of traditional confidence levels, specifically used to measure the degree of "conflict" or "contradiction" in the model's internal decisions, and is a strong signal for identifying HCFPs; the above definition of conflict level comprehensively characterizes the degree of conflict by summing the pairwise interaction terms between different categories of evidence. The degree of evidence conflict for each object query across mutually exclusive categories is used to quantify the level of evidence conflict in the detection result, which is then used as a measure of the risk level of the detection result.

[0107] Specifically, the conflict quantification module based on evidence theory is the core of the methodology of this invention. Its main task is to transform the original Logits vector extracted by the previous module into a quantitative indicator that can accurately measure the risk of false detection with high confidence—the degree of evidence conflict. Based on the framework of evidence theory, this module innovatively constructs a quantitative path from neural network output to scenario risk assessment.

[0108] The Softplus function is represented as:

[0109]

[0110] in, To input real-valued variables, in this invention, the Softplus function operates element-wise on the vector. Compared to the commonly used ReLU function, the Softplus function has the smooth property of being differentiable everywhere, which can better preserve the small negative values ​​in Logits (which may represent weak opposing evidence), thus more finely characterizing the complex relationships between evidence and providing richer input for subsequent conflict calculations.

[0111] Furthermore, in S3, the optimal timing for intervention in the target detection process is defined through the comprehensive risk review submodule. In order to achieve a thorough review of all potential risks, the optimal timing is designed after the target detector outputs the original prediction results of all N candidate queries, but before performing the final confidence threshold filtering. This crucial timing selection ensures that no potential HCFP candidate box will be discarded prematurely due to the initial confidence level not being met, thereby avoiding the risk omissions that may be caused by the "filter first, then analyze" approach in the traditional process.

[0112] The risk score integration submodule integrates the degree of evidence conflict. The original detection results are bound to the data structure; that is, for each candidate element i, its original detection results are bound to the data structure. Includes bounding box Confidence level and prediction categories , the degree of conflict of evidence As a new key attribute, it is added to the original detection result. This results in an enhanced detection outcome. ;

[0113] Enhanced detection results Represented as:

[0114] = { , , , }

[0115] According to the dual-standard decision-making rule, the high-risk target early warning submodule performs a final risk classification on all detection results filtered through the confidence threshold. Detection results that simultaneously meet the following two conditions are judged as high-risk, high-confidence false detections: 1) The confidence level of the current detection result. Higher than the preset decision threshold (In this embodiment, =0.7); 2) The degree of evidence conflict of the current test results Risk alarm threshold higher than preset , The warning logic can be determined by statistical analysis of the Conflict values ​​of a large amount of validation data (in this embodiment, the 90th percentile of the distribution is used). ,in, This represents the logical AND operation, and is used for alert logic. If true, the current target triggers a high-risk warning and outputs a warning signal; otherwise, no high-risk warning is issued.

[0116] Specifically, the conflict quantification process is integrated into the inference pipeline of the target detector through the risk assessment and results integration module. Without affecting the original functions, a reliable risk score is attached to each detection result to achieve "comprehensive risk review".

[0117] Furthermore, in step S4, the warning signal is directly input into the autonomous driving decision-making and planning system through the safety strategy adjustment submodule. This submodule is used to dynamically adjust the conservatism of the driving strategy, for example, defining the system's driving state as... When the perceived result One of them is marked as high risk When faced with a target, even if the target's classification and confidence level are traditionally credible, the decision system immediately shifts it to a conservative state. In a conservative state, the decision-making and planning system will proactively execute actions such as increasing following distance, reducing cruising speed, or temporarily prohibiting high-dynamic operations such as lane changing and overtaking, thereby reserving more safety redundancy to deal with potential perception errors.

[0118] Utilizing high degree of evidence conflict As an efficient triggering mechanism, the value initiates a cross-validation process among multiple sensors through the sensor redundancy verification submodule. When the main sensor (such as a camera) outputs a high degree of evidence conflict... When detecting a target value, its three-dimensional spatial position (through its corresponding bounding box) is determined. The data (obtained from depth estimation) is used as a high-priority verification request and distributed to redundant sensors (such as LiDAR and millimeter-wave radar). The redundant sensors will perform a matching search within their corresponding spatial ROIs. If other sensors fail to find any physical entities within the current spatial ROI, or if the attributes of the found entities (such as size and velocity) are significantly inconsistent with the detection results of the main sensor, the confidence level for classifying the main sensor's detection result as a false detection can be greatly increased. This is achieved through weight fusion. False positive adjustment is performed to obtain the final fusion confidence level. Reject high-risk, high-confidence false positive decision results;

[0119] Final fusion confidence Represented as:

[0120] = *

[0121] When sensor collisions occur A value close to 0 will be taken, thereby effectively "rejecting" this high-risk HCFP test at the fusion level and preventing it from affecting the final decision.

[0122] Specifically, the downstream application and decision support module, as the application layer of the technical solution of this invention, is responsible for consuming the output of the upstream module with evidence conflict level. The risk score detection results are transformed into specific decisions and actions that can improve vehicle driving safety. As a bridge connecting perceived uncertainty and vehicle planning and control, it enables the risk quantification capability of this invention to be truly implemented.

[0123] Example 2: Reference Figure 3 This embodiment is described in detail as follows: The autonomous driving target detection reliability enhancement system based on evidence conflict analysis is used to execute the autonomous driving target detection reliability enhancement method based on evidence conflict analysis described in Embodiment 1. It includes a target detection and original evidence extraction module 1, a conflict quantification module 2, a risk assessment and result integration module 3, and a downstream application and decision support module 4 connected in sequence.

[0124] The target detection and original evidence extraction module 1 includes a model forward inference submodule and an original evidence localization and extraction submodule connected in sequence.

[0125] The conflict quantification module 2 includes an evidence transformation submodule, a Dirichlet parameter generation submodule, and an evidence conflict degree calculation submodule connected in sequence.

[0126] The risk assessment and result integration module 3 includes a comprehensive risk review submodule, a risk score integration submodule, and a high-risk target early warning submodule connected in sequence.

[0127] The downstream application and decision support module 4 includes a security policy adjustment submodule and a sensor redundancy verification submodule connected in sequence.

[0128] Although the invention has been described with reference to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of interpreting or limiting the subject matter of the invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.

Claims

1. An automatic driving target detection reliability enhancement method based on evidence conflict analysis, characterized in that, The method comprises the following steps: S1. input the preprocessed image data into the target detection and original evidence extraction module to obtain a feature map, call a target detector to perform forward reasoning according to a cross attention mechanism, and extract original evidence; S2. convert the original evidence into an evidence vector through a conflict quantification module, and map the evidence vector into a parameter of a Dirichlet distribution to calculate an evidence conflict degree; S3. integrate the conflict quantification process into the reasoning pipeline of the target detector through a risk assessment and result integration module according to the evidence conflict degree, combine a double standard decision rule, perform final risk grading on the detection results filtered through a confidence threshold, and output a warning signal; S4. directly input the warning signal into the decision planning system of the autonomous driving through a downstream application and decision support module, convert the driving state in which a target is marked as a high-risk target into a conservative state, perform false detection adjustment through a fusion weight to obtain a final fusion confidence, and veto a high-risk high-confidence false detection decision result.

2. The automatic driving target detection reliability enhancement method based on evidence conflict analysis according to claim 1, characterized in that, In S1, the model forward reasoning submodule in the target detection and original evidence extraction module is used to process the input data and perform a core detection algorithm, and the specific steps are as follows: The input data is a preprocessed RGB image, which is represented as a three-dimensional tensor Wherein, H, W, 3 are the height, width and depth of the RGB image respectively, the deep visual features of the preprocessed RGB image are extracted by the backbone network Backbone of the convolutional neural network to obtain the feature map F img containing rich spatial and semantic information; The feature map F img containing rich spatial and semantic information is represented as: ； wherein, , , is the height and width of the feature map, and D is the number of feature channels. The feature map Fimg is input into the Transformer encoder for self-attention computation, generating the enhanced feature sequence. The decoder uses a set of N learnable objects to query data through multiple iterations. Image features output by the encoder The system interacts with the target, gradually locating and identifying it. Each decoder layer uses a cross-attention mechanism for model inference, ensuring that each object query contains precise information about the specific target. The cross attention mechanism is represented as: ； wherein Attention denotes attention, Q denotes a query vector matrix, K denotes a key vector matrix, and V denotes a value vector matrix, denotes a transpose of the matrix K, denotes a dimension of the key vector, and softmax denotes a function for normalizing the attention score; After the model reasoning is completed, the original evidence positioning and extraction submodule is used to extract the original evidence, and the specific steps are as follows: From the classification prediction head of the L-1 layer decoder Decoder, for each element of all N object queries extract its corresponding Logits vector, i.e., the original evidence where, ; Original evidence is represented as: ； wherein, is the L-th object query, is the state vector after the L-1-th layer decoder optimization, is the classification prediction head of the L-1-th layer decoder Decoder, the original evidence is a C-dimensional real number vector, C is the total number of predefined target categories, represents each element in the Logits vector, which represents the original evidence that the element corresponding to the target determined by the model belongs to category c, = 1, 2, …, C.

3. The automatic driving target detection reliability enhancement method based on evidence conflict analysis according to claim 2, characterized in that, In the S2, the original evidence is transformed into non-negative evidence quantity, i.e. evidence vector, conforming to the evidence theory form by the evidence transformation submodule ;​ evidence vector is represented as: ； wherein, is a reference value for the centralised translation of the original evidence s is a positive hyperparameter scaling factor, is a non-linear activation function; The evidence vector transformed by the evidence transformation submodule is transformed by a Dirichlet parameter generation submodule parameters of the Dirichlet distribution parameters of the Dirichlet distribution in the context of the evidence theory i.e. the evidence count corresponding to each class; Parameters of Dirichlet distribution is expressed as: ； whereby , we obtain , represents the evidence of the class c corresponding to the element i, represents the evidence vector of the class c corresponding to the element i, whereby the parameters of the Dirichlet distribution as a complete evidence distribution, we obtain the total evidence amount for the , , represents the sum of the evidence of all classes except the class , represents the sum of the evidence of all classes except the class , Utilizing the generated Dirichlet parameters The evidence conflict degree is calculated by an evidence conflict degree calculation submodule ; Evidence conflict is represented as: ； where, denotes the th object query, is the total number of predefined target categories, and is the category index satisfying j, k ∈ and , and denote the Dirichlet distribution parameters of the th object query on category and category , respectively.

4. The evidence conflict analysis based autonomous driving target detection reliability enhancement method of claim 3, wherein, In S3, a comprehensive risk review submodule is used to define the best time to intervene in the target detection process, and the best time is designed after the target detector outputs all N original prediction results of candidate queries but before performing final confidence threshold filtering; The risk score integration submodule integrates the degree of evidence conflict. The original detection results are bound to the data structure; that is, for each candidate element i, its original detection results are bound to the data structure. Includes bounding box Confidence level and prediction categories , the degree of conflict of evidence As a new key attribute, it is added to the original detection result. This results in an enhanced detection outcome. ; Enhanced detection results is represented as: ={ , , , }； According to the double standard decision rule, all detection results filtered through the confidence threshold are finally risk rated by the high-risk target early warning sub-module, and the detection results that meet the following two conditions are judged as high-risk high-confidence false detection: 1) the confidence of the current detection result is higher than the preset decision threshold ;​ 2) the evidence conflict degree of the current detection result higher than the preset risk warning threshold i.e. the early warning logic wherein, represents a logical AND operation, when the early warning logic is true, the current target triggers a high-risk early warning and outputs an early warning signal, otherwise no high-risk early warning is performed.

5. The automatic driving target detection reliability enhancement method based on evidence conflict analysis according to claim 4, characterized in that, In the S4, the early warning signal is directly input into the decision planning system of the automatic driving through the security policy adjustment submodule, and the driving state of the system is defined as When the perception result contains a target marked as high risk , the decision system immediately converts it to a conservative state ; The cross-validation process between multiple sensors is started by the sensor redundancy verification sub-module, when the main sensor outputs a high evidence conflict value When the detection target is a high-priority verification request, its three-dimensional spatial position is distributed to the redundant sensor as a high-priority verification request, and the redundant sensor will perform a matching search within the corresponding spatial ROI. If other sensors fail to find any physical entity within the current spatial ROI, or the properties of the found entity are significantly inconsistent with the detection result of the main sensor, the confidence value of the judgment that the detection result of the main sensor is a false detection is increased, i.e., the fusion weight The final fusion confidence is obtained by false detection adjustment , and the high-risk high-confidence false detection decision result is vetoed. Final fusion confidence is represented as: = * 。 6. An autonomous driving target detection reliability enhancement system based on evidence conflict analysis, characterized in that, The method for enhancing the reliability of autonomous driving target detection based on evidence conflict analysis comprises a target detection and original evidence extraction module (1), a conflict quantification module (2), a risk assessment and result integration module (3), and a downstream application and decision support module (4) connected in sequence; The target detection and original evidence extraction module (1) comprises a model forward reasoning submodule and an original evidence positioning and extraction submodule connected in sequence; The conflict quantification module (2) comprises an evidence conversion submodule, a Dirichlet parameter generation submodule, and an evidence conflict degree calculation submodule connected in sequence; The risk assessment and result integration module (3) comprises a comprehensive risk review submodule, a risk score integration submodule, and a high-risk target warning submodule connected in sequence; The downstream application and decision support module (4) comprises a safety strategy adjustment submodule and a sensor redundancy verification submodule connected in sequence.

Images