Automatic driving-oriented Monte Carlo uncertainty quantization target detection method

Abstract

The invention relates to a Monte Carlo uncertainty quantization target detection method for automatic driving, and belongs to the field of complex scene target detection algorithms, and the method comprises the following steps: constructing a Transform framework-oriented multivariate determinant point process Monte Carlo DDPP-MC-Dropout; and detecting a to-be-detected target in automatic driving based on a DETR algorithm of uncertainty quantization. The invention provides a method for calibrating the Softmax confidence coefficient based on an uncertainty result. According to the method, the output end of the target detection model is optimized based on the uncertainty quantification result, and the over-confidence problem and the interpretability problem of the Softmax confidence coefficient are relieved; according to the method, the DDPP-MC-Dropout method is provided for cognitive uncertainty quantification, the calculation complexity is reduced to the maximum extent, and the uncertainty quantification method is assisted to be applied to real-time tasks.

Description

Technical Field

[0001] This invention relates to the field of target detection algorithms in complex scenes, specifically to a target detection method for Monte Carlo uncertainty quantification for autonomous driving. Background Technology

[0002] Generally speaking, uncertainty can be broadly categorized into two types based on its source: accidental uncertainty and cognitive uncertainty. These uncertainties, when introduced into neural networks, directly impact the network's performance. Existing methods for quantifying uncertainty in deep learning can be divided into three categories: Bayesian methods, ensemble network methods, and single deterministic network methods.

[0003] Bayesian methods require approximate inference of the posterior distribution, but while pursuing quantization accuracy, they also increase computational complexity. Ensemble methods offer better robustness and superior overall performance compared to Bayesian methods, but they cannot accurately distinguish the sources of uncertainty and have poorer real-time performance. Single deterministic network methods offer good real-time performance but suffer from high network complexity. In the field of autonomous driving, the computational complexity issue brought about by uncertainty methods is particularly prominent. How to achieve good uncertainty quantification results with minimal computational cost has become a pressing problem to be solved for the industrial application of this method. Furthermore, in open-world scenarios, sample sources are complex and varied and susceptible to noise interference. Random uncertainty noise is difficult to quantify accurately, while out-of-distribution detection relies on the quantification of cognitive uncertainty. Applying uncertainty quantification methods to open-world object detection requires addressing the impact of noise on the quantification of cognitive uncertainty.

[0004] After searching, we found application publication number CN120526405A, which discloses a target detection method, device, and autonomous vehicle based on uncertainty. A comparison of this invention with the aforementioned prior art CN120526405A reveals the following differences: 1. Differences in Data Sources and Fusion Methods: This invention primarily targets object detection using single visual source data (such as images) and improves detection performance by quantifying the internal uncertainty of a deep learning model. The comparative document, however, focuses on fusing data from two different sensors, radar and vision, where uncertainty stems from the heterogeneity of multimodal data. The two differ fundamentally in their data processing sources and fusion strategies. 2. Differences in Uncertainty Quantification Methods: This invention employs uncertainty quantification techniques based on Monte Carlo Dropout and Diverse Determinantal Point Process (DDPP), focusing on reducing computational complexity while improving the model's ability to quantify cognitive uncertainty. The comparative document, however, does not detail its uncertainty quantification method. Based on its description, it can be inferred that its method may not specifically address the quantification of cognitive uncertainty within the model, but rather focuses more on uncertainty processing at the data fusion level.

[0005] 3. Computational Efficiency and Real-Time Considerations: This invention optimizes computational efficiency through DDPP-MC-Dropout, ensuring the feasibility of uncertainty quantification in real-time tasks, which is particularly important for scenarios requiring immediate response, such as autonomous driving. In contrast, the multimodal fusion methods used in the comparison documents, especially in the uncertainty fusion stage, may introduce additional computational burdens and affect real-time processing capabilities.

[0006] 4. Model calibration and confidence level adjustment: This invention proposes a confidence calibration method with adjustable temperature scaling, which can effectively improve the calibration of the softmax confidence of the model output and reduce the overconfidence problem. In contrast, the comparative documents mention less about confidence calibration or adjustment of the model output probability distribution, possibly because they lack a dedicated mechanism for optimizing model confidence. Summary of the Invention

[0007] This invention aims to solve the problems of the prior art. It proposes a target detection method based on Monte Carlo uncertainty quantification for autonomous driving.

[0008] This invention optimizes the output of the target detection model based on the uncertainty quantification results, mitigating the overconfidence and interpretability issues of Softmax confidence. This invention proposes the DDPP-MC-Dropout method for cognitive uncertainty quantification, minimizing computational complexity and facilitating the application of uncertainty quantification methods in real-time tasks. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the Monte Carlo Dropout algorithm according to a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of the DDPP Monte Carlo Dropout algorithm. Figure 3 This is a schematic diagram of the DETR algorithm based on uncertainty quantization; Figure 4 This is a schematic diagram of a temperature scaling method; Figure 5 Represents the confidence level histogram and reliability plot; Figure 6 The calibration curves are shown under different positioning parameters. Figure 7 Represents calibration heatmaps under different coordinate parameters; Figure 8 This represents the calibration heatmap under different frame size parameters. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0011] The technical solution of the present invention to solve the above-mentioned technical problems is: 1. Uncertainty Modeling in Target Detection (1) Quantification of uncertainty Cognitive uncertainty arises from a lack of training data or limitations in the model's structure. This uncertainty can be effectively modeled using the distribution of the model's parameters. Referring to Bayesian theory, a Gaussian distribution can be used as the prior distribution of the weight parameters in regression models. The output of the obtained Bayesian network is defined as Then the model likelihood is In a given training dataset Below, the object detection output (including bounding box and class probability) can be obtained by marginalizing the model parameters: (1) Where w represents the trained model weights, x represents the input image, and y represents the output result. Calculating the above integral is difficult because the posterior distribution cannot be solved analytically. This posterior distribution can capture a reasonable set of model parameters. This is the key to training a Bayesian neural network.

[0012] In this invention, Monte Carlo Dropout is used to approximate the posterior distribution: during the testing phase, model parameters are sampled from the approximate posterior distribution through T forward propagations (each time randomly discarding neurons). Each forward propagation can be considered as a process from... Sampling a parameter Thus, T models with different parameters were obtained, such as Figure 1 As shown, after obtaining the output vectors of T models, taking the variance or entropy yields the uncertainty. Therefore, the marginalization result in (1) can be approximated as: (2) Where T is the number of forward propagations. For the first i Dropout sampling parameters during the next forward pass.

[0013] In object detection, the output results are... y Typically includes bounding box coordinates With category probability For bounding box coordinates, Monte Carlo Dropout can be used to calculate the variance and characterize its uncertainty. (3) in This represents the bounding box coordinates of the i-th propagation. This represents the inherent noise in the observed data. The mean value of the bounding box parameters can be calculated using formula (4): (4) It can be considered The four terms in the equation are equally important in uncertainty quantification, and their summation yields the total regression uncertainty: (5) The uncertainty of class probabilities can be measured by the entropy of the probability vector: (6) in The mean of the class probabilities can be used as an approximation, referring to formula (3.2). Further calculations can be performed in classification tasks: (7) (2) Uncertainty loss function The MC-Dropout loss function obtained by minimizing the KL divergence is: (8) Where N is the total number of data points. p This represents the Dropout probability. From the distribution A sample drawn from the middle, This represents the predicted output of the model during the i-th forward propagation. This is the model weight parameter matrix. is a constant, which is the weight coefficient of the regularization term.

[0014] In regression tasks, the output with uncertainty quantification results is: (9) In regression tasks, the negative log-likelihood in formula (8) can be further simplified: (10) The constant part does not affect model training and can be ignored. The loss function can be reduced to: (11) in N The total number of samples, i The sample index is used. The loss can be viewed as two parts: the first part is the residual of the regression model, used to capture the cognitive uncertainty of the model; the second part is the random uncertainty caused by inherent noise, which acts as a regularization term. This method does not directly separate the random uncertainty, but rather treats it as an adaptive weight. To enable the model to treat samples with different noise levels differently. Random uncertainty is difficult to quantify directly, but through the design of adaptive weights in (11), the model can implicitly learn random uncertainty while explicitly learning the mean.

[0015] In classification tasks, the first i The output prediction vector for each sample A set of probabilities will be obtained through the Softmax layer. p ,assumed The predicted vector follows a Gaussian distribution, that is: (12) Here , This indicates that when the model parameters are W Output at the time.

[0016] If the total number of output categories is C ,sample i Category k ,have Based on the cross-entropy loss function commonly used in classification tasks, we can obtain: (13) The loss function is obtained by averaging the results of T executions: (14) The training function based on the hybrid loss function of (11) and (14) can theoretically improve the robustness of model detection and the reliability of calibration. The loss function can also dynamically adjust the sample weights according to the noise level, reducing the interference of accidental uncertainty on model training.

[0017] 2. DDPP-MC-Dropout for the Transformer Framework The Transformer model includes Dropout layers in multiple locations: after the embedding layer, inside the encoder and decoder, and before the output layer. Previous research has shown that Monte Carlo Dropout-based uncertainty quantization methods significantly outperform Softmax methods only when Dropout is activated in multiple layers of the Transformer architecture; however, this incurs significant computational overhead.

[0018] Because of the randomness of Dropout, Monte Carlo Dropout samples may suffer from redundancy and insufficient diversity. Multivariate determinant point process Monte Carlo Dropout (DDPP-MC-Dropout) effectively solves this problem, guaranteeing better uncertainty quantization results than Softmax even when only the last Dropout layer of the output layer is enabled. The principle of the DDPP-MC-Dropout algorithm is as follows: Figure 2 As shown, after obtaining T models through MC-Dropout, the Diverse Determinantal Point Process (DDPP) is used to select the most diverse subset. Specifically, this involves constructing a positive definite similarity matrix. L Its elements are determined by the similarity between points. Several subsets are obtained by randomizing sampling through T models, and the determinant is calculated based on the corresponding similarity submatrix. The subset with the largest determinant value is the subset with the strongest diversity.

[0019] DDPP is a probabilistic model used to model the diversity of subsets. Given a set S, DDPP defines a probability distribution. p , making any S subset of A All conditions are met: (15) in L It is a kernel matrix. yes L Corresponding to subset A A submatrix, given a dataset Its kernel matrix L It is Symmetric matrices (which can be obtained through...) (to build), its elements Indicates the similarity between data points: (16) By maximizing the determinant value Choose a subset of diversity and utilize that subset. AThe statistics are used to estimate the uncertainty of the model, for example: (17) in For diverse subsets A The mean of the values. This method tends to select subsets with high diversity and avoid selecting similar distributions, thus enabling Monte Carlo Dropout to more effectively capture the cognitive uncertainty of the model.

[0020] Formula (16) is the key in DDPP, namely the choice of kernel function. Common kernel functions include cosine similarity kernel, Gaussian kernel, linear kernel, etc. This section uses the Gaussian kernel function to calculate distribution similarity, as shown in (18). This is a kernel function that is better able to capture the nonlinear relationship between continuous values ​​and performs well in cases where the data distribution is complex and nonlinear.

[0021] (18) in and It is the output of the regression task, which is the bounding box parameter output in object detection.

[0022] 3. DETR Algorithm Flow Based on Uncertainty Quantization like Figure 3 As shown, this invention retains all feature extraction modules, encoders, and decoders of the Transformer framework, modifying only the feedforward neural network (FFN) of its output part. This section uses the DETR object detection model as an example. Its FFN consists of two parts: a class prediction FFN and a bounding box regression FFN. Both have similar structures: the FFN is composed of linear layers, a ReLU activation function, and another linear layer. The class prediction FFN adds a Softmax layer after the last linear layer to achieve multi-class normalization. DETR (DEtection Transformer) is a model name, representing the first application of the Transformer model for object detection.

[0023] Following the setup of Chen et al., this paper adds a Dropout layer after the last linear layer of each of the two FFNs. The DDPP-MC-Dropout model is used to quantify the cognitive uncertainty contained in the neural network model. During training, the Dropout layer is enabled, eliminating the need for explicit calculation of uncertainty; training is optimized by modifying the loss function. During testing, the Dropout layer is also enabled for T forward propagations. The variance of the localization information is calculated, and a weighted average of the variances of the four parameters is taken to obtain the localization reliability. The category confidence score is obtained by calculating the entropy of the category information. Finally, two additional terms are added to the output sequence to represent the category confidence and the location confidence.

[0024] As can be seen, compared with the previous operation of adding Dropout layers to multiple linear layers in the neural network, DDPP-MC-Dropout makes very little change to the neural network and has a limited increase in computational cost. The testing process of the Transformer model based on DDPP-MC-Dropout is shown in Table 1. Localization uncertainty and classification uncertainty are handled independently, and uncertainty information is incorporated into the output vector by concatenation.

[0025] Table 1. Flowchart of the DETR algorithm based on uncertainty quantization 4. Adjustable temperature compression and reliability calibration Temperature scaling (TS) is a commonly used confidence calibration technique in post-processing, which adjusts temperature parameters... T Smooth Softmax distribution as Figure 4 As shown, this approach mitigates the overconfidence problem of the model and improves the accuracy of the model's output probability. Building upon this research, this invention further proposes an uncertainty-based temperature control method, Temp-U, to achieve confidence level calibration.

[0026] Temperature scaling adjusts the output of Softmax by introducing a temperature parameter: (19) in For the model to class i The logits output, T For temperature parameters, n The number of categories. This method is simple and efficient, and effectively alleviates the problem of overconfidence in Softmax. Temperature parameter. T Directly affecting the model's output probability distribution, it typically requires training with a negative log-likelihood loss function to determine its specific value, but it is fixed. T The value has limitations: While this can mitigate the model's overconfidence, it also correspondingly reduces the confidence level it gives for correctly identified samples. The Temp-U method, after obtaining the class confidence parameters... Then, the temperature can be adjusted using uncertain parameters. T : (20) in and This is a hyperparameter used to control the effect of uncertainty on temperature; specific values ​​are given in the Effects section of this invention. This temperature parameter enables the model to operate under significant uncertainty. To alleviate model overconfidence; when uncertainty is low The model can provide more reliable results.

[0027] Key points of the invention The DDPP-MC-Dropout is optimized based on Gaussian process kernels, which reduces the computational complexity of MC-Dropout by selecting the most diverse samples; the DETR method for uncertainty quantification based on DDPP-MC-Dropout is also implemented.

[0028] A method for calibrating the Softmax confidence score based on uncertainty results is proposed.

[0029] 1. Experimental Details To verify the performance of the proposed algorithm, experiments will be conducted to compare computational complexity, algorithm accuracy, and confidence reliability. The computing platform used for model training is an Nvidia A100, and the platform for inference and visualization analysis is an Nvidia RTX 3090. Platform parameters are shown in Table 3, and all experiments maintained a consistent environment. For object detection, the Deformable DETR model was used as the baseline model, evaluated on the COCO 2017 and Cityscapes datasets. The backbone network used a ResNet-50 pre-trained on ImageNet-1k, maintaining the same network structure and training parameters as Deformable DETR, without using additional data. To achieve a higher inference frame rate, the forward propagation times of DDPP-MC-Dropout were set to 20, and the temperature-adjusted hyperparameter was adjusted. and The values ​​were set to 0.5 and 0.9 respectively. The calibration and visualization framework used net:calv1.3. See Table 3.4 for specific parameters.

[0030] Table 3 Computing Platform Parameter Settings Table 4 Experimental Parameter Settings 2. Basic performance comparison experiment (1) Comparison of experimental setup and results To evaluate the impact of MC-Dropout on computational complexity, FPS, parameter count, and detection accuracy, this section compares the performance of UD-DETR (our method), DeformableDETR (the baseline method), DETR, and Faster R-CNN using the COCO dataset and DeformableDETR as the benchmark model. Performance under different Dropout configurations is also compared: 1 layer, 2 layers, 5 layers, and a combination of 1 layer and DDPP. With DDPP enabled, the number of forward propagation iterations T is set to 20, and the number of random samples for uncertainty calculation is set to 5 to 20, while other MC-Dropout models are fixed at 20. Experimental results are shown in Table 5.

[0031] Table 5 Comparison of target detection performance on the COCO dataset (2) Comparative analysis of experimental results Experimental results show that the UD-DETR model using DDPP and single-layer MC-Dropout performs best. It completes uncertainty quantification and confidence calibration while maintaining FPS, and slightly improves detection accuracy, showing significant advantages over the multi-layer MC-Dropout method.

[0032] ① Computational complexity (GFLOPS) Traditional MC-Dropout leads to a linear increase in model computational complexity; that is, setting the number of forward propagations to T increases computational complexity by a factor of T. This high computational complexity results in a decrease in inference frame rate, which is unacceptable for real-time tasks. Experiments show that when applying a single layer of MC-Dropout and DDPP, the model's computational complexity increased from 173 to 184 compared to the baseline method, a 6% increase. In contrast, using one layer of MC-Dropout increased computational complexity by 11%, using two layers by 21%, and using five layers by 54%. The experimental results indicate that increasing the number of Dropout layers leads to an approximately multiple increase in computational complexity, while the increase in computational complexity brought by DDPP is insignificant.

[0033] ② Inference Frame Rate (FPS) Theoretically, MC-Dropout increases computational complexity while correspondingly reducing the inference frame rate. Experimental results confirm that as the number of MC-Dropout layers increases from 0 to 1, 2, and 5, the inference frame rate gradually decreases, even falling below that of the traditional detection model Faster R-CNN. DDPP enables UD-DETR to maintain real-time performance comparable to Deformable DETR. This is thanks to DDPP's ability to effectively reduce computational redundancy and eliminate the need for excessive forward propagation. Experimental results show that using a single Dropout layer combined with DDPP has almost no impact on the real-time performance of the object detection network.

[0034] ③ Number of model parameters (M) Theoretical analysis shows that using MC-Dropout uncertainty quantization does not introduce additional training parameters; that is, MC-Dropout itself does not increase the number of parameters. DDPP, on the other hand, is a probabilistic model used to generate diverse sample sets. It primarily affects the selection of the Dropout mask, not the model weights themselves, and does not affect the number of model parameters. The number of model parameters in UD-DETR remains comparable to the baseline methods.

[0035] ④ Mean Precision (mAP) UD-DETR's accuracy is slightly better than the baseline method. The combination of 1 Dropout layer plus DDPP achieves the best performance in four aspects, but it is not significantly correlated with the number of MC-Dropout layers. Theoretically, increasing the number of MC-Dropout layers will not increase or even slightly reduce the detection accuracy. However, the uncertainty-based training strategy will improve the model's detection accuracy and training convergence speed.

[0036] 3. Calibration performance comparison experiment (1) Comparison of experimental setup and results To verify the effectiveness of the algorithm for confidence calibration, this section conducts confidence calibration experiments on the COCO and Cityscapes datasets. Using Deformable-DETR as the baseline, it compares the algorithm with TS, MDCA, MbLS, TCD, and UD-DETR (our method), using D-ECE and average precision as evaluation metrics. In the COCO dataset experiments, COCO was set as the in-domain dataset and CorCOCO as the out-of-domain dataset; the results are shown in Table 6. In the Cityscapes dataset experiments, Cityscapes was set as the in-domain dataset, and FoggyCityscapes and RainCityscapes were used as out-of-domain datasets; the results are shown in Table 7.

[0037] (2) Comparative analysis of experimental results ① Experimental results of COCO-CorCOCO This experiment primarily verifies calibration performance, specifically the D-ECE index. UD-DETR outperforms all other methods, reducing the D-ECE by 3.7 on COCO and 0.8 on CorCOCO compared to the baseline. In terms of detection accuracy, the methods are similar, with this method achieving the best accuracy on CorCOCO and second only to MbLS on COCO. However, MbLS exhibits the worst D-ECE performance among all methods. Overall, this method significantly improves calibration performance and maintains good accuracy in out-of-domain scenarios.

[0038] ②Cityscapes-FoggyCityscapes-RainCityscapes Experiment Results This experiment targets complex autonomous driving scenarios to verify the calibration performance and robustness of the proposed method. The method demonstrates significantly superior calibration performance on both in-domain (Cityscapes) and out-of-domain (FoggyCityscapes and RainCityscapes) datasets, with accuracy second only to the TCD method. Under various weather conditions, the proposed method performs best in foggy weather, showing the largest lead over the baseline. Overall, the proposed method exhibits leading calibration performance and good accuracy, with minimal impact from weather conditions.

[0039] Table 3.6 Experimental results on the COCO dataset Table 3.7 Experimental results on the Cityscapes dataset 4. Ablation test To verify the effectiveness of each innovative module, ablation experiments were conducted based on the UD-DETR model. The ablation experiments included three variables: an uncertainty training strategy (Loss_U), the DDPP-MC-Dropout module (DDPP-MCD), and a temperature-scaling-based confidence correction module (Temp). All experiments were performed on the COCO dataset, and the results are shown in Table 8. The results indicate that each innovative module improves the D-DETR target detection algorithm. Detection accuracy is only affected by the Loss_U module. All three modules reduce the expected calibration error, with the Temp module showing the most significant improvement. The DDPP-MCD module further enhances the calibration effect of the Temp module. Overall, the module combination slightly improves detection accuracy and significantly improves calibration performance.

[0040] Table 3.8 Ablation Experiment Results 5. Visual Analysis (1) Confidence calibration performance analysis To visually demonstrate the effect of uncertainty confidence calibration, this section compares the confidence distribution histograms and reliability plots of the baseline method and our method. The confidence distribution histogram uses confidence as the x-axis and the percentage of the total sample as the y-axis. The two vertical dashed lines correspond to the average accuracy and average confidence, respectively; the closer the two lines are, the better the calibration effect. The reliability plot uses confidence as the x-axis and accuracy as the y-axis. The diagonal line represents the expected accuracy, i.e., the perfect calibration line. The difference between accuracy and the perfect calibration line (Gap) is marked in red; the smaller the Gap area, the better the model calibration performance. To highlight the algorithm's calibration effect, the IoU threshold is set to 0.5 (a lower IoU threshold will lead to a larger Gap). The confidence histograms and reliability plots of three randomly selected categories are shown below. Figure 5 As shown.

[0041] The confidence histogram shows that the confidence level output by this method (b) is closer to the mean accuracy. In the reliability plot, the confidence level of the baseline method (a) is generally low, while the accuracy of this method is evenly distributed on both sides of the ideal calibration line, with a gap area much smaller than the baseline. The results indicate that this method can improve the overcorrection of the model for known class confidence levels, and the predicted confidence level is more reliable.

[0042] (2) Analysis of calibration performance-frame positioning parameter curves To analyze the variation of calibration performance with bounding box parameters, this section plots different... The confidence level, precision, and ECE curves under (x-coordinate of the box center point, y-coordinate of the box center point, box height, and box width) are as follows: Figure 6 As shown in the figure, the black, blue, and red dashed lines represent accuracy, confidence level, and ECE, respectively. The closer the value is to 0 or 1, the closer the target is to the image edge. The closer h and w are to 0, the smaller the target's proportion in the image; the closer h and w are to 1, the larger the proportion of the detected target in the image. The area under the ECE curve of this method (b) is smaller than that of the baseline method (a), indicating that the overall calibration performance of this method is superior. Parameter analysis shows that this method is generally better than the baseline, but there is still room for improvement in the calibration of image edge targets; parameter Analysis shows that this method is effective in... At smaller values, ECE performance significantly outperforms the baseline, while... The ECE value steadily approaches 0 when the target size is large, indicating that the proposed method is accurate and stable in predicting large targets. Theoretically, fixed-temperature calibration may unnecessarily reduce the confidence level of high-determinism targets, while the proposed method provides lower uncertainty and a lower temperature parameter T for the main target, thus achieving a more accurate confidence level.

[0043] (3) Calibration performance - thermographic analysis of frame positioning parameters Figure 7 Heatmaps showcasing calibration performance; the x and y axes of each heatmap represent the box positioning parameters, respectively. The pixel color mapping represents the average accuracy, average confidence, and ECE, with red indicating high values ​​and dark blue indicating low values. The baseline method (a) has significantly higher ECE values ​​in the left and lower right corners, indicating that the baseline method has mediocre calibration performance when dealing with edge targets. In contrast, the proposed method (b) has a higher ECE value in the upper right corner, and its overall performance is significantly better than the baseline.

[0044] Figure 8 Demonstrated reliability - Heatmap, and Figure 7 Similarly, the horizontal and vertical axes of each heatmap represent the bounding box positioning parameters. The pixel color mapping represents the average accuracy, average confidence, and ECE, with red indicating high values ​​and dark blue indicating low values. The baseline method (a) has a high ECE in the lower left corner, indicating that its calibration performance in small target detection is average. In contrast, the proposed method (b) significantly reduces the overall ECE, demonstrating significantly better calibration performance in small target detection than the original method.

[0045] Through comprehensive visualization analysis of ECE metrics, the superiority of this method relative to the baseline was verified, while also revealing the advantages and limitations of the method: this method better fits the accuracy and confidence, has excellent ECE performance on large targets, and significantly improves the calibration performance for small targets, but has limited improvement on the calibration of image edge regions.

[0046] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0047] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A target detection method based on Monte Carlo uncertainty quantification for autonomous driving, characterized in that, Includes the following steps: Constructing a multivariate determinant point process Monte Carlo DDPP-MC-Dropout for the Transformer framework; The DETR algorithm, based on uncertainty quantization, is used to detect targets in autonomous driving.

2. The target detection method for Monte Carlo uncertainty quantification for autonomous driving according to claim 1, characterized in that, The construction of the Monte Carlo DDPP-MC-Dropout process for the Transformer framework specifically includes: After obtaining T models through MC-Dropout, the Multivariate Determinant Point Process (DDPP) is used to select the most diverse subset. Specifically, this involves constructing a positive definite similarity matrix. L Its elements are determined by the similarity between points. Several subsets are obtained by random sampling through T models, and the determinant is calculated based on the corresponding similarity submatrix. The subset with the largest determinant value is the subset with the strongest diversity. DDPP is a probabilistic model used to model the diversity of subsets; given a set S, DDPP defines a probability distribution. p , making any S subset of A All conditions are met: （15） in L It is a kernel matrix. yes L Corresponding to subset A A submatrix, given a dataset Its kernel matrix L It is A symmetric matrix whose elements Indicates the similarity between data points: （16）； By maximizing the determinant value Choose a subset of diversity and utilize that subset. A The statistics are used to estimate the uncertainty of the model. （17） in For diverse subsets A The mean of the distribution is used; the distribution similarity is calculated using the Gaussian kernel function, as shown in (18). （18） in and It is the output of the regression task, which is the bounding box parameter output in object detection.

3. The target detection method for Monte Carlo uncertainty quantification for autonomous driving according to claim 1 or 2, characterized in that, All feature extraction modules, encoders, and decoders of the Transformer framework are retained, and only the feedforward neural network (FFN) of its output part is modified. In the DETR object detection model, its FFN consists of two parts: the category prediction FFN and the bounding box regression FFN. The two have similar structures: both consist of a linear layer, a ReLU activation function, and a linear layer in the FFN. The category prediction FFN adds a Softmax layer after the last linear layer to achieve multi-class normalization. A Dropout layer was added after the last linear layer of each of the two FFNs. The DDPP-MC-Dropout model was used to quantify the cognitive uncertainty contained in the neural network model. During the training phase, the Dropout layer was enabled for training, eliminating the need to explicitly calculate uncertainty. Training was optimized by modifying the loss function. During testing, the Dropout layer was also enabled for T forward propagations. The variance of the location information was calculated, and the variances of the four parameters were weighted and averaged to obtain the location reliability. The category confidence score is obtained by calculating the entropy of the category information. Finally, two additional terms are added to the output sequence to represent the category confidence and the location confidence.

4. The target detection method for Monte Carlo uncertainty quantification for autonomous driving according to claim 1, characterized in that, It also includes the following steps: The Temp-U method, a temperature control method based on uncertainty, is used to achieve confidence calibration. Temperature scaling adjusts the output of Softmax by introducing a temperature parameter: （19） in For the model to class i The logits output, T For temperature parameters, n The Temp-U method obtains the class confidence parameter, which represents the number of categories. Then, the temperature can be adjusted using uncertain parameters. T : （20）； in and This is a hyperparameter.

5. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the target detection method for Monte Carlo uncertainty quantification for autonomous driving as described in any one of claims 1 to 4.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the target detection method for Monte Carlo uncertainty quantification for autonomous driving as described in any one of claims 1 to 4.

Images