A real-time detection method for elastic threshold of vehicle-mounted visual adversarial samples

Abstract

The application discloses a kind of vehicle-mounted visual confrontation sample's elasticity threshold real-time detection method, belong to computer vision technical field, it includes obtaining the continuous multiple frames real-time image data of vehicle-mounted visual sensor acquisition, generates original image sequence to carry out space-time feature extraction and disturbance analysis, generates visual disturbance feature vector, combined with the dynamic environment evaluation of obtained vehicle-mounted auxiliary sensor data, generates vehicle environment state vector;According to vehicle environment state vector and visual disturbance feature vector, the dynamic generation of elasticity adjustment parameter is fused and calculated elasticity threshold, visual disturbance feature vector is compared with elasticity threshold and is judged, generates confrontation sample detection result.Adopting the environment state evaluation of multi-sensor fusion adaptively generates dynamic detection threshold, can distinguish normal image change and malicious confrontation disturbance according to the real-time complexity of driving scene, to reliably and robustly realize real-time safety detection under dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and in particular to a real-time method for detecting adversarial examples in vehicle-mounted vision using an elastic threshold. Background Technology

[0002] In intelligent driving systems, onboard vision sensors are a core component for environmental perception, and the image data they input directly affects the safety of vehicle decision-making and control. In recent years, adversarial example attacks targeting deep neural network vision models have emerged as a new security threat. Attackers can cause models to misidentify data by adding perturbations imperceptible to the human eye to the input image, potentially leading to serious driving safety hazards. Therefore, real-time detection of adversarial examples on the input visual data at the vehicle end has become an important direction for security technology.

[0003] In existing technologies, adversarial example detection schemes are mostly designed in static or controlled environments. One common approach is to analyze image feature distribution offline and set a globally fixed anomaly threshold. When the feature deviation of the input image exceeds this threshold, it is identified as an adversarial example. Another approach focuses on using the statistical properties of the image itself or the activation state of neurons within the model to identify abnormal patterns. These schemes mostly rely on a single visual data stream for analysis and decision-making.

[0004] However, fixed thresholds cannot adapt to the natural fluctuations in image features under different lighting, weather, and motion conditions. This leads to overly lenient thresholds in simple scenarios that may result in missed detections, while overly strict thresholds in complex scenarios generate a large number of false alarms. Relying solely on visual data stream analysis makes it difficult to distinguish between drastic image changes caused by normal driving behaviors such as severe weather and sharp turns, and disturbances introduced by malicious attacks. The detection logic lacks generalization ability and scene adaptability. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a real-time elastic threshold detection method for vehicle-mounted visual adversarial examples. It employs multi-sensor fusion environmental state assessment to adaptively generate dynamic detection thresholds, enabling the differentiation between normal image changes and malicious adversarial disturbances based on the real-time complexity of the driving scenario. This achieves reliable and robust real-time security detection in dynamic environments.

[0006] The above objectives can be achieved through the following approach: A real-time elastic threshold detection method for adversarial examples in vehicle-mounted vision includes: acquiring multiple consecutive frames of real-time image data collected by a vehicle-mounted vision sensor to generate an original image sequence; performing spatiotemporal feature extraction and perturbation analysis on the original image sequence to generate a visual perturbation feature vector; acquiring vehicle-mounted auxiliary sensor data and combining it with the visual perturbation feature vector to perform dynamic environment assessment and generate a vehicle environment state vector; performing fusion calculation based on the vehicle environment state vector and the visual perturbation feature vector to obtain an elastic adjustment parameter, and dynamically generating an elastic threshold using the elastic adjustment parameter; comparing the visual perturbation feature vector with the elastic threshold to generate an adversarial example detection result.

[0007] Optionally, the step of acquiring continuous multi-frame real-time image data collected by the vehicle-mounted vision sensor and generating an original image sequence includes: acquiring the original output signal of the vehicle-mounted vision sensor, generating continuous multi-frame real-time image data and applying filtering processing to remove noise interference to obtain denoised image data; aligning the denoised image data with historical frame data through motion compensation to generate an aligned multi-frame image sequence; and performing brightness equalization and resolution normalization processing on the aligned multi-frame image sequence to generate an original image sequence.

[0008] Optionally, the step of acquiring the original output signal of the vehicle-mounted vision sensor, generating multiple consecutive frames of real-time image data, and applying filtering to remove noise interference to obtain denoised image data includes: acquiring the original output signal of the vehicle-mounted vision sensor and generating multiple consecutive frames of real-time image data; applying adaptive filtering to the multiple consecutive frames of real-time image data to remove noise interference and generating intermediate filtered image data; and performing multi-channel color space conversion correction on the intermediate filtered image data to generate denoised image data.

[0009] Optionally, performing spatiotemporal feature extraction and perturbation analysis on the original image sequence to generate a visual perturbation feature vector includes: performing inter-frame pixel difference operations on the original image sequence to generate an inter-frame difference feature map; extracting local texture and global statistical characteristics from the inter-frame difference feature map to generate a preliminary perturbation representation vector; and aggregating the preliminary perturbation representation vector using temporal weighting to generate a visual perturbation feature vector.

[0010] Optionally, performing inter-frame pixel difference operations on the original image sequence to generate an inter-frame difference feature map includes: selecting the current frame and the previous historical frame from the original image sequence to generate adjacent frame image pairs; performing pixel-level difference calculations on the adjacent frame image pairs and extracting the absolute difference values ​​to generate an original difference image; and applying a smoothing filter to the original difference image to eliminate isolated noise points to generate an inter-frame difference feature map.

[0011] Optionally, the step of acquiring vehicle-mounted auxiliary sensor data and combining it with the visual disturbance feature vector to perform dynamic environment assessment and generate a vehicle environment state vector includes: acquiring output data from the vehicle-mounted IMU sensor, GPS sensor, and weather perception unit to generate original parameters of the multi-source sensors; performing normalization and fusion processing on the original parameters of the multi-source sensors to generate fused sensor features; and performing complex quantitative assessment by combining the visual disturbance feature vector and the fused sensor features to generate a vehicle environment state vector.

[0012] Optionally, the step of acquiring the output data of the vehicle-mounted IMU sensor, GPS sensor, and weather sensing unit to generate multi-source sensor raw parameters includes: acquiring real-time motion data of the vehicle-mounted IMU sensor to generate motion state parameters; acquiring positioning and speed data of the vehicle-mounted GPS sensor to generate position and speed parameters; acquiring weather and illumination data of the external environment sensing sensor and associating and mapping them with the motion state parameters and position and speed parameters to generate multi-source sensor raw parameters.

[0013] Optionally, the generation of adversarial example detection results includes: extracting dynamic weight factors based on the vehicle environment state vector to generate environment adaptation weights; performing weighted fusion operations on the environment adaptation weights and the visual disturbance feature vector to generate elastic adjustment parameters; adaptively calibrating the disturbance intensity based on the elastic adjustment parameters to obtain an elastic threshold; and applying the elastic threshold to the threshold comparison of the visual disturbance feature vector to generate adversarial example detection results.

[0014] Optionally, the step of extracting dynamic weight factors from the vehicle environment state vector and generating environment adaptation weights includes: parsing the illumination condition component and dynamic motion component from the vehicle environment state vector to generate an environment component set; calculating a scene complexity quantification value for the environment component set to generate a complexity score; and dynamically mapping and calculating weight coefficients based on the complexity score to generate environment adaptation weights.

[0015] Based on the same inventive concept, this invention also provides a real-time elastic threshold detection system for vehicle-mounted visual adversarial examples. The system includes: an image acquisition module for acquiring multiple consecutive frames of real-time image data from a vehicle-mounted visual sensor to generate an original image sequence; a feature extraction module for performing spatiotemporal feature extraction and perturbation analysis on the original image sequence to generate a visual perturbation feature vector; an environment assessment module for acquiring vehicle-mounted auxiliary sensor data and combining it with the visual perturbation feature vector to perform dynamic environment assessment and generate a vehicle environment state vector; and a threshold judgment module for fusing the vehicle environment state vector and the visual perturbation feature vector to obtain an elastic adjustment parameter, dynamically generating an elastic threshold using the elastic adjustment parameter, comparing the visual perturbation feature vector with the elastic threshold, and generating an adversarial example detection result.

[0016] Compared with the prior art, the present invention has the following advantages: This invention can dynamically adjust the judgment boundary based on the real-time environmental conditions, so that the detection standard matches the complexity of the current scene, thereby maintaining a high detection rate under changing conditions.

[0017] This invention not only analyzes the perturbation features of the image sequence itself, but also simultaneously integrates vehicle motion status, geographical location, and external environmental information to construct a unified environmental state assessment. This cross-modal information integration enables the system to distinguish between image differences caused by normal environmental changes and genuine malicious attack perturbations, reducing false positives. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a real-time detection method for the elastic threshold of adversarial samples in vehicle-mounted vision, according to an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the normalized distribution of the sensor parameters in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram illustrating the dynamic change of the elastic threshold with environmental complexity according to an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the structure of a real-time detection system for elastic thresholds of adversarial samples in vehicle-mounted vision, according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Reference Figure 1 One embodiment of the present invention proposes a real-time elastic threshold detection method for vehicle-mounted visual adversarial examples. It adopts multi-sensor fusion environmental state assessment to adaptively generate dynamic detection thresholds, which can distinguish between normal image changes and malicious adversarial disturbances based on the real-time complexity of the driving scene, thereby achieving reliable and robust real-time security detection in dynamic environments.

[0025] The method in this embodiment specifically includes: S1. Acquire multiple consecutive frames of real-time image data collected by the vehicle-mounted vision sensor and generate the original image sequence; In one embodiment of the present invention, step S1 includes the following steps: Acquire the raw output signal of the vehicle-mounted vision sensor and generate multiple consecutive frames of real-time image data; Adaptive filtering is applied to remove noise interference from multiple consecutive frames of real-time image data to generate intermediate filtered image data. Multi-channel color space conversion and correction are performed on the intermediate filtered image data to generate denoised image data; The denoised image data is aligned with the historical frame data through motion compensation to generate an aligned multi-frame image sequence. The aligned multi-frame image sequence is subjected to brightness equalization and resolution normalization to generate the original image sequence.

[0026] Specifically, the unprocessed electrical signals output by vision sensors installed on the vehicle are acquired and converted into a standard-format matrix. Each element in the matrix represents the color and brightness value of a point in the image, thus obtaining multiple consecutive frames of real-time image data. Adaptive filtering is then performed on these multiple frames of real-time image data to remove noise interference, generating intermediate filtered image data. The core of this adaptive filtering is to calculate the statistical distribution of pixel values ​​within the neighborhood of each pixel and dynamically adjust the filtering intensity based on the noise level. Let the input image pixel values ​​be... ,in Let be the pixel coordinates, and let the size of the local window centered on be . Calculate the mean value of the pixels within the window. and standard deviation :

[0027] in, Indicates the local average brightness. Characterizes local contrast and noise intensity. Filtered pixel values. It can be obtained through the following formula:

[0028] in, As an adjustment factor, based on actual measurements from 200 sets of industrial sensors, the value range is set to [value range missing]. , used to fine-tune the filtering intensity; This is an estimate of the global noise standard deviation pre-calibrated based on sensor characteristics; It is a very small positive number, used to prevent the denominator from being zero. This formula guarantees that in flat regions... Small size, strong filtering effect to suppress noise, especially in edge regions. Large filter size results in weak filtering to preserve details.

[0029] Multi-channel color space conversion and correction are performed on the intermediate filtered image data to generate denoised image data. Considering the potential deviation in sensor color response under different lighting conditions, the image is converted from the sensor's original color space to a standard color space. This process involves a... Correction matrix Implementation. Let the original three-channel pixel vector be... Corrected pixel vector The calculation is as follows:

[0030] Among them, matrix The element values ​​are predetermined by regression analysis of color chart images taken by this type of sensor under standard light source. The purpose is to minimize the difference between the corrected color values ​​and the standard color chart reference values ​​under the color difference formula.

[0031] The denoised image data is aligned with pre-stored historical frame data through motion compensation to generate an aligned multi-frame image sequence. The most recent historical frame is selected as the reference frame, and the overall translational motion of the current frame relative to the reference frame is calculated. The translational motion is calculated using a phase correlation-based method. Let the current frame be... Historical reference frame is Their two-dimensional discrete Fourier transforms are respectively and Calculate the cross power spectrum:

[0032] in yes The complex conjugate of . For The impulse response function is obtained by performing an inverse Fourier transform, and the coordinates of its peak position are the translation amount. Use this translation amount to adjust the current frame. Perform a translation interpolation transformation to obtain the current frame image aligned with the reference frame. Repeat this alignment process for all frames to be processed to ensure that each frame in the sequence is spatially aligned with a common reference frame.

[0033] The aligned multi-frame image sequence is subjected to brightness equalization and resolution normalization to generate the original image sequence. Brightness equalization uses a histogram matching method to adjust the brightness histogram of each frame to match the brightness histogram distribution of the reference frame, eliminating inter-frame brightness differences caused by gradual changes in illumination. Resolution normalization scales all frame images to a uniform size using bilinear interpolation, for example... Pixels are used to ensure data dimensionality consistency in subsequent processing. The resulting image data set after all the above processing constitutes the original image sequence for subsequent steps.

[0034] For example, the onboard vision sensor outputs at a rate of 30 frames per second. The image is pixel-level. Ten consecutive frames of data are acquired, first undergoing adaptive filtering, and then a local window is set. Sensor calibration ,Pick For the coordinates in the first frame image Calculate its pixels Neighborhood obtained The original value of this point Substituting into the filtering formula, we get Color correction uses a pre-calibrated matrix: , For the filtered pixel vector of that point After conversion, the corrected vector is approximately Motion compensation calculations yield the translation amount of the current frame relative to the historical reference frame. Pixel alignment is achieved through interpolation. Finally, the brightness histograms of all 10 aligned images are matched to the histogram distribution of the first frame, and then uniformly scaled using bilinear interpolation. Pixels, thus generating a raw image sequence containing 10 frames of normalized images.

[0035] S2. Perform spatiotemporal feature extraction and perturbation analysis on the original image sequence to generate a visual perturbation feature vector; In one embodiment of the present invention, step S2 includes the following steps: Select the current frame and the previous historical frame from the original image sequence to generate adjacent frame image pairs; Pixel-level difference calculation is performed on adjacent frame image pairs and the absolute difference is extracted to generate the original difference image; A smoothing filter is applied to the original difference image to eliminate isolated noise points, generating an inter-frame difference feature map; Local texture and global statistical properties are extracted from the inter-frame difference feature map to generate a preliminary perturbation representation vector; The initial perturbation representation vector is aggregated using time-series weighting to generate a visual perturbation feature vector.

[0036] Specifically, two frames with adjacent timestamps are selected from the original image sequence in chronological order, and designated as the current frame. and its direct preceding frame This process generates adjacent frame image pairs. Pixel-level difference calculations are then performed on these pairs, involving subtracting the brightness values ​​of pixels at the same spatial coordinates in both frames and taking the absolute value of the result to generate the original difference image. For each pixel location in the image The difference is calculated as follows: , A smoothing filter is applied to the original difference image to eliminate isolated noise points caused by sensor noise or small illumination fluctuations, generating an inter-frame difference feature map. Smoothing is achieved using Gaussian filtering, with the Gaussian kernel function being... Standard deviation Based on actual measurements using 200 industrial sensors, the pixel size was set to 1.5. Smoothing calculations are represented as convolution operations. , in The core radius is usually taken as... Rounded up. Local texture and global statistical properties are extracted from the inter-frame difference feature map to generate a preliminary perturbation representation vector. Local texture characteristics are obtained by calculating the difference feature map around each pixel. Contrast features of the gray-level co-occurrence matrix in the neighborhood The global statistical properties are obtained by calculating the mean of the entire difference feature map. Standard deviation and entropy Divide the image into For non-overlapping grids, calculate the value within each grid cell. The average value is used to obtain 16 local texture values, which are then concatenated with 3 global statistical values ​​to form a 19-dimensional preliminary perturbation representation vector. .

[0037] The initial perturbation representation vector is aggregated using time-weighted aggregation to generate a visual perturbation feature vector. Considering that recent disturbances have a greater impact on the current state, for continuous... Each time step Perform weighted fusion. Let the current time be... ,past The set of vectors at each time step is The corresponding weight It decays exponentially over time:

[0038] in The attenuation coefficient is set to 0.2. The weights are then normalized. Visual perturbation feature vector This is the weighted sum of these vectors: .

[0039] For example, the 5th frame is taken from the 10-frame original image sequence generated in step S1. Frame 4 as Calculate coordinates The pixel difference value at that location is set as follows: ,but .right application Gaussian smoothing filter. Global statistical properties are calculated on the smoothed difference feature map. , ; division The mesh calculates the mean local texture contrast, resulting in 16 values; for example, the mean of the first mesh is 1.5. After stitching, a 19-dimensional image is obtained. .Pick Aggregates the current and previous two moments. Its original weight is After normalization, we get The three 19-dimensional vectors are weighted and summed to generate a final 19-dimensional visual perturbation feature vector. .

[0040] S3. Acquire vehicle auxiliary sensor data and combine it with visual disturbance feature vector to perform dynamic environment assessment and generate vehicle environment state vector. In one embodiment of the present invention, step S3 includes the following steps: Acquire real-time motion data from the vehicle-mounted IMU sensor and generate motion state parameters; Acquire positioning and speed data from the vehicle's GPS sensors and generate position and speed parameters; Acquire weather and illumination data from external environment sensing sensors and correlate and map them with motion state parameters and position and velocity parameters to generate raw parameters from multiple sensor sources; The original parameters of the multi-source sensors are normalized and fused to generate fused sensor features; By combining visual disturbance feature vectors with fused sensor features, a complex quantitative evaluation is performed to generate a vehicle environmental state vector.

[0041] Specifically, real-time readings of triaxial acceleration and triaxial angular velocity output from the vehicle-mounted IMU sensor are acquired, and after calibration and coordinate transformation, a model containing the vehicle's longitudinal acceleration is generated. lateral acceleration yaw rate Motion state parameters The system acquires the latitude and longitude coordinates and speed information output by the vehicle's GPS sensor, and calculates the vehicle's instantaneous speed on the ground plane using differential calculations. and rate of change of heading angle Generate position and velocity parameters The ambient light intensity is obtained by acquiring data from external environmental sensors, such as light sensors and temperature and humidity sensors. Weather status codes (e.g., 0 for sunny days, 1 for rainy days, etc.). The motion state parameters... Position and velocity parameters With environmental parameters Alignment and association mapping based on timestamps generate a multi-dimensional vector of raw parameters from multiple sources of sensors. .

[0042] The raw parameters from multiple sensors are normalized and fused to generate fused sensor features. Since the physical meanings and dimensions of the various parameters are different, a minimum-maximum normalization is first performed to unify their numerical range. Interval. For parameters Its normalized value for:

[0043] in and These are the empirical minimum and maximum values ​​of this parameter, obtained based on historical driving data statistics. of The normalized vector is Subsequently, principal component analysis was used to analyze... Dimensionality reduction is performed, retaining principal components with a cumulative contribution rate of over 95%, to form more compact fused sensor features. .

[0044] Combining visual perturbation feature vectors Features of fused sensors Perform complex quantification assessments to generate vehicle environmental state vectors. .Will and These are concatenated into a joint feature vector. A scenario complexity score is then calculated using a pre-trained complexity evaluation model. The model is a feedforward network containing two fully connected layers, and its formula is expressed as: , in For model parameters, It is a linear rectified activation function. Compress the output of the Sigmoid function to between. The closer the value is to 1, the more complex the environment. Ultimately, the vehicle's environment state vector... Defined as including lighting Exercise intensity and complexity score The vector, i.e. .in Depend on The norm of the normalized values ​​of acceleration and angular velocity in the equation is calculated, for example... .

[0045] For example, such as Figure 2 As shown, the IMU output is GPS output is Light sensor output Weather code (Sunny day). Composition Use predefined Perform normalization, for example for ,but .right ,but .get After dimensionality reduction through principal component analysis, a 3D model was obtained. Assuming Compare it with the 19-dimensional [structure] obtained in step S2. Concatenate the data into a 22-dimensional vector. Input the complexity evaluation model, assuming the model output... Calculate exercise intensity Assuming ,but Finally, a vehicle environment state vector is generated. .

[0046] S4. The vehicle environment state vector and the visual disturbance feature vector are fused to obtain the elastic adjustment parameter, and the elastic threshold is dynamically generated using the elastic adjustment parameter. The visual disturbance feature vector is compared with the elastic threshold to generate the adversarial example detection result.

[0047] In one embodiment of the present invention, step S4 includes the following steps: The illumination condition component and dynamic motion component are extracted from the vehicle environment state vector to generate an environment component set; Calculate a quantified value of scene complexity for each environmental component set and generate a complexity score; The weight coefficients are dynamically mapped based on the complexity score to generate environmental adaptation weights. The environmental adaptation weights and visual perturbation feature vectors are weighted and fused to generate elastic adjustment parameters. The elastic threshold is obtained by adaptively calibrating the disturbance intensity based on the elastic adjustment parameters. The elastic threshold is applied to the threshold comparison of the visual perturbation feature vector to generate adversarial example detection results.

[0048] Specifically, from the vehicle environment state vector The illumination condition component is directly analyzed. With dynamic motion components This generates an environment component set. A scene complexity quantification value is then calculated for the environment component set; this value is already included in... In the middle, that is, complexity score Scoring based on complexity Dynamic mapping calculates weight coefficients to generate environment-adaptive weights. The mapping relationship is designed as a linear function to ensure that the weights increase with increasing complexity. , in and The boundary weights are determined based on a large amount of adversarial attack test data, and are set accordingly. .when (Simplest) time ,when (Most complex) time .

[0049] Environmental adaptation weight With visual perturbation feature vector Perform weighted fusion calculations to generate elastic adjustment parameters. The fusion operation focuses on the overall strength of the perturbation vector and calculates... L2 norm and with weight Multiplication:

[0050] in , yes The Each component. It is a scalar that comprehensively reflects the magnitude of the current perceived disturbance and has been adaptively adjusted for environmental complexity.

[0051] Based on elastic adjustment parameters The elastic threshold is obtained by adaptively calibrating the disturbance intensity. Threshold calculation introduces a base threshold. This threshold is determined during the model training phase based on the perturbation norm distribution of clean samples, for example, by taking its 95th quantile. Elastic threshold. It is a weighted sum of the base threshold and the elastic adjustment parameter: , in, It is a mixing coefficient between 0 and 1, used to balance the ratio of the basic static threshold to the dynamic adjustment component, and is set to 0.6 based on the principle of optimal performance on the validation set. This formula allows the threshold to change dynamically with environmental complexity and the current level of perturbation.

[0052] like Figure 3 As shown, an elastic threshold is applied to the threshold comparison of visual perturbation feature vectors to generate adversarial example detection results. Specifically, the L2 norm of the visual perturbation feature vectors is compared. With elasticity threshold Size relationship: if The detection result is then considered an adversarial example; if If the calculated norm is greater than the dynamically generated elastic threshold, the current frame image is determined to contain adversarial perturbations, and the output is an adversarial sample; otherwise, it is determined to be a normal sample.

[0053] For example, the result obtained from step S3 Parsing Complexity score Calculate the environmental adaptation weight according to the formula. Assume that step S2 yields... The L2 norm is calculated as follows Then the elastic adjustment parameter Let the base threshold obtained during training be... Mixing coefficient Then the elastic threshold Finally, comparison ,because Therefore, adversarial sample detection results are generated, and the current frame is determined to be an adversarial sample.

[0054] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides a real-time elastic threshold detection system for vehicle-mounted visual adversarial examples, the system comprising: The image acquisition module is used to acquire multiple consecutive frames of real-time image data collected by the vehicle-mounted vision sensor and generate the original image sequence. The feature extraction module is used to extract spatiotemporal features and perform perturbation analysis on the original image sequence to generate visual perturbation feature vectors; The environmental assessment module is used to acquire data from onboard auxiliary sensors and combine it with visual disturbance feature vectors to perform dynamic environmental assessment and generate a vehicle environmental state vector. The threshold judgment module is used to calculate the elastic adjustment parameter by fusing the vehicle environment state vector and the visual disturbance feature vector, and dynamically generate the elastic threshold using the elastic adjustment parameter. The visual disturbance feature vector is compared with the elastic threshold to generate the adversarial example detection result.

[0055] It should be noted that the data interaction connections between the various units described above do not necessarily represent direct or indirect connections. Any method of indirect connection can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0056] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A real-time detection method for elastic threshold of adversarial examples in vehicle-mounted vision, characterized in that, The method includes: Acquire continuous multi-frame real-time image data collected by the vehicle-mounted vision sensor and generate the original image sequence; Spatiotemporal feature extraction and perturbation analysis are performed on the original image sequence to generate a visual perturbation feature vector; The vehicle auxiliary sensor data is acquired and combined with the visual disturbance feature vector to perform dynamic environment assessment and generate a vehicle environment state vector. The vehicle environment state vector and the visual disturbance feature vector are fused to obtain the elastic adjustment parameter, and the elastic threshold is dynamically generated using the elastic adjustment parameter. The visual disturbance feature vector is compared with the elastic threshold to generate the adversarial example detection result.

2. The method for real-time detection of elastic thresholds for adversarial examples in vehicle-mounted vision according to claim 1, characterized in that, The step of acquiring multiple consecutive frames of real-time image data collected by the vehicle-mounted vision sensor and generating the original image sequence includes: The raw output signal of the vehicle vision sensor is acquired, multiple frames of real-time image data are generated, and filtering is applied to remove noise interference to obtain denoised image data. The denoised image data is aligned with the historical frame data by motion compensation to generate an aligned multi-frame image sequence. The aligned multi-frame image sequence is subjected to brightness equalization and resolution normalization processing to generate the original image sequence.

3. The method for real-time detection of elastic thresholds for adversarial examples in vehicle-mounted vision according to claim 2, characterized in that, The process of acquiring the raw output signal of the vehicle-mounted vision sensor, generating multiple consecutive frames of real-time image data, and applying filtering to remove noise interference to obtain denoised image data includes: Acquire the raw output signal of the vehicle-mounted vision sensor and generate multiple consecutive frames of real-time image data; Adaptive filtering is applied to the continuous multi-frame real-time image data to remove noise interference and generate intermediate filtered image data; The intermediate filtered image data is subjected to multi-channel color space conversion correction to generate denoised image data.

4. The method for real-time detection of elastic thresholds for adversarial examples in vehicle-mounted vision according to claim 2, characterized in that, Spatiotemporal feature extraction and perturbation analysis are performed on the original image sequence to generate a visual perturbation feature vector, including: Perform inter-frame pixel difference operation on the original image sequence to generate an inter-frame difference feature map; Local texture and global statistical properties are extracted from the inter-frame difference feature map to generate a preliminary perturbation representation vector; The preliminary perturbation representation vector is aggregated using time-series weighting to generate a visual perturbation feature vector.

5. The method for real-time detection of elastic thresholds for adversarial examples in vehicle-mounted vision according to claim 4, characterized in that, Performing inter-frame pixel difference operations on the original image sequence to generate inter-frame difference feature maps includes: Select the current frame and the previous historical frame from the original image sequence to generate adjacent frame image pairs; Pixel-level difference calculation is performed on the adjacent frame image pairs and the absolute difference is extracted to generate the original difference image; A smoothing filter is applied to the original difference image to eliminate isolated noise points, generating an inter-frame difference feature map.

6. The method for real-time detection of elastic thresholds for adversarial examples in vehicle-mounted vision according to claim 4, characterized in that, The step of acquiring vehicle-mounted auxiliary sensor data and combining it with the visual disturbance feature vector to perform dynamic environment assessment and generate a vehicle environment state vector includes: Acquire output data from the vehicle-mounted IMU sensor, GPS sensor, and weather sensing unit to generate raw parameters from multiple sensor sources; The original parameters of the multi-source sensors are normalized and fused to generate fused sensor features; The vehicle environmental state vector is generated by combining the visual disturbance feature vector with the fused sensor features through a complex quantification evaluation.

7. The method for real-time detection of elastic thresholds for adversarial examples in vehicle-mounted vision according to claim 6, characterized in that, The process of acquiring the output data from the vehicle-mounted IMU sensor, GPS sensor, and weather sensing unit to generate raw parameters from multiple sensor sources includes: Acquire real-time motion data from the vehicle-mounted IMU sensor and generate motion state parameters; Acquire positioning and speed data from the vehicle's GPS sensors and generate position and speed parameters; The weather and illumination data from the external environment sensing sensor are acquired and correlated with the motion state parameters and position and velocity parameters to generate the original parameters of the multi-source sensor.

8. The method for real-time detection of elastic thresholds for adversarial examples in vehicle-mounted vision according to claim 6, characterized in that, The generated adversarial example detection results include: Dynamic weighting factors are extracted from the vehicle environment state vector to generate environmental adaptation weights. The environmental adaptation weights and the visual perturbation feature vectors are weighted and fused to generate elastic adjustment parameters. Based on the elastic adjustment parameters, the disturbance intensity is adaptively calibrated to obtain the elastic threshold. The elastic threshold is applied to the threshold comparison of the visual perturbation feature vector to generate adversarial example detection results.

9. A real-time detection method for elastic threshold of adversarial examples in vehicle-mounted vision according to claim 8, characterized in that, The step of extracting dynamic weight factors and generating environmental adaptation weights based on the vehicle environment state vector includes: The illumination condition component and dynamic motion component are extracted from the vehicle environment state vector to generate an environment component set; Calculate the scene complexity quantification value for the environmental component set and generate a complexity score; The weight coefficients are calculated based on the complexity score dynamic mapping to generate the environment adaptation weights.

10. A real-time elastic threshold detection system for vehicle-mounted visual adversarial examples, characterized in that, The system includes: The image acquisition module is used to acquire multiple consecutive frames of real-time image data collected by the vehicle-mounted vision sensor and generate the original image sequence. The feature extraction module is used to perform spatiotemporal feature extraction and perturbation analysis on the original image sequence to generate a visual perturbation feature vector; The environmental assessment module is used to acquire vehicle auxiliary sensor data and combine it with the visual disturbance feature vector to perform dynamic environmental assessment and generate a vehicle environmental state vector. The threshold judgment module is used to perform fusion calculation based on the vehicle environment state vector and the visual disturbance feature vector to obtain the elastic adjustment parameter, and use the elastic adjustment parameter to dynamically generate the elastic threshold. The visual disturbance feature vector is compared with the elastic threshold to generate the adversarial example detection result.

Images