Method for automatic scene calibration based on machine vision

Abstract

The application discloses a scene automatic calibration method based on machine vision, which comprises six steps: first, image acquisition and quantitative scoring, low-quality image enhancement and retry; second, target core feature extraction, effective judgment according to weight table; third, at least three heterogeneous models are used for parallel inference and result verification; fourth, space-time and behavior consistency verification is performed on the recognition results of 10 continuous frames, and manual review is performed if the results are inconsistent; finally, the threshold is dynamically adjusted combined with business scene and environmental parameters, and the results are output if the threshold is met, otherwise manual review is performed. Through multi-level verification, etc., the model recognition accuracy is improved from ≤98% to ≥99.8%, and the small target detection rate is improved from ≤95% to ≥98.5%; by fusing laser radar point cloud, etc., the ranging error is reduced from ≥10 cm to ≤3 cm; by three-dimensional scoring and screening of images, combined with enhancement and infrared fusion technology, the misjudgment rate under bad weather is reduced from ≥5% to ≤1%.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine vision, and in particular relates to a scene automatic calibration method based on machine vision. Background Art

[0002] With the rapid development of the global aviation industry, intelligent airport management has become the key to improving operational efficiency. In large aviation hub scenarios, the number of airports with passenger throughput exceeding 10 million has increased significantly, and the average daily take-off and landing volume on the apron often exceeds 500. The accuracy requirements for aircraft type identification and equipment monitoring are as follows:

[0003] Aircraft model recognition: error ≤ 0.2%; detection rate of personnel and equipment (size ≤ 1m) ≥ 98%; error ≤ 5cm (such as monitoring the distance between the aircraft nose wheel and the jet bridge).

[0004] Existing technologies (such as patents CN114863228A and CN114254696A) have significant deficiencies in the following scenarios and cannot meet industry needs:

[0005] In extreme environments, such as heavy rain with an image signal-to-noise ratio ≤20dB, and low light conditions, such as nighttime illumination ≤50lux, the traditional calibration method has an error rate of ≥5%.

[0006] The camera is affected by aircraft takeoff and landing vibrations (frequency 1050 Hz, amplitude 0.52g), and external parameter drift causes a trajectory error of ≥ 50 cm (for example, the probability of misjudging the luggage cart's position and causing a safety accident is ≥ 3%).

[0007] The missed detection rate of a single visual model (such as YOLOv8) for small targets (1m×1m luggage carts) is ≥5%, which cannot meet the safety requirement of "zero missed detection".

[0008] The fixed threshold has poor versatility in peak (100 flights / hour) and off-peak scenarios, with a misjudgment rate ≥ 10% (for example, a narrow-body aircraft is misjudged as a wide-body aircraft, and the probability of occupying a dedicated parking space is ≥ 8%). Summary of the Invention

[0009] The purpose of the present invention is to provide a scene automatic calibration method based on machine vision to solve the problems of insufficient image quality and data reliability, missing features and single model in the existing technology mentioned in the background technology.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0011] The automatic scene calibration method based on machine vision includes the following steps:

[0012] Step 1, image acquisition: obtain image data of the scene to be identified through a data acquisition device;

[0013] Step 2, image quality judgment: the collected image data is quantitatively scored; and based on the scoring results, it is determined whether the image is a valid image. If it is, it proceeds to step 3; if not, it starts the image enhancement mechanism and tries again;

[0014] Step 3, feature extraction and integrity judgment: Extract the target core features from the screened valid images, and judge the validity of the target core features by constructing a feature importance weight table and calculating the integrity score;

[0015] Step 4, multi-model cross-validation: deploy at least three heterogeneous recognition models to perform parallel inference on the target core features, collect statistics on the inference results of all heterogeneous recognition models, and determine whether the inference results meet the preset values;

[0016] Step 5: Spatiotemporal consistency judgment: The recognition results of 10 consecutive frames of valid images are checked for spatial, temporal, and behavioral consistency. If the verification results are consistent, the process proceeds to step 6. If they are inconsistent, the case is marked as an abnormality and triggers manual review.

[0017] Step 6, scenario-adaptive threshold verification: Dynamically adjust the recognition threshold based on the business scenario and environmental parameters. The recognition result meets the current scenario threshold and outputs the final recognition result. If it does not meet the threshold, it is marked as an abnormal case and triggers manual review.

[0018] According to the above technical solution, in step 2, the image quality assessment algorithm is based on the three-dimensional indicators of clarity, signal-to-noise ratio, and exposure, and uses Gaussian blur difference (DoG) and entropy value calculation to achieve quantitative scoring. The images after the quantitative scoring are screened, and images greater than or equal to the threshold are identified as valid images, and images less than the threshold are identified as low-quality images, triggering image enhancement or backup camera switching and re-calibration.

[0019] According to the above technical solution, in step 3, the feature importance weight table weights the core features, extracts features through a convolutional neural network (CNN), and calculates the completeness score. Missing features are retried after multi-frame fusion or GAN completion.

[0020] According to the above technical solution, in step 4, the inference results of the heterogeneous recognition model are output with confidence through Softmax, and the voting mechanism is used to arbitrate the results. When comparing fine-grained features, the model difference library is called to perform pixel-level matching. Specifically:

[0021] The test passes when the confidence level is ≥95% and ≥2 of the deployed heterogeneous recognition models are consistent.

[0022] When there is disagreement or low confidence <80%, the fine-grained feature library (model difference parameter ≤1cm) is called for arbitration or lidar point cloud registration.

[0023] According to the above technical solution, in step 5, the spatiotemporal consistency check includes:

[0024] Spatial consistency: target position trajectory fitting deviation ≤ 30cm;

[0025] Temporal consistency: result fluctuation variance ≤ threshold;

[0026] Behavioral consistency: The target motion complies with physical laws, and the motion state is predicted through Kalman filtering.

[0027] According to the above technical solution, in step 6, the dynamic threshold adjustment model generates a scene threshold strategy through a decision tree algorithm based on scene metadata and environmental sensor data.

[0028] According to the above technical solution, after being marked as an abnormal case, it is manually reviewed and placed in the misjudgment case library for iteration of the model.

[0029] According to the above technical solution, in step 2, the image enhancement retry mechanism uses a super-resolution algorithm to improve the clarity of low-quality images, and the backup camera switching realizes multi-camera redundancy through the RTSP protocol.

[0030] According to the above technical solution, in step 3, the feature completion model is based on the generative adversarial network (GAN), which predicts the missing part through the feature distribution of similar models and generates the completed feature map.

[0031] According to the above technical solution, in step 4, multimodal fusion uses a point cloud image registration algorithm to convert the lidar point cloud into an image coordinate system, which is then fused with visual features and input into the recognition model to improve the detection rate of small targets.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The technical solution of the present invention forms a systematic solution to the core pain points of existing technologies in complex scenarios through multi-level verification, dynamic compensation and closed-loop iteration. Through feature completion), multi-model cross-validation and fine-grained feature library arbitration, the accuracy of model recognition is improved from ≤98% of the existing technology to ≥99.8%, and the small target detection rate is improved from ≤95% to ≥98.5%, solving the problems of traditional methods misjudging similar models and missing small targets.

[0034] By integrating LiDAR point cloud and visual measurement, and combining Kalman filtering to dynamically compensate for vibration errors, the distance measurement error between the aircraft's nose wheel and the stop line is reduced from ≥10cm with existing technology to ≤3cm, meeting the "centimeter-level" positioning requirements of the apron (such as safe distance control for jet bridge docking and luggage cart parking).

[0035] Valid images are screened through three-dimensional quantitative scoring (clarity, signal-to-noise ratio, and exposure). Low-quality images are enhanced through ESRGAN super-resolution. Combined with infrared image fusion technology, the misjudgment rate in heavy rain and low-light scenes is reduced from ≥5% with existing technologies to ≤1%, solving the problem of recognition failure of traditional methods in severe weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the automatic calibration method of the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] Example 1

[0039] like Figure 1 As shown, the automatic scene calibration method based on machine vision includes the following steps:

[0040] Step 1, image acquisition: obtain image data of the scene to be identified through a data acquisition device;

[0041] Step 2, image quality judgment: the collected image data is quantitatively scored; and based on the scoring results, it is determined whether the image is a valid image. If it is, it proceeds to step 3; if not, it starts the image enhancement mechanism and tries again;

[0042] Step 3, feature extraction and integrity judgment: Extract the target core features from the screened valid images, and judge the validity of the target core features by constructing a feature importance weight table and calculating the integrity score;

[0043] Step 4, multi-model cross-validation: deploy at least three heterogeneous recognition models to perform parallel inference on the target core features, collect statistics on the inference results of all heterogeneous recognition models, and determine whether the inference results meet the preset values;

[0044] Step 5: Spatiotemporal consistency judgment: The recognition results of 10 consecutive frames of valid images are checked for spatial, temporal, and behavioral consistency. If the verification results are consistent, the process proceeds to step 6. If they are inconsistent, the case is marked as an abnormality and triggers manual review.

[0045] Step 6, scenario-adaptive threshold verification: Dynamically adjust the recognition threshold based on the business scenario and environmental parameters. The recognition result meets the current scenario threshold and outputs the final recognition result. If it does not meet the threshold, it is marked as an abnormal case and triggers manual review.

[0046] The technical solution of the present invention forms a systematic solution to the core pain points of existing technologies in complex scenarios through multi-level verification, dynamic compensation and closed-loop iteration. Through feature completion), multi-model cross-validation and fine-grained feature library arbitration, the model recognition accuracy is improved from ≤98% of the existing technology to ≥99.8%, and the small target detection rate is improved from ≤95% to ≥98.5%, solving the problems of traditional methods in misjudging similar models and missing small targets.

[0047] By integrating LiDAR point cloud and visual measurement, and combining Kalman filtering to dynamically compensate for vibration errors, the distance measurement error between the aircraft's nose wheel and the stop line is reduced from ≥10cm with existing technology to ≤3cm, meeting the "centimeter-level" positioning requirements of the apron (such as safe distance control for jet bridge docking and luggage cart parking).

[0048] Valid images are screened through three-dimensional quantitative scoring (clarity, signal-to-noise ratio, and exposure). Low-quality images are enhanced through ESRGAN super-resolution. Combined with infrared image fusion technology, the misjudgment rate in heavy rain and low-light scenes is reduced from ≥5% with existing technologies to ≤1%, solving the problem of recognition failure of traditional methods in severe weather.

[0049] Example 2

[0050] This embodiment is a further refinement of the first embodiment.

[0051] In step 1, image acquisition uses a high-definition camera (frame rate ≥ 25fps, resolution ≥ 1080p) to obtain apron images, and supports multi-camera redundancy (≥ 2 channels, RTSP protocol switching, response ≤ 100ms) to capture images of the application scenario.

[0052] In step 2, the quantitative score is based on clarity (C), signal-to-noise ratio (SNR), and exposure (E). The comprehensive score is calculated using DoG (Formula 1), entropy (Formula 2), and signal-to-noise ratio (Formula 3). , where the threshold S≥70 is a valid image.

[0053] Specifically, the calculation formula of clarity C is:

[0054]

[0055] Where, is the gradient operator, Take the spatial derivative of the signal; The modulus of the gradient, that is, the length of the gradient vector, reflects the pixel The edge strength at (the larger the gradient, the more obvious the edge); It represents the sum of the squared gradients of n samples (such as n pixels of an image); N represents the normalization factor (usually the total number of samples, i.e. (N≈n), ensuring that the result is the average value).

[0056] Specifically, the calculation formula for exposure E is:

[0057]

[0058] Where k is the value of a discrete variable (such as the grayscale value of an image, 0 to 255 corresponds to the black and white gradient of 8-bit grayscale); represents the probability of the variable taking k; It represents the self-information of k, which measures the amount of information obtained when k is observed.

[0059] Specifically, the calculation formula of the signal-to-noise ratio SNR is:

[0060]

[0061] Where, represents the mean value of the useful signal, represents the variance of the noise.

[0062] When S<70, ESRGAN super-resolution enhancement (Formula 4) or switching to a backup camera and recalibration are performed, specifically:

[0063]

[0064] Where, represents the enhanced (super-resolution processed) high-resolution image, that is, the output enhanced image; Represents the input low-resolution image, which is the original image data to be processed; ESRGAN is a deep learning model for image super-resolution tasks, which means that the model is used to process low-resolution images. Processing to obtain high-resolution enhanced images , which reflects the process of using the ESRGAN model to achieve the conversion from low-resolution images to high-resolution enhanced images.

[0065] The three-dimensional scoring system of DoG, entropy, and signal-to-noise ratio (Formula 1-Formula 3) was used to screen valid images (≥70 points). Low-quality images were enhanced using ESRGAN (Formula 4) or processed using a backup camera (RTSP switching, ≤100ms) to ensure data reliability.

[0066] In step 3, feature extraction and integrity judgment are specifically as follows:

[0067] First, CNN (such as ResNet50) extracts features, AHP constructs a weight table (key features ≥ 30%), and calculates the integrity score. The integrity score is calculated using the following formula: Among them, F represents The sum represents the weight of the i-th variable, Represents the value of the i-th variable. The final result is obtained by multiplying the weight of each variable by the value and then summing them.

[0068] Then, it is generated through multi-frame fusion (joining visible features of 5 frames) or GAN (Formula 5, accuracy ≤ 0.5 pixels). The specific calculation formula is:

[0069]

[0070] Where, Indicates the final completed data result after model processing. Generative Adversarial Network (GAN) is an artificial intelligence algorithm framework that uses adversarial training between a generator and a discriminator to achieve data generation and completion. Represents the original data with missing values, which is the part to be completed in the model input. Refers to complete reference data of the same type, distribution, or category as the missing data, which is used to assist GAN in learning data patterns and completing the completion logic.

[0071] In step 4, the multi-model cross-validation is as follows:

[0072] Perform reasoning on at least three heterogeneous models, including but not limited to YOLOv8, FasterRCNN, and ViT in parallel.

[0073] Set the confidence level P to: , where P represents the confidence output after calculation by the Softmax function, is the Softmax activation function, y represents the original score vector output by the model; it passes when the threshold P ≥ 95% and ≥ 2 of the deployed heterogeneous recognition models are consistent.

[0074] For disagreement resolution, arbitration is performed using a fine-grained feature library (model differences ≤ 1 cm) or ICP point cloud registration (Formula 6, accuracy ≤ 5 cm). The specific calculation formula is:

[0075] Where, The rotation matrix (describing the rotation in three-dimensional space) and the translation vector (describing the translation in space) to be optimized together define the transformation relationship between two point clouds or coordinate systems. argmin represents the variable value that minimizes the following expression, i.e., finding the R and T values ​​that minimize the error sum. are corresponding points, such as points in the source point cloud and the corresponding points in the target point cloud , The point of the target point cloud Use transformation Mapped to the source point cloud coordinate system, Calculate the error, Represents the covariance matrix of the observation noise, reflecting the statistical characteristics of the noise in the observation data, Represents the translation of a point set in space.

[0076] Core features (such as the number of engines and tail fin markings) are extracted using a CNN. A weighted completeness score (≥80) is calculated using the AHP method (pass). Missing features are supplemented using a GAN (Formula 5, accuracy ≤ 0.5 pixel) or multi-frame fusion (5-frame stitching). Three heterogeneous models are inferred in parallel, with Softmax outputting confidence scores (passed if confidence ≥ 95% and ≥ 2 of the deployed heterogeneous recognition models agree). In case of disagreement, a fine-grained feature library (aircraft model differences ≤ 1 cm) or ICP point cloud registration (Formula 6, accuracy ≤ 5 cm) is used to improve model robustness.

[0077] In step 5, the time, space and scene adaptation are specifically as follows:

[0078] Spatial consistency: After RANSAC anti-noise, the least squares method is used to fit the trajectory (deviation ).

[0079] Temporal consistency: sliding window (≥5 frames) detection variance .

[0080] Behavioral consistency: Kalman filtering (Equations 7 and 8) predicts motion (velocity v ≤ baseline, covariance dynamically adjusted).

[0081] Formula 7 is specifically:

[0082] Where, The system state vector at discrete time k+1 reflects the characteristics of the system at that moment; Represents the transfer matrix, which reflects the transfer relationship of the internal state of the system in discrete time steps and determines the current state How to affect the next moment state . Represents the system state vector at discrete time k, which is the state at the previous moment. Represents the input matrix, which describes how the external input affects the system state. Associated with state changes; Represents the control input (or external input) vector at discrete time k, such as the control voltage of the motor, the driving force of the robotic arm, etc., which is a signal applied to the system artificially or externally. Represents the process noise (or disturbance) vector, which represents the inevitable random interference and modeling errors in system operation. It is a random factor that affects the accuracy of the state and is usually assumed to be a random variable that satisfies certain statistical characteristics (such as Gaussian distribution).

[0083] Formula 8 is specifically:

[0084] Where, The observation data at discrete time k is the actual measured data; H represents the observation matrix, which reflects the mapping relationship between the system state and the observation value. The quantity converted into the observation space is determined by the characteristics of the observation equipment and the observation method; represents the system state vector at discrete time k; The observation noise vector at discrete time k is the random error introduced during the observation process.

[0085] Update equation:

[0086] Where, The optimal state estimate at time k is the best guess of the system's true state after integrating predictions and observations. Represents the optimal state estimate at time k-1, which serves as the basis for prediction at the current moment. K represents the Kalman gain, which determines the observed data. and predicted value The weight when updating the state, balancing the impact of prediction error and observation noise, is calculated through covariance; It represents the observation value at time k, reflecting the actual observation data of the system state; H represents the observation matrix.

[0087] in, , where represents the covariance matrix of the system state estimate at time k-1, reflecting the statistical characteristics of the state prediction error. Represents the transpose of the observation matrix H. The observation matrix is ​​used to map the system state space to the observation space. The transpose is a conventional operation of matrix operations, which is adapted to the covariance matrix and the observation noise covariance matrix. The covariance matrix of the observation noise reflects the statistical characteristics of the noise in the observation data.

[0088] In step 6, the scene adaptive threshold verification is specifically to generate a strategy based on scene metadata (peak / low peak, wide body / narrow body) and environmental parameters (lighting, weather) through a decision tree algorithm. The threshold is stored in the configuration point table (remote update ≤50ms).

[0089] In steps 5 and 6, 10 consecutive frames of data are spatiotemporally verified: trajectory deviation after RANSAC noise reduction is ≤ 30 cm (least squares fitting), the variance of the sliding window detection result fluctuation is ≤ the threshold, and Kalman filtering is used to predict motion (speed ≤ baseline, covariance adjusted for the environment). A decision tree dynamically adjusts the threshold (configured point table, remote update ≤ 50ms) based on scene metadata (high / low peak, wide / narrow body) and environmental parameters (lighting, weather), adapting to different scenarios.

[0090] In this invention, misjudgment cases are stored in MongoDB, and PyTorch incremental training (loss function , where It represents the total deviation between the predicted value and the true value. represents the true value, represents the model prediction value, Represents the square of the error of a single sample. The optimized model is deployed on edge nodes (such as Jetson AGX), and the accuracy is updated by +0.5% every week to achieve continuous evolution of the system.

[0091] Example 3

[0092] This embodiment provides a specific implementation method, taking a rainstorm day with the tail partially blocked as an example to perform aircraft model identification.

[0093] Step 1, image acquisition: A 4K camera acquires rainstorm images (original score 60 points, <70 points), triggering ESRGAN enhancement (score 75 points, effective).

[0094] Step 2, feature processing: The tail wing is obscured by the rain curtain, and GAN completes the tail wing logo (completeness score 75 → 85 points).

[0095] Step 3, model verification: YOLOv8 (94%), FasterRCNN (96%), ViT (97%), among the deployed heterogeneous recognition models, ≥ 2 models are consistent, and passed.

[0096] Step 4, spatiotemporal verification: The trajectory deviation for 10 consecutive frames is 20 cm (≤30 cm, which complies with the in-place motion law).

[0097] Step 5, scene adaptation: The threshold for heavy rain scenes is relaxed (matching degree ≥ 98.5%, normal 99%), and the model is output (correct, misjudgment rate 0%).

[0098] Technical advantages: The traditional method has a misjudgment rate of ≥5% in this scenario. This invention reduces it to 0% through multi-level verification, verifying the effectiveness of the solution and improving flight punctuality by ≥2%.

[0099] This embodiment provides a specific deployment method of the system, including the following steps:

[0100] 1. Hardware deployment

[0101] Camera: 4K resolution (3840×2160), multi-camera (≥2 channels), RTSP protocol transmission, deployed high on the apron (height 1.5 to 8.5 meters, angle 30° to 60°), supports anti-vibration bracket (vibration attenuation ≥20dB).

[0102] Edge computing: Runs PyTorch inference models (YOLOv8, FasterRCNN, ViT) and integrates a MEMS accelerometer to monitor vibration in real time (frequency 1050Hz, amplitude 0.52g).

[0103] Storage and training: The MongoDB cluster stores misjudgment cases, the PyTorch cluster performs incremental training weekly, and model parameters are updated to edge nodes via OTA (over-the-air download), with an update time of ≤10 seconds.

[0104] 2. Software Implementation

[0105] Image enhancement: The ESRGAN library (implemented in PyTorch) improves low-quality images, and OpenCV implements multi-frame fusion, supporting real-time processing (≤100ms / frame).

[0106] Features and models: ResNet50 (PyTorch) feature extraction, AHP weight calculation, GAN completion (PyTorch, generation speed ≤ 50ms / frame), heterogeneous model parallel inference (multi-threaded acceleration, inference speed ≤ 30ms / model).

[0107] Spatiotemporal verification: RANSAC (OpenCV) for noise reduction (≤20ms / frame), Kalman filter (Python library, ≤10ms / frame) for prediction, and decision tree (Scikitlearn) for threshold strategy generation (≤50ms / scene switch).

[0108] Iterative optimization: MongoDB CRUD operations (≤10ms / case), PyTorch Lightning for incremental training (≤2 hours / week), and edge node hot updates (≤10 seconds).

[0109] This invention addresses the industry pain points of airport aprons (extreme environments, dynamic drift, multimodal fusion, and scene adaptation) and proposes a multi-level verification and closed-loop iteration solution, achieving fully automated, high-precision scene calibration. Compared with existing technologies, the accuracy of aircraft model recognition has been significantly improved, with the small target detection rate increased by 3.5% compared to existing technologies, and the false alarm rate in extreme environments has also been significantly reduced compared to existing technologies, meeting the safety and efficiency requirements of smart airports. Through hardware and software collaborative design (anti-vibration brackets, edge computing, OTA updates), the system ensures long-term stable operation in complex dynamic scenarios. It has significant technological innovation and industry application value and can be extended to fields such as intelligent transportation and industrial testing.

[0110] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0111] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A scene automatic calibration method based on machine vision, characterized by: The following steps are involved: Step 1, image acquisition: obtain image data of the scene to be identified through a data acquisition device; Step 2, image quality judgment: quantitatively score the collected image data; And judge whether the image is a valid image based on the scoring result. If it is, go to step 3. If not, start the image enhancement mechanism and try again; Step 3, feature extraction and integrity judgment: Extract the target core features from the screened valid images, and judge the validity of the target core features by constructing a feature importance weight table and calculating the integrity score; Step 4, multi-model cross-validation: deploy at least three heterogeneous recognition models to perform parallel inference on the target core features, collect statistics on the inference results of all heterogeneous recognition models, and determine whether the inference results meet the preset values; The inference results of the heterogeneous recognition model are output with confidence through Softmax. A voting mechanism is used to arbitrate the results. When performing fine-grained feature comparison, the model difference library is called to perform pixel-level matching. Specifically: The test passes when the confidence level is ≥95% and ≥2 of the deployed heterogeneous recognition models are consistent. When there is disagreement or low confidence <80%, fine-grained feature library arbitration or lidar point cloud registration is called; Multimodal fusion uses a point cloud image registration algorithm to convert the lidar point cloud into an image coordinate system, which is then integrated with visual features and input into the recognition model to improve the detection rate of small targets. Step 5: Spatiotemporal consistency judgment: The recognition results of 10 consecutive frames of valid images are checked for spatial, temporal, and behavioral consistency. If the verification results are consistent, the process proceeds to step 6. If they are inconsistent, the case is marked as an abnormality and triggers manual review. Step 6, scenario-adaptive threshold verification: Dynamically adjust the recognition threshold based on the business scenario and environmental parameters. The recognition result meets the current scenario threshold and outputs the final recognition result. If it does not meet the threshold, it is marked as an abnormal case and triggers manual review.

2. The method for automatic scene calibration based on machine vision according to claim 1, characterized in that: The image quality assessment algorithm in step 2 is based on the three-dimensional indicators of clarity, signal-to-noise ratio, and exposure, and uses Gaussian blur difference and entropy calculation to achieve quantitative scoring. The images after the quantitative scoring are screened, and images greater than or equal to the threshold are identified as valid images, and images less than the threshold are identified as low-quality images, triggering image enhancement or backup camera switching and re-calibration.

3. The method for automatic scene calibration based on machine vision according to claim 1, characterized in that: In step 3, the feature importance weight table weights the core features, extracts features through a convolutional neural network, and calculates the completeness score. Missing features are retried after multi-frame fusion or GAN completion.

4. The method for automatic scene calibration based on machine vision according to claim 1, characterized in that: The spatiotemporal consistency check in step 5 includes: Spatial consistency: target position trajectory fitting deviation ≤ 30cm; Temporal consistency: result fluctuation variance ≤ threshold; Behavioral consistency: The target motion complies with physical laws, and the motion state is predicted through Kalman filtering.

5. The method for automatic scene calibration based on machine vision according to claim 1, characterized in that: In step 6, the dynamic threshold adjustment model generates a scene threshold strategy through a decision tree algorithm based on scene metadata and environmental sensor data.

6. The method for automatic scene calibration based on machine vision according to claim 1, characterized in that: In step 5, after being marked as an abnormal case, it is manually reviewed and placed in the misjudgment case library for iteration of the model.

7. The method for automatic scene calibration based on machine vision according to claim 1, characterized in that: In step 2, the image enhancement retry mechanism uses a super-resolution algorithm to improve the clarity of low-quality images, and the backup camera switching realizes multi-camera redundancy through the RTSP protocol.

8. The method for automatic scene calibration based on machine vision according to claim 1, characterized in that: In step 3, the feature completion model is based on the generative adversarial network, which predicts the missing part through the feature distribution of similar models and generates the completed feature map.

Images