An intelligent detection system and method for track foreign matter based on image recognition

Abstract

The present invention discloses an intelligent track foreign object detection system and method based on image recognition, which specifically relates to the field of image recognition technology. The system obtains real-time images and extracts regions of interest in combination with image segmentation algorithms to identify potential foreign objects. The system analyzes the spatiotemporal data of track foreign objects to verify the spatiotemporal consistency of detection results to determine real threats. Multidimensional features such as shape, texture, color, and size are extracted for real threat targets, and classification confidence is output to evaluate model classification stability. The system evaluates detection accuracy by combining spatiotemporal consistency and classification stability, classifies detection results into different categories, and predicts the degree of abnormality of subsequent detection based on incomplete accuracy detection results. The system dynamically adjusts the image acquisition resolution or frame rate to improve detection clarity. The system solves the problems of missed detection of small and medium-sized foreign objects and frequent false alarms in the prior art, significantly improves the accuracy and reliability of track foreign object detection, and ensures the safety of high-speed train operation.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to an intelligent detection system and method for track foreign matter based on image recognition. Background Art

[0002] Image recognition-based intelligent track foreign object detection uses image recognition technology and intelligent algorithms to analyze images of the track to identify and determine whether there are any foreign objects on the track that could affect traffic safety. This technology typically combines high-resolution cameras, sensors, and artificial intelligence models (such as deep learning) to achieve real-time monitoring and anomaly detection, ensuring smooth and safe track operation. For example, in high-speed rail or subway systems, cameras installed along the track continuously capture track images. When foreign objects (such as fallen tools, debris, branches, or other obstacles) enter the track, the system quickly identifies these anomalies using a trained deep learning model and triggers an alarm, notifying relevant personnel to clean up the scene or take emergency measures. For example, during railway operation in a certain area, the system detected a fallen steel plate on the track. Using an image recognition model, it promptly reported the problem, preventing a train derailment.

[0003] The existing technology has the following shortcomings:

[0004] Intelligent detection systems may fail to detect small foreign objects or objects with a color similar to the track background (such as bolts trapped in rusty rails) due to the high concealment of the target or the limited camera angle. Although these small foreign objects are not large in size, they can pose a major safety hazard during high-speed train operation, such as the risk of derailment. For example, a black bolt dropped on a dark rail may be mistaken for a normal track due to its similar texture to the background and thus missed. In addition, if the image recognition algorithm is inaccurate in foreign object detection and mistakenly identifies harmless objects (such as leaves and plastic bags) as dangerous foreign objects, resulting in frequent system alarms, this will not only increase maintenance costs but may also reduce staff's trust in the alarms, causing them to ignore real threats. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent detection system and method for track foreign matter based on image recognition to solve the shortcomings of the background technology.

[0006] In order to achieve the above object, the present invention provides the following technical solution: a method for intelligent detection of foreign objects on a track based on image recognition, comprising the following steps:

[0007] S1: High-definition cameras installed along the track are used to acquire real-time image data of the track area and perform preprocessing on the acquired image data, including denoising, contrast enhancement, and color correction.

[0008] S2: Segmenting the track area using an image segmentation algorithm to extract regions of interest within the track area, performing target detection on the extracted regions of interest based on a deep learning model including a convolutional neural network to identify potential track foreign objects;

[0009] S3: Obtain the spatiotemporal data of track foreign objects in different frames, analyze the position changes and movement directions of track foreign objects in consecutive frames, verify the spatiotemporal consistency of track foreign object detection results, and determine whether the detected target is a real threat;

[0010] S4: After determining that the detected target is a real threat, extract the multi-dimensional features of the detected foreign object, including shape, texture, color, and size. Use the deep learning model to output the classification confidence, analyze the proportion of low-confidence targets, and evaluate the classification stability of the deep learning model algorithm.

[0011] S5: Evaluate the accuracy of the image recognition algorithm for track foreign object detection based on the spatiotemporal consistency of the track foreign object detection results and the classification stability of the deep learning model algorithm. Based on the evaluation results, classify the track foreign object detection results into accurate detection results, incompletely accurate detection results, and inaccurate detection results.

[0012] S6: For incomplete accuracy detection results, predict the degree of abnormality in the accuracy of the image recognition algorithm for track foreign object detection in the subsequent detection time period. If the degree of abnormality is high, increase the resolution or frame rate of image acquisition to enhance the clarity of the track image.

[0013] Preferably, in S2, the track area is segmented using an image segmentation algorithm, and the rail edges and straight line structures are extracted through edge detection and Hough transform to distinguish the left and right sides of the track; a semantic segmentation model is used to classify each pixel in the image, separate the track area from the background area, and annotate the pixel labels of the track area for the model to learn track features; based on the segmentation results, the image portion of the track area is retained;

[0014] Define the region of interest (ROI) within the track area and delineate the ROI boundary based on the geometric characteristics of track width and spacing. Use a convolutional neural network model to perform target detection on the extracted ROI area. Select the input track area image and ROI label, and train the model based on the annotated foreign object location and category. Use classification loss and regression loss to optimize the model. Output the foreign object's bounding box coordinates, classification label, and confidence score.

[0015] Preferably, in S3, the position change and movement direction of the track foreign matter in consecutive frames are analyzed to generate a foreign matter spatiotemporal consistency index, and the method for obtaining the foreign matter spatiotemporal consistency index is:

[0016] Set the target trajectory sequence to , represents the position sequence of the detected track foreign objects in consecutive frames, and sets the reference trajectory sequence , represents the standard motion pattern of an orbiting object; the distance between two trajectory points is calculated using the Euclidean distance , the expression is: ;in, is the position coordinate of the target trajectory point, For the position coordinates of the reference trajectory points, construct the cumulative distance matrix D: ;in, From the starting point of the trajectory to The cumulative minimum distance of the points, is the Euclidean distance between the current points, Represents the optimal path selection from the previous position to the current point, and the boundary conditions are: ; ; ; In the cumulative distance matrix, find the optimal path P from the starting point (1,1) to the end point (m,n), the path consists of a series of points Composition, satisfying: ; It means that the path P is selected so that the cumulative distance of all points on the path is minimized, and the spatiotemporal consistency index of foreign objects is calculated. The expression is: ; Where L is the length of the optimal path, is the Euclidean distance on the path, and HAK is the spatiotemporal consistency index of foreign objects.

[0017] Preferably, in S4, the low confidence target proportion fluctuation index is generated after analyzing the proportion of low confidence targets. The low confidence target proportion fluctuation index is obtained as follows:

[0018] Collect a time series of the proportion of low-confidence targets ; Indicates the proportion of low-confidence targets in the T-th frame image, where T is the total number of frames in the time series. The time series R is segmented into w intervals: ; Indicates the boundary of the wth interval; counts the number of data points in each interval , the expression is: ; Calculate the probability of the i-th interval: ; is the probability that the low confidence ratio falls into the i-th interval, and the low confidence ratio target fluctuation index is calculated. The expression is: ; Where GXH is the fluctuation index of the proportion of low-confidence targets.

[0019] Preferably, in S5, the accuracy of the image recognition algorithm for detecting track foreign objects is evaluated based on the spatiotemporal consistency of the track foreign object detection results and the classification stability of the deep learning model algorithm;

[0020] The foreign body spatiotemporal consistency index and the low-confidence target proportion fluctuation index are converted into comprehensive feature vectors, and the comprehensive feature vectors are used as the input of the machine learning model. The machine learning model uses each set of comprehensive feature vectors to predict the accuracy value label of the image recognition algorithm for track foreign body detection as the prediction target, and takes minimizing the sum of the prediction errors of the accuracy value labels of all image recognition algorithms for track foreign body detection as the training target. The machine learning model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The accuracy value of the image recognition algorithm for track foreign body detection is determined according to the model output results. Among them, the machine learning model is a polynomial regression model.

[0021] Preferably, in S5, based on the evaluation results, the track foreign object detection results are divided into accuracy detection results, incomplete accuracy detection results and inaccuracy detection results, specifically:

[0022] Comparing the acquired accuracy value of the image recognition algorithm for detecting foreign objects on the track with a gradient standard threshold, wherein the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and comparing the accuracy value of the image recognition algorithm for detecting foreign objects on the track with the first standard threshold and the second standard threshold respectively;

[0023] If the accuracy value of the image recognition algorithm for detecting foreign objects on the track is greater than the second standard threshold, it indicates that the image recognition algorithm has high accuracy in detecting foreign objects on the track. In this case, a high-accuracy detection signal is generated, and the track foreign object detection result is classified as an accuracy detection result.

[0024] If the accuracy value of the image recognition algorithm for detecting foreign objects on the track is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it means that the accuracy of the image recognition algorithm for detecting foreign objects on the track is moderate, and a moderate accuracy detection signal is generated, and the track foreign object detection result is classified as an incomplete accuracy detection result;

[0025] If the accuracy value of the image recognition algorithm for detecting foreign objects on the track is less than the first standard threshold, it means that the accuracy of the image recognition algorithm for detecting foreign objects on the track is low. At this time, a low-accuracy detection signal is generated, and the track foreign object detection result is classified as an inaccurate detection result.

[0026] Preferably, in S6, for the incomplete accuracy detection result, the abnormal degree of accuracy of the image recognition algorithm for detecting foreign objects on the track in the subsequent detection time period is predicted, specifically:

[0027] For incomplete accuracy detection results, that is, the accuracy value of the image recognition algorithm for track foreign object detection generated within a fixed detection time period is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, the accuracy values ​​greater than or equal to the first standard threshold and less than or equal to the second standard threshold in the subsequent detection time period are collected, and a corresponding data set is established. The mean and standard deviation of the data set are calculated, and after analysis, the degree of abnormality in the accuracy of the image recognition algorithm for track foreign object detection in the subsequent detection time period is predicted based on the analysis results.

[0028] Preferably, if the mean accuracy value in the data set is greater than or equal to the reference threshold value of the mean accuracy value, and the standard deviation of the accuracy value is less than the reference threshold value of the standard deviation of the accuracy value, the detection result is stable and accurate, and no warning signal is generated at this time, no adjustment is required, and the existing detection settings can be maintained;

[0029] If the mean accuracy value is greater than or equal to the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is greater than or equal to the reference threshold of the standard deviation of the accuracy value, the detection accuracy fluctuates greatly, and a third-level warning signal is generated at this time, and the algorithm parameters are adjusted or the data is increased;

[0030] If the mean accuracy value is less than the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is greater than or equal to the reference threshold of the standard deviation of the accuracy value, the detection accuracy is low and fluctuates significantly. In this case, a level 1 warning signal is generated and improvement measures should be taken, including optimizing the model or improving the image acquisition quality.

[0031] If the mean accuracy value is less than the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is less than the reference threshold of the standard deviation of the accuracy value, the detection accuracy is low. At this time, a secondary warning signal is generated, and the model should be optimized or the algorithm should be switched.

[0032] The present invention also provides an intelligent detection system for track foreign objects based on image recognition, which includes an image acquisition and preprocessing module, an image segmentation and target detection module, a spatiotemporal consistency analysis module, a classification stability assessment module, a detection result assessment and classification module, and a dynamic optimization and anomaly prediction module;

[0033] Image acquisition and preprocessing module: This module uses high-definition cameras installed along the track to acquire real-time image data of the track area and performs preprocessing on the acquired image data, including denoising, contrast enhancement, and color correction.

[0034] Image segmentation and object detection module: Uses image segmentation algorithms to segment the track area, extracts regions of interest within the track area, and performs object detection on the extracted regions of interest based on a deep learning model that includes a convolutional neural network to identify potential track foreign objects.

[0035] Spatiotemporal consistency analysis module: This module obtains the spatiotemporal data of track foreign objects in different frames and analyzes the position changes and movement directions of track foreign objects in consecutive frames to verify the spatiotemporal consistency of the track foreign object detection results and determine whether the detected target is a real threat.

[0036] Classification Stability Assessment Module: After determining that the detected target is a real threat, it extracts multi-dimensional features of the detected foreign object, including shape, texture, color, and size. The deep learning model outputs the classification confidence, analyzes the proportion of low-confidence targets, and evaluates the classification stability of the deep learning model algorithm.

[0037] Detection result evaluation and classification module: This module evaluates the accuracy of the image recognition algorithm for detecting track foreign objects based on the spatiotemporal consistency of the detection results and the classification stability of the deep learning model algorithm. Based on the evaluation results, the detection results are classified into accurate detection results, incompletely accurate detection results, and inaccurate detection results.

[0038] Dynamic optimization and anomaly prediction module: For incomplete accuracy detection results, the accuracy abnormality of the image recognition algorithm for track foreign object detection in the subsequent detection time period is predicted. If the abnormality is high, the resolution or frame rate of image acquisition is increased to enhance the clarity of the track image.

[0039] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0040] 1. This invention achieves efficient detection and classification of foreign objects within the track area by combining image segmentation algorithms with deep learning models. To address the issues of small foreign objects being highly concealed and the complex track background, the invention introduces a foreign object spatiotemporal consistency index and a low-confidence target ratio fluctuation index to comprehensively analyze the spatiotemporal characteristics and classification stability of the detected targets, thereby improving the detection accuracy of small targets and those with similar backgrounds. A polynomial regression model further converts multidimensional feature vectors into detection accuracy predictions. Combined with dynamic adjustment of image acquisition resolution or frame rate, this model effectively addresses fluctuations in detection performance and ensures the real-time and adaptability of the system.

[0041] 2. This invention significantly reduces the problems of missed detection of small foreign objects and false alarms of harmless objects, improving the accuracy and reliability of track foreign object detection. Based on accuracy assessment, the detection results are classified and processed, and an early warning mechanism promptly identifies abnormal detection situations, providing targeted optimization strategies to ensure the long-term stability of the system. By dynamically optimizing image acquisition and deep learning models, the system's adaptability to complex track environments is improved, significantly reducing safety hazards and operating and maintenance costs, and providing a solid guarantee for rail transit operation safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0043] Figure 1 Flow chart of the method of the present invention.

[0044] Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0046] Example 1, please refer to Figure 1 As shown, the intelligent detection method for track foreign objects based on image recognition described in this embodiment includes the following steps:

[0047] S1: High-definition cameras installed along the track are used to acquire real-time image data of the track area and perform preprocessing on the acquired image data, including denoising, contrast enhancement, and color correction.

[0048] S2: Segmenting the track area using an image segmentation algorithm to extract regions of interest within the track area, performing target detection on the extracted regions of interest based on a deep learning model including a convolutional neural network to identify potential track foreign objects;

[0049] S3: Obtain the spatiotemporal data of track foreign objects in different frames, analyze the position changes and movement directions of track foreign objects in consecutive frames, verify the spatiotemporal consistency of track foreign object detection results, and determine whether the detected target is a real threat;

[0050] S4: After determining that the detected target is a real threat, extract the multi-dimensional features of the detected foreign object, including shape, texture, color, and size. Use the deep learning model to output the classification confidence, analyze the proportion of low-confidence targets, and evaluate the classification stability of the deep learning model algorithm.

[0051] S5: Evaluate the accuracy of the image recognition algorithm for track foreign object detection based on the spatiotemporal consistency of the track foreign object detection results and the classification stability of the deep learning model algorithm. Based on the evaluation results, classify the track foreign object detection results into accurate detection results, incompletely accurate detection results, and inaccurate detection results.

[0052] S6: For incomplete accuracy detection results, predict the degree of abnormality in the accuracy of the image recognition algorithm for track foreign object detection in the subsequent detection time period. If the degree of abnormality is high, increase the resolution or frame rate of image acquisition to enhance the clarity of the track image.

[0053] In S1, use high-resolution cameras (e.g., 4K or higher) to ensure that track details, such as the outline and texture of small foreign objects, are captured. Select cameras with wide dynamic range (WDR) to accommodate complex lighting conditions (such as shadows and direct sunlight). Install cameras at intervals along the track to ensure full coverage. Adjust the camera angle to ensure a panoramic view of the track area and avoid blind spots. The acquisition frequency must match the train speed; for example, high-speed rail can use 30 frames per second or higher to ensure that track details are not missed.

[0054] Image preprocessing removes noise from the image caused by sensor or environmental factors, improving image quality. Common noise types include: Gaussian noise: caused by random variations in the camera sensor. Illumination noise: caused by reflections or uneven light sources from the surrounding environment. Gaussian filtering: uses a Gaussian smoothing algorithm to reduce high-frequency noise while preserving image edge information. Median filtering: suitable for removing salt and pepper noise (such as bright or dark spots) and protecting the sharpness of track edges. Non-local mean (NL-Means) filtering: performs a weighted average of similar pixel blocks to effectively remove noise while preserving detail.

[0055] Improves image clarity in the track area, making target foreign objects (such as bolts and gravel) more visible in the image. The track area may have low contrast, such as insufficient lighting or an overly complex background. Enhances the brightness contrast in the image by redistributing pixel intensity values. Applicable to situations where the overall brightness distribution in the track area is uniform but the details are not obvious. Adaptive Histogram Equalization (CLAHE): Processes the image in blocks to enhance local contrast and avoid noise amplification caused by over-enhancement. Applicable to situations where the track background is complex and the lighting is uneven. Gamma Correction: Enhances dark details or reduces overexposure in highlight areas by adjusting the gamma value (non-linear transformation). Commonly used to optimize track details in shadow areas.

[0056] Correct color deviations in images to make them closer to the real scene and ensure that the color characteristics of foreign objects (such as rust and plastic bags) are displayed correctly. The camera may cause image color distortion under different light conditions (such as daylight, shadows, and night lights). White balance correction: Adjust the white point in the image to make the image color distribution more natural. For example, by adjusting the ratio of the RGB channels, the color deviation caused by different light sources can be corrected. Color mapping: Use a reference color card to map the color values ​​in the image to a standard color space. Applicable to track monitoring environments with color reference benchmarks. Color smoothing: Apply a filtering algorithm to eliminate local color abnormalities in the image to make the color distribution of the track and foreign objects more uniform.

[0057] In this application, using a high-definition camera to capture and perform real-time track image denoising, contrast enhancement, and color correction can significantly improve image quality and the accuracy of foreign object identification, providing clear and reliable image data for subsequent target detection and classification. These preprocessing steps are the key foundation of the intelligent track foreign object detection system.

[0058] S2: Segment the track area using an image segmentation algorithm, extract regions of interest within the track area, perform target detection on the extracted regions of interest based on a deep learning model, and identify potential track foreign objects. The deep learning model includes a convolutional neural network.

[0059] Use image segmentation algorithms to segment the track area, separating the track area (such as rails and sleepers) from the background area (such as the ground and grass), and clearly mark the track area as the focus of analysis. The specific steps include:

[0060] Use Canny edge detection or the Sobel operator to extract track edges. Detect the straight lines and edge contours of the track in the preprocessed image. Use the Hough transform to detect the straight lines of the track. Fit the detected edge points into straight lines to distinguish the left and right sides of the track. Segment the track area based on region growing or threshold segmentation. Use a semantic segmentation model (such as U-Net or DeepLab) to classify each pixel in the image and distinguish the track area from the background. Label the pixel labels of the track area during training so that the model can learn the track characteristics. Based on the segmentation results, retain only the image portion of the track area and remove irrelevant background.

[0061] Extract the region of interest (ROI) within the track area. The region of interest (ROI) refers to the key area in the image that may contain the target track foreign matter, and the system uses it as the focus of subsequent detection. Examples include the track surface, track joints, and the area between the rails and sleepers. Combined with the track segmentation results, the ROI boundaries are defined using the geometric characteristics of the track width and spacing. For example, set the area of ​​a certain range (such as 10 pixels) on both sides of the rail as the ROI. Dynamically adjust the ROI range according to the track environment, such as expanding the ROI in the track curve area to cover more possible foreign matter distribution areas. Analyze historical track foreign matter detection data, focus on areas where foreign matter often appears, and optimize the ROI range.

[0062] Utilize deep learning models to detect objects within the ROI. Convolutional Neural Network (CNN) models include: Object Detection Model: YOLO (You Only Look Once): Offers strong real-time detection performance, suitable for the high-impact requirements of track foreign object detection. Faster R-CNN: High-precision object detection algorithm, suitable for scenarios requiring precise detection. SSD (Single Shot Detector): Balances speed and accuracy, suitable for multi-object scenarios.

[0063] Collect a large amount of track foreign object data, including various environments (sunny, rainy) and foreign object types (such as bolts, gravel, tools).

[0064] Label the location (bounding box) and category (e.g., tool, plastic bag) of foreign objects.

[0065] Input: Track region image and ROI label.

[0066] Loss function: Use classification loss (cross entropy) and regression loss (L1 or L2 loss) to optimize the model.

[0067] Data enhancement: Simulate various real-world environments through rotation, flipping, and lighting changes to enhance model robustness.

[0068] The extracted ROI area is used as model input for target detection.

[0069] Output detection results: Location: The bounding box coordinates of the foreign object. Category: The classification label of the foreign object (such as tool, gravel). Confidence: The confidence probability of each detection result, indicating the model's confidence in the detection result.

[0070] Improve the detection capability of small targets (such as bolts and gravel) through multi-scale feature extraction (such as FPN). Combine the detection confidence and preset threshold to screen out highly reliable targets as potential foreign objects. For example, if the detection confidence is greater than the preset threshold, the target is considered a potential foreign object.

[0071] If the same foreign object is detected in multiple frames, a spatiotemporal fusion method is used to improve the accuracy of the detection results. Output: The foreign object's location, category, and confidence level are visually displayed overlaid on the track image. An alarm message is sent to the monitoring center, marking the specific danger points in the track area.

[0072] S3: Obtain the spatiotemporal data of track foreign objects in different frame images, analyze the position changes and movement directions of track foreign objects in consecutive frames, verify the spatiotemporal consistency of track foreign object detection results, and determine whether the detected target is a real threat.

[0073] Obtain spatiotemporal data on track foreign objects, including: Time data: refers to the timestamps of the foreign object's appearance in different frames of the image. Spatial data: refers to the coordinate position of the foreign object in the image, including 2D image coordinates (such as x, y) and possible depth information (z, if using LiDAR or stereo cameras).

[0074] In consecutive frames, objects are correlated across frames based on their appearance features (such as shape, texture, and color) or proximity. Kalman filtering or optical flow is used to track the object's trajectory. Only detected objects within the trajectory area are retained, eliminating interference from non-trajectory areas (such as dynamic background objects). Each frame records the target's timestamp, spatial coordinates, and detection confidence, forming a spatiotemporal data set for the foreign object.

[0075] Analyze the position changes and movement direction of foreign objects on the track in consecutive frames, and use the coordinate positions of the foreign objects in consecutive frame images to calculate the motion trajectory. If the position change in consecutive frames is close to 0, it is judged to be a stationary target (such as a bolt or gravel). If the position change changes significantly and the direction changes continuously over time, it may be a dynamic target (such as a plastic bag). Use trajectory curve fitting (such as the least squares method) to evaluate whether the motion trajectory is smooth and reasonable. Unreasonable situations: sudden changes in position and large jumps in direction. Classification of trajectory characteristics: Stationary trajectory: a stable target on the track, such as a fallen tool. Regular trajectory: a leaf or plastic bag moving with the wind, the motion trajectory is random but continuous. Abnormal trajectory: the target has a sudden change in position or does not conform to the laws of physics (possibly a false detection).

[0076] After analyzing the position change and movement direction of track foreign objects in consecutive frames, the foreign object spatiotemporal consistency index is generated. According to the foreign object spatiotemporal consistency index, it is judged whether the detected target is a real threat. The method for obtaining the foreign object spatiotemporal consistency index is as follows:

[0077] Set the target trajectory sequence to , represents the position sequence of the detected track foreign objects in continuous frames (such as two-dimensional coordinates (x, y)), setting the reference trajectory sequence , which represents the possible standard motion mode of the orbital foreign body (such as stationary, regular movement). The distance between two trajectory points is calculated using the Euclidean distance. , the expression is: ;in, is the position coordinate of the target trajectory point, For the position coordinates of the reference trajectory points, construct the cumulative distance matrix D: ;in, From the starting point of the trajectory to The cumulative minimum distance of the points, is the Euclidean distance between the current points, Represents the optimal path selection from the previous position to the current point, and the boundary conditions are: ; ; ; In the cumulative distance matrix, find the optimal path P from the starting point (1,1) to the end point (m,n), the path consists of a series of points Composition, satisfying: ; It means that the path P is selected so that the cumulative distance of all points on the path is minimized, and the spatiotemporal consistency index of foreign objects is calculated. The expression is: ; Where L is the length of the optimal path, is the Euclidean distance on the path, and HAK is the spatiotemporal consistency index of foreign objects.

[0078] A low spatiotemporal consistency index for a foreign object indicates a high degree of match between the detected trajectory and the reference trajectory, a smooth trajectory, and consistent motion patterns with the actual scenario. For example, a stationary foreign object (such as a bolt or gravel) may maintain a nearly constant position on the track for multiple consecutive frames, or a slowly moving foreign object (such as a sliding tool) may have a linear trajectory with a consistent direction. In these cases, the detected object exhibits high spatiotemporal consistency and is therefore considered a real threat, as such objects often pose a real risk to track safety.

[0079] A high spatiotemporal consistency index for a foreign object indicates a poor match between its trajectory and the reference trajectory, an uneven trajectory, or unusual variations. This often indicates that the detected object is a dynamic interference object (such as a plastic bag fluttering in the wind or a fast-moving shadow). Its trajectory may exhibit irregular jumps, sudden changes in direction, or dramatic changes in speed. In this case, the detected object exhibits low spatiotemporal consistency and is not considered a true threat, so the alarm can be suppressed to avoid false positives.

[0080] S4: After determining that the detection target is a real threat, extract the multi-dimensional features of the foreign object, including shape, texture, color, and size. Output the classification confidence through the deep learning model, analyze the proportion of low-confidence targets, and evaluate the classification stability of the deep learning model algorithm.

[0081] After confirming that the detected target is a real threat, multidimensional features are extracted for detecting foreign objects. These include shape features, which describe the geometric structure of the foreign object, such as edges, outlines, and corners. Shape feature vectors (such as roundness and aspect ratio) are extracted using Hough transforms and edge detection (e.g., Canny or Sobel). Texture features describe the details and structural patterns on the foreign object's surface, such as roughness and smoothness. Texture features of the foreign object are calculated using local binary patterns (LBP) and gray-level co-occurrence matrices (GLCMs). Color features describe the color distribution of the foreign object and its contrast with the background. A color histogram is extracted to analyze the distribution of each channel in RGB, HSV, or Lab color space. Size features describe the physical size of the foreign object and the area it occupies. Size features (e.g., normalized area) are calculated based on the length, width, and area of ​​the bounding box. Shape, texture, color, and size features are combined into a feature vector, which serves as input to the deep learning model.

[0082] Choose a convolutional neural network (CNN)-based deep learning model, such as YOLO, Faster R-CNN, or SSD, for foreign object classification.

[0083] Input features: The extracted feature vector is input into the deep learning model and classified through the feature extraction layer (convolutional layer) and classification layer (fully connected layer).

[0084] Output: The category probability of each target, indicating the confidence that the target belongs to a certain category.

[0085] Classification confidence S is the model's assessment of the reliability of the detection results, and its value range is [0,1].

[0086] Set the confidence threshold T (e.g., T=0.5) and define targets with classification confidence lower than T as low-confidence targets. Set the total number of detected targets to N and the number of low-confidence targets to Nlow.

[0087] The proportion of low confidence targets R: The higher the R value, the more unstable the model classification result or the higher the classification uncertainty. The lower the R value, the more reliable the model classification result and the stronger the classification stability.

[0088] After analyzing the proportion of low-confidence targets, a low-confidence target proportion fluctuation index is generated. Based on the low-confidence target proportion fluctuation index, the classification stability of the deep learning model algorithm is evaluated. The method for obtaining the low-confidence target proportion fluctuation index is as follows:

[0089] Collect a time series of the proportion of low-confidence targets ; Indicates the proportion of low-confidence targets in the T-th frame image (as a percentage). T is the total number of frames in the time series. The time series R is segmented into w intervals: ; The granularity w of the interval division can be determined based on the data range and application requirements (for example, 5 intervals, 10 intervals).

[0090] Count the number of data points in each interval , the expression is: ; Calculate the probability of the i-th interval: ; is the probability that the low confidence ratio falls into the i-th interval, and the low confidence ratio target fluctuation index is calculated. The expression is: ; In the formula, GXH is the fluctuation index of the proportion of low confidence targets; if the probability of a certain interval is ,but: .

[0091] When the fluctuation index of the low-confidence target ratio is small, it indicates that the time series distribution of the low-confidence target ratio is concentrated and has little variation, and the system's classification results remain consistent across consecutive detection frames. In other words, the deep learning model demonstrates strong robustness and stability in classification tasks, maintaining high classification confidence across diverse environments and conditions, and minimizing the risk of false positives and missed detections. This demonstrates the model's high reliability in extracting and classifying target features, making it suitable for scenarios requiring high real-time performance and accuracy.

[0092] A high fluctuation index for the low-confidence target ratio indicates a dispersed and highly variable time series distribution of the low-confidence target ratio, leading to inconsistent classification results across consecutive detection frames. This reflects the significant uncertainty inherent in deep learning models in classification tasks, which can be significantly affected by external interference (such as lighting variations and background complexity), leading to increased false or missed detections. This indicates poor model classification stability, requiring algorithm optimization, enhanced training data, or improved feature extraction methods to enhance reliability and accuracy.

[0093] S5: Based on the spatiotemporal consistency of the track foreign object detection results and the classification stability of the deep learning model algorithm, the accuracy of the image recognition algorithm for track foreign object detection is evaluated. Based on the evaluation results, the track foreign object detection results are divided into accurate detection results, incompletely accurate detection results, and inaccurate detection results.

[0094] The foreign body spatiotemporal consistency index and the low-confidence target proportion fluctuation index are converted into comprehensive feature vectors, and the comprehensive feature vectors are used as the input of the machine learning model. The machine learning model uses each set of comprehensive feature vectors to predict the accuracy value label of the image recognition algorithm for track foreign body detection as the prediction target, and takes minimizing the sum of the prediction errors of the accuracy value labels of all image recognition algorithms for track foreign body detection as the training target. The machine learning model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The accuracy value of the image recognition algorithm for track foreign body detection is determined according to the model output results. Among them, the machine learning model is a polynomial regression model.

[0095] The accuracy value of the image recognition algorithm for detecting foreign objects on the track is obtained by obtaining the corresponding function expression from the comprehensive feature vector training data of the trained machine learning model: Where, is the output function of the model, HAK is the spatiotemporal consistency index of foreign bodies, GXH is the fluctuation index of the proportion of low-confidence targets, is the accuracy value of the image recognition algorithm for detecting foreign objects on the track.

[0096] Comparing the acquired accuracy value of the image recognition algorithm for detecting foreign objects on the track with a gradient standard threshold, wherein the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and comparing the accuracy value of the image recognition algorithm for detecting foreign objects on the track with the first standard threshold and the second standard threshold respectively;

[0097] If the accuracy value of the image recognition algorithm for detecting foreign objects on the track is greater than the second standard threshold, it indicates that the image recognition algorithm has high accuracy in detecting foreign objects on the track. In this case, a high-accuracy detection signal is generated, and the foreign object detection result is classified as an accuracy detection result, without any further processing.

[0098] If the accuracy value of the image recognition algorithm for detecting foreign objects on the track is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it means that the accuracy of the image recognition algorithm for detecting foreign objects on the track is moderate. In this case, a moderate accuracy detection signal is generated, and the foreign object detection result is classified as an incomplete accuracy detection result. There may be a risk of false detection or missed detection, and further verification or review is required.

[0099] If the accuracy value of the image recognition algorithm for detecting foreign objects on the track is less than the first standard threshold, it means that the image recognition algorithm has low accuracy in detecting foreign objects on the track. At this time, a low-accuracy detection signal is generated, and the track foreign object detection result is classified as an inaccurate detection result. The result may be a false detection or insufficient model performance, and the model needs to be optimized or the algorithm needs to be replaced.

[0100] When the image recognition algorithm's accuracy for detecting track foreign objects falls below the first standard threshold, manifested by significant fluctuations in classification results and frequent false positives and missed detections, this indicates that the model's adaptability to feature extraction or classification tasks is insufficient. In this case, the existing model can be optimized to improve its robustness and accuracy by increasing training data (especially for complex scenarios or anomalous samples), improving the model architecture (such as by introducing a multi-layer attention mechanism), optimizing hyperparameters (such as the learning rate and regularization coefficient), and employing data augmentation techniques.

[0101] If the existing model still fails to meet detection requirements after multiple optimizations, or performs poorly in specific scenarios (such as extreme lighting conditions or when there are many dynamic interference objects), it is necessary to replace the algorithm with a more appropriate one. For example, you can try more advanced object detection models (such as YOLOv8 and DETR), or introduce multimodal algorithms that combine images and other sensor data (such as lidar and thermal imaging) for detection to fundamentally improve detection performance.

[0102] S6: For incomplete accuracy detection results, predict the degree of abnormality in the accuracy of the image recognition algorithm for track foreign object detection in the subsequent detection time period. If the degree of abnormality is high, increase the resolution or frame rate of image acquisition to enhance the clarity of the track image.

[0103] For incomplete accuracy detection results, that is, the accuracy value of the image recognition algorithm for track foreign object detection generated within a fixed detection time period is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, the accuracy values ​​greater than or equal to the first standard threshold and less than or equal to the second standard threshold in the subsequent detection time period are collected, and a corresponding data set is established. The mean and standard deviation of the data set are calculated, and after analysis, the degree of abnormality in the accuracy of the image recognition algorithm for track foreign object detection in the subsequent detection time period is predicted based on the analysis results.

[0104] If the mean accuracy value in the data set is greater than or equal to the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is less than the reference threshold of the standard deviation of the accuracy value, the detection result is stable and accurate. At this time, no warning signal is generated, indicating that the system performance is good and no adjustment is required. The existing detection settings can be maintained.

[0105] If the mean accuracy value is greater than or equal to the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is greater than or equal to the reference threshold of the standard deviation of the accuracy value, the detection accuracy fluctuates greatly, but the overall accuracy is acceptable. At this time, a third-level warning signal is generated, indicating that the system detection is not stable enough and may be affected by the external environment (such as changes in lighting and interference). The volatility problem should be optimized, such as adjusting algorithm parameters or adding data.

[0106] If the mean accuracy value is less than the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is greater than or equal to the reference threshold of the standard deviation of the accuracy value, the detection accuracy is low and fluctuates significantly. At this time, a first-level warning signal is generated, indicating that the system performance is poor and improvement measures should be taken, including optimizing the model or improving the image acquisition quality (such as increasing the resolution or frame rate).

[0107] If the mean accuracy value is less than the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is less than the reference threshold of the standard deviation of the accuracy value, the detection accuracy is low, but the result is relatively stable. At this time, a secondary warning signal is generated, indicating that the model may not be adaptable enough to specific scenarios, and model optimization or algorithm switching should be performed.

[0108] In this embodiment, track images are acquired in real time by high-definition cameras installed along the track. The images are preprocessed by denoising, contrast enhancement, and color correction. Regions of interest in the track area are extracted using an image segmentation algorithm, and target detection is performed based on a convolutional neural network to identify potential track foreign objects. The position changes and movement directions of track foreign objects in consecutive frames are further analyzed to verify the spatiotemporal consistency of the detection results and determine whether they are real threats. Multidimensional features such as shape, texture, color, and size are extracted for real threat targets, and classification confidence is output. The proportion of low-confidence targets is analyzed to evaluate the classification stability of the model. The accuracy of the image recognition algorithm is evaluated by combining spatiotemporal consistency and classification stability, and the detection results are divided into accuracy detection, incomplete accuracy detection, and inaccuracy detection. For incomplete accuracy detection results, the degree of abnormality is predicted by analyzing the accuracy mean and fluctuation of subsequent detections. If the degree of abnormality is high, the image acquisition resolution or frame rate is increased to enhance the clarity of the track image, thereby improving detection performance and reliability.

[0109] Example 2, please refer to Figure 2 As shown, the track foreign object intelligent detection system based on image recognition described in this embodiment includes an image acquisition and preprocessing module, an image segmentation and target detection module, a spatiotemporal consistency analysis module, a classification stability evaluation module, a detection result evaluation and classification module, and a dynamic optimization and anomaly prediction module;

[0110] Image acquisition and preprocessing module: This module uses high-definition cameras installed along the track to acquire real-time image data of the track area and performs preprocessing on the acquired image data, including denoising, contrast enhancement, and color correction.

[0111] Image segmentation and object detection module: Uses image segmentation algorithms to segment the track area, extracts regions of interest within the track area, and performs object detection on the extracted regions of interest based on a deep learning model that includes a convolutional neural network to identify potential track foreign objects.

[0112] Spatiotemporal consistency analysis module: This module obtains the spatiotemporal data of track foreign objects in different frames and analyzes the position changes and movement directions of track foreign objects in consecutive frames to verify the spatiotemporal consistency of the track foreign object detection results and determine whether the detected target is a real threat.

[0113] Classification Stability Assessment Module: After determining that the detected target is a real threat, it extracts multi-dimensional features of the detected foreign object, including shape, texture, color, and size. The deep learning model outputs the classification confidence, analyzes the proportion of low-confidence targets, and evaluates the classification stability of the deep learning model algorithm.

[0114] Detection result evaluation and classification module: This module evaluates the accuracy of the image recognition algorithm for detecting track foreign objects based on the spatiotemporal consistency of the detection results and the classification stability of the deep learning model algorithm. Based on the evaluation results, the detection results are classified into accurate detection results, incompletely accurate detection results, and inaccurate detection results.

[0115] Dynamic optimization and anomaly prediction module: For incomplete accuracy detection results, the accuracy abnormality of the image recognition algorithm for track foreign object detection in the subsequent detection time period is predicted. If the abnormality is high, the resolution or frame rate of image acquisition is increased to enhance the clarity of the track image.

[0116] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. An intelligent detection method for track foreign objects based on image recognition, characterized by: The following steps are involved: S1: High-definition cameras installed along the track are used to acquire real-time image data of the track area and perform preprocessing on the acquired image data, including denoising, contrast enhancement, and color correction. S2: Segmenting the track area using an image segmentation algorithm, extracting a region of interest within the track area, and performing target detection on the extracted region of interest based on a first deep learning model to identify potential track foreign objects, wherein the first deep learning model includes a convolutional neural network; S3: Obtain the spatiotemporal data of the track foreign object in different frames, analyze the position change and movement direction of the track foreign object in consecutive frames, verify the spatiotemporal consistency of the track foreign object detection results, and determine whether the detected target is a real threat based on the spatiotemporal consistency of the foreign object; Specifically, the foreign object temporal and spatial consistency index is generated after analyzing the position change and movement direction of the foreign object in the continuous frames. The method for obtaining the foreign object temporal and spatial consistency index is as follows: set the target trajectory sequence as , represents the position sequence of the detected track foreign objects in consecutive frames, and sets the reference trajectory sequence , represents the standard motion mode of the track foreign body, uses the Euclidean distance to calculate the distance between two trajectory points, constructs the cumulative distance matrix, and finds the optimal path P in the cumulative distance matrix. The path consists of a series of points Composition, calculate the foreign body spatiotemporal consistency index, the expression is: ; Where L is the length of the optimal path, is the Euclidean distance on the path; S4: After determining that the detected target is a real threat, extract multidimensional features of the detected foreign object, including shape, texture, color, and size, and use the multidimensional features as input to the second deep learning model. The second deep learning model outputs the classification confidence, analyzes the proportion of low-confidence targets, generates a low-confidence target proportion fluctuation index, and evaluates the classification stability of the second deep learning model algorithm based on the low-confidence target proportion fluctuation index; Specifically, the method for obtaining the low confidence target ratio fluctuation index is as follows: collect the time series of the low confidence target ratio, divide the time series R into w intervals, count the number of data points in each interval, and calculate the probability that the low confidence ratio falls into the i-th interval. , calculate the low confidence target proportion fluctuation index, the expression is: ; S5: Based on the spatiotemporal consistency of the track foreign object detection results and the classification stability of the second deep learning model algorithm, the accuracy of the image recognition algorithm for track foreign object detection is evaluated. Based on the evaluation results, the track foreign object detection results are divided into accurate detection results, incompletely accurate detection results, and inaccurate detection results; S6: For incomplete accuracy detection results, predict the degree of abnormality in the accuracy of the image recognition algorithm for track foreign object detection in the subsequent detection time period. If the degree of abnormality is high, increase the resolution or frame rate of image acquisition to enhance the clarity of the track image.

2. The method for intelligent detection of foreign objects on a track based on image recognition according to claim 1, characterized in that: In S2, an image segmentation algorithm is used to segment the track area. Edge detection and Hough transform are used to extract rail edges and straight line structures, distinguishing the left and right sides of the track. A semantic segmentation model is used to classify each pixel in the image, separating the track area from the background area. Pixel labels of the track area are also annotated for the model to learn track features. Based on the segmentation results, the image portion of the track area is retained; Define the region of interest (ROI) within the track area and delineate the ROI boundary based on the geometric characteristics of track width and spacing. Use a convolutional neural network model to perform target detection on the extracted ROI area. Select the input track area image and ROI label, and train the model based on the annotated foreign object location and category. Use classification loss and regression loss to optimize the model. Output the foreign object's bounding box coordinates, classification label, and confidence score.

3. The method for intelligent detection of foreign objects on a track based on image recognition according to claim 2, characterized in that: In S3, the position change and movement direction of the track foreign matter in the continuous frames are analyzed to generate the foreign matter spatiotemporal consistency index. The method for obtaining the foreign matter spatiotemporal consistency index is as follows: Set the target trajectory sequence to , represents the position sequence of the detected track foreign objects in consecutive frames, and sets the reference trajectory sequence , represents the standard motion pattern of an orbiting object; the distance between two trajectory points is calculated using the Euclidean distance , the expression is: ;in, is the position coordinate of the target trajectory point, For the position coordinates of the reference trajectory points, construct the cumulative distance matrix D: ;in, From the starting point of the trajectory to The cumulative minimum distance of the points, is the Euclidean distance between the current points, Represents the optimal path selection from the previous position to the current point, and the boundary conditions are: ; ; ; In the cumulative distance matrix, find the distance from the starting point (1,1) to the end point The optimal path P consists of a series of points Composition, satisfying: ; It means that the path P is selected so that the cumulative distance of all points on the path is minimized, and the spatiotemporal consistency index of foreign objects is calculated. The expression is: ; Where L is the length of the optimal path, is the Euclidean distance on the path, and HAK is the spatiotemporal consistency index of foreign objects.

4. The method for intelligent detection of foreign objects on tracks based on image recognition according to claim 3, characterized in that: In S4, the proportion of low-confidence targets is analyzed to generate a low-confidence target proportion fluctuation index. The method for obtaining the low-confidence target proportion fluctuation index is: Collect a time series of the proportion of low-confidence targets ; Indicates the proportion of low-confidence targets in the T-th frame image, where T is the total number of frames in the time series. The time series R is segmented into w intervals: ; Indicates the boundary of the wth interval; counts the number of data points in each interval , the expression is: ; Calculate the probability of the i-th interval: ; is the probability that the low confidence ratio falls into the i-th interval, and the low confidence ratio target fluctuation index is calculated. The expression is: ; Where GXH is the fluctuation index of the proportion of low-confidence targets.

5. The method for intelligent detection of foreign objects on a track based on image recognition according to claim 4, characterized in that: In S5, the accuracy of the image recognition algorithm for track foreign object detection is evaluated based on the spatiotemporal consistency of the track foreign object detection results and the classification stability of the second deep learning model algorithm; The foreign body spatiotemporal consistency index and the low-confidence target proportion fluctuation index are converted into comprehensive feature vectors, and the comprehensive feature vectors are used as the input of the machine learning model. The machine learning model uses each set of comprehensive feature vectors to predict the accuracy value label of the image recognition algorithm for track foreign body detection as the prediction target, and takes minimizing the sum of the prediction errors of the accuracy value labels of all image recognition algorithms for track foreign body detection as the training target. The machine learning model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The accuracy value of the image recognition algorithm for track foreign body detection is determined according to the model output results. Among them, the machine learning model is a polynomial regression model.

6. The method for intelligent detection of foreign objects on a track based on image recognition according to claim 5, characterized in that: In S5, based on the evaluation results, the track foreign object detection results are divided into accuracy detection results, incomplete accuracy detection results, and inaccuracy detection results, specifically: Comparing the acquired accuracy value of the image recognition algorithm for detecting foreign objects on the track with a gradient standard threshold, wherein the gradient standard threshold includes a first standard threshold and a second standard threshold, and the first standard threshold is less than the second standard threshold, and comparing the accuracy value of the image recognition algorithm for detecting foreign objects on the track with the first standard threshold and the second standard threshold respectively; If the accuracy value of the image recognition algorithm for detecting foreign objects on the track is greater than the second standard threshold, it indicates that the image recognition algorithm has high accuracy in detecting foreign objects on the track. In this case, a high-accuracy detection signal is generated, and the track foreign object detection result is classified as an accuracy detection result. If the accuracy value of the image recognition algorithm for detecting foreign objects on the track is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, it means that the accuracy of the image recognition algorithm for detecting foreign objects on the track is moderate, and a moderate accuracy detection signal is generated, and the track foreign object detection result is classified as an incomplete accuracy detection result; If the accuracy value of the image recognition algorithm for detecting foreign objects on the track is less than the first standard threshold, it means that the accuracy of the image recognition algorithm for detecting foreign objects on the track is low. At this time, a low-accuracy detection signal is generated, and the track foreign object detection result is classified as an inaccurate detection result.

7. The method for intelligent detection of foreign objects on a track based on image recognition according to claim 1, characterized in that: In S6, for the incomplete accuracy detection results, the accuracy abnormality of the image recognition algorithm for track foreign object detection in the subsequent detection period is predicted, specifically: For incomplete accuracy detection results, that is, the accuracy value of the image recognition algorithm for track foreign object detection generated within a fixed detection time period is greater than or equal to the first standard threshold and less than or equal to the second standard threshold, the accuracy values ​​greater than or equal to the first standard threshold and less than or equal to the second standard threshold in the subsequent detection time period are collected, and a corresponding data set is established. The mean and standard deviation of the data set are calculated, and after analysis, the degree of abnormality in the accuracy of the image recognition algorithm for track foreign object detection in the subsequent detection time period is predicted based on the analysis results.

8. The method for intelligent detection of foreign objects on a track based on image recognition according to claim 7, characterized in that: If the mean accuracy value in the data set is greater than or equal to the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is less than the reference threshold of the standard deviation of the accuracy value, the test result is stable and accurate. In this case, no warning signal is generated and no adjustment is required. The existing test settings can be maintained. If the mean accuracy value is greater than or equal to the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is greater than or equal to the reference threshold of the standard deviation of the accuracy value, the detection accuracy fluctuates greatly, and a third-level warning signal is generated at this time, and the algorithm parameters are adjusted or the data is increased; If the mean accuracy value is less than the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is greater than or equal to the reference threshold of the standard deviation of the accuracy value, the detection accuracy is low and fluctuates significantly. In this case, a level 1 warning signal is generated and improvement measures should be taken, including optimizing the model or improving the image acquisition quality. If the mean accuracy value is less than the reference threshold of the mean accuracy value, and the standard deviation of the accuracy value is less than the reference threshold of the standard deviation of the accuracy value, the detection accuracy is low. At this time, a secondary warning signal is generated, and the model should be optimized or the algorithm should be switched.

9. An intelligent track foreign object detection system based on image recognition, used to implement the intelligent track foreign object detection method based on image recognition according to any one of claims 1 to 8, characterized in that: It includes image acquisition and preprocessing module, image segmentation and target detection module, spatiotemporal consistency analysis module, classification stability assessment module, detection result evaluation and classification module, and dynamic optimization and anomaly prediction module; Image acquisition and preprocessing module: This module uses high-definition cameras installed along the track to acquire real-time image data of the track area and performs preprocessing on the acquired image data, including denoising, contrast enhancement, and color correction. Image segmentation and object detection module: uses an image segmentation algorithm to segment the track area, extracts the region of interest within the track area, and performs object detection on the extracted region of interest based on a first deep learning model, which includes a convolutional neural network, to identify potential track foreign objects; Spatiotemporal consistency analysis module: This module obtains the spatiotemporal data of track foreign objects in different frames, analyzes the position changes and movement directions of track foreign objects in consecutive frames, verifies the spatiotemporal consistency of track foreign object detection results, and determines whether the detected target is a real threat based on the spatiotemporal consistency of the foreign object; Classification stability assessment module: After determining that the detected target is a real threat, it extracts multidimensional features of the detected foreign object, including shape, texture, color, and size. These multidimensional features are used as input to the second deep learning model. The second deep learning model outputs the classification confidence, analyzes the proportion of low-confidence targets, generates a low-confidence target proportion fluctuation index, and evaluates the classification stability of the second deep learning model algorithm based on the low-confidence target proportion fluctuation index; Detection result evaluation and classification module: This module evaluates the accuracy of the image recognition algorithm for detecting track foreign objects based on the spatiotemporal consistency of the detection results and the classification stability of the second deep learning model algorithm. Based on the evaluation results, the detection results are classified into accurate detection results, incompletely accurate detection results, and inaccurate detection results. Dynamic optimization and anomaly prediction module: For incomplete accuracy detection results, the accuracy abnormality of the image recognition algorithm for track foreign object detection in the subsequent detection time period is predicted. If the abnormality is high, the resolution or frame rate of image acquisition is increased to enhance the clarity of the track image.

Images