Rail transit clearance intruding object detection method and system based on feature extraction

Abstract

The application provides a rail transit clearance foreign matter intrusion detection method and system based on feature extraction, and relates to the technical field of rail transit safety. In view of the low detection precision of foreign matter intrusion in the prior art, the application acquires original image information in the target clearance and performs pretreatment; feature extraction is performed on the pretreated image information to obtain a multilevel image feature set; feature unwinding is performed on the multilevel image feature set to separate out category recognition degree features and interference features, and the interference features are removed; the category recognition degree features are input into a feature purification module, a classification submodule is trained and key discriminant features are extracted; a reconstruction submodule is trained and redundant information is removed; the classification submodule and the reconstruction submodule are updated cooperatively during the training process, and a final feature purification module is determined, purified features which reserve the key discriminant features and remove the redundant information are output; and foreign matter intrusion detection is performed based on the purified features. The application has high detection precision.

Description

Technical Field

[0001] This invention relates to the field of rail transit safety technology, and in particular to a method and system for detecting foreign object intrusion within rail transit clearance based on feature extraction. Background Technology

[0002] With the continuous expansion of urban rail transit networks and the increasing intensity of operations year by year, operational safety within the clearance zone becomes particularly important. The clearance zone includes critical infrastructure such as tracks, overhead contact lines, and signaling equipment, and the safe operation of these facilities directly affects the operational safety of rail transit.

[0003] Currently, foreign object detection within clearance zones still mainly relies on manual inspections or automated methods based on traditional image recognition. While these methods can detect some anomalies under normal conditions, they have significant limitations in complex real-world scenarios. Factors such as changes in lighting, surface contamination, background interference, and differences in camera angles can easily lead to a decrease in detection accuracy. Moreover, new types of anomalies frequently appear within clearance zones, such as foreign object accumulation on tracks, accidentally dropped objects, or abnormal attachments on contact wire poles or lines. Existing detection methods struggle to consistently identify these out-of-distribution (OOD) anomalies.

[0004] In summary, existing technologies suffer from numerous interferences in complex scenarios and difficulty in maintaining stable discrimination against out-of-distribution anomalies, resulting in low detection accuracy.

[0005] Therefore, developing a method and system for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction is of great significance for improving the accuracy of foreign object intrusion detection. Summary of the Invention

[0006] To address the issue of low accuracy in foreign object intrusion detection in existing technologies, this invention proposes a feature extraction-based method for detecting foreign object intrusion within rail transit clearances, specifically including the following steps:

[0007] S1. Obtain the original image information within the target clearance of the rail transit, and preprocess the original image information;

[0008] S2. Extract features from the preprocessed image information to obtain a multi-level image feature set;

[0009] S3. Perform feature unwrapping processing on the multi-level image feature set to separate the category recognition features and interference features, and remove the interference features;

[0010] S4. Refine the category discrimination features to obtain purified features; specifically including:

[0011] The category discrimination features are input into the feature purification module, which includes a classification sub-module and a reconstruction sub-module.

[0012] The classification submodule is trained using a first preset function, and key discriminant features of category recognition characteristics are extracted based on the trained classification submodule.

[0013] The reconstruction submodule is trained using the second preset function, and redundant information in the category recognition features is removed based on the trained reconstruction submodule.

[0014] By combining the first and second preset functions, the classification submodule and the reconstruction submodule are updated collaboratively to determine the final classification submodule and the final reconstruction submodule;

[0015] The final feature purification module is determined based on the final classification submodule and the final reconstruction submodule;

[0016] The category identification features are input into the final feature purification module to obtain purified features that retain key discriminative features and remove redundant information.

[0017] S5. Detect foreign object intrusion within the target clearance of rail transit based on the purification characteristics;

[0018] Specifically, this includes: inputting the purification features into the final classification submodule to obtain a predicted confidence score that the image within the target boundary is in a normal state; determining whether the predicted confidence score is less than a preset threshold; if so, determining that there is a foreign object intruding into the target boundary, and that the foreign object is an unknown foreign object outside the distribution range.

[0019] Furthermore, in S3, the feature unwrapping process for the multi-level image feature set includes: normalizing the multi-level image feature set; decomposing the normalized multi-level image feature set to obtain amplitude components and phase components; separating category recognition features and interference features by retaining the phase components and suppressing the amplitude components, and removing interference features.

[0020] Furthermore, the normalized multi-level image feature set is decomposed to obtain amplitude and phase components, including: performing a discrete Fourier transform on the normalized multi-level image feature set to decompose it into amplitude and phase components.

[0021] Furthermore, the multi-level image feature set is normalized, including: performing batch normalization and instance normalization on each group of multi-level image features in the multi-level image feature set; and performing learnable weighted fusion on the results of batch normalization and instance normalization to obtain the normalized multi-level image feature set.

[0022] Furthermore, in step S2, feature extraction is performed on the preprocessed image information to obtain a multi-level image feature set, including: using a convolutional neural network to extract features from the preprocessed image information to obtain a multi-level image feature set.

[0023] Furthermore, in step S4, the classification sub-module is trained using a first preset function, including: obtaining core information related to category discrimination from the category discrimination features based on the first preset function; extracting key discrimination features based on the core information, the key discrimination features including the predicted category probability distribution; labeling the original image information, using the labeled information as label data, and training the classification sub-model based on the predicted category probability distribution and the label data.

[0024] Furthermore, the first preset function is a cross-entropy loss function. Based on the predicted class probability distribution and label data, a classification sub-model is trained, including: inputting the predicted class probability distribution and label data into the cross-entropy loss function to calculate the classification loss value; and adjusting the classification sub-model parameters based on the classification loss value.

[0025] Furthermore, in step S4, the classification submodule and the reconstruction submodule are collaboratively updated by combining the first preset function and the second preset function, including: reconstructing the category recognition features based on the second preset function and outputting the reconstructed image corresponding to the category recognition features; inputting the reconstructed image and preprocessed image information into the second preset function to calculate the reconstruction loss value; deriving the adversarial loss value based on the reconstruction loss value; combining the adversarial loss value with the classification loss value to construct the total loss function; and collaboratively updating the classification submodule and the reconstruction submodule based on the total loss function.

[0026] Furthermore, the second preset function includes a mean squared error function or an L1 loss function.

[0027] The present invention also provides a foreign object intrusion detection system within rail transit clearance based on feature extraction, the system being used to execute the foreign object intrusion detection method within rail transit clearance based on feature extraction described in any of the preceding claims, the system comprising:

[0028] The information acquisition module is used to acquire image information of the rail transit target clearance and to preprocess the image information.

[0029] The feature extraction module is used to extract features from the preprocessed image information to obtain a multi-level image feature set.

[0030] The feature unwrapping module is used to perform feature unwrapping processing on multi-level image feature sets to separate category recognition features and interference features, and remove interference features;

[0031] The feature purification module is used to purify the category discrimination features to obtain purified features. Specifically, it includes: inputting the category discrimination features into the feature purification module, which comprises a classification sub-module and a reconstruction sub-module; training the classification sub-module using a first preset function, and extracting key discriminative features of the category discrimination features based on the trained classification sub-module; training the reconstruction sub-module using a second preset function, and removing redundant information from the category discrimination features based on the trained reconstruction sub-module; combining the first and second preset functions to collaboratively update the classification sub-module and the reconstruction sub-module to determine the final classification sub-module and the final reconstruction sub-module; determining the final feature purification module based on the final classification sub-module and the final reconstruction sub-module; and inputting the category discrimination features into the final feature purification module to obtain purified features that retain key discriminative features and remove redundant information.

[0032] The detection module is used to detect foreign object intrusion within the target clearance of rail transit based on the purification features; specifically, it includes: inputting the purification features into the final classification submodule to obtain a predicted confidence score that the image within the target clearance is in a normal state; determining whether the predicted confidence score is less than a preset threshold; if so, determining that there is foreign object intrusion within the target clearance, and that the foreign object is an unknown foreign object outside the distribution.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] This invention acquires raw image information within the target boundary of rail transit, preprocesses the raw image information, and extracts features from the preprocessed image information to obtain a multi-level image feature set. The multi-level image feature set undergoes feature unwrapping processing to separate category-discriminating features and interference features, and the interference features are removed. The category-discriminating features are then purified to obtain cleaned features. Specifically, this includes: inputting the category-discriminating features into a feature purification module, which includes a classification sub-module and a reconstruction sub-module; training the classification sub-module using a first preset function, and extracting key discriminative features based on the trained classification sub-module; training the reconstruction sub-module using a second preset function, and removing redundant information from the category-discriminating features based on the trained reconstruction sub-module; combining the first and second preset functions to collaboratively update the classification sub-module and the reconstruction sub-module to determine the final classification sub-module and the final reconstruction sub-module; determining the final feature purification module based on the final classification sub-module and the final reconstruction sub-module; and inputting the category-discriminating features into the final feature purification module to obtain cleaned features that retain key discriminative features and remove redundant information. Based on the aforementioned purification features, foreign object intrusion within the target clearance of rail transit is detected. Through feature unwrapping processing, category identification features and interference features are separated and eliminated, directly reducing the impact of irrelevant factors on detection. This allows subsequent feature processing and detection stages to focus only on the core features related to foreign object intrusion, avoiding interference features from masking the anomaly and facilitating accurate detection.

[0035] Meanwhile, the feature purification stage employs a dual-path design with a classification submodule and a reconstruction submodule, achieving dual optimization by preserving key features and eliminating redundant information. Through training with a first preset function, key discriminative features are specifically extracted from the category identification features, enhancing the feature purification module's ability to distinguish between normal and abnormal categories. Through training with a second preset function, redundant information in the category identification features is selectively eliminated, reducing the interference of useless details on the detection results. Combining the first and second preset functions, the classification and reconstruction submodules are updated collaboratively, resulting in more coordinated parameter adaptation between the two submodules. The final feature purification module outputs purified features that possess both strong discriminative power and sufficient purity, improving detection accuracy from the source. Attached Figure Description

[0036] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0037] Figure 1This is a flowchart of a foreign object intrusion detection method based on feature extraction within the rail transit clearance provided in an embodiment of the present invention;

[0038] Figure 2 This is a flowchart of feature purification provided in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of the structure of the foreign object intrusion detection system within the rail transit clearance based on feature extraction provided in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0041] The specific embodiments of the present invention will be described below.

[0042] To address the issue of low accuracy in foreign object intrusion detection in existing technologies, this invention acquires original image information within the target clearance of rail transit, preprocesses the original image information, and extracts features from the preprocessed image information to obtain a multi-level image feature set. Feature unwrapping is performed on a multi-level image feature set to separate category-discriminating features and interfering features, and the interfering features are removed. The category-discriminating features are then purified to obtain cleaned features. Specifically, this involves: inputting the category-discriminating features into a feature purification module, which includes a classification sub-module and a reconstruction sub-module; training the classification sub-module using a first preset function, and extracting key discriminative features from the category-discriminating features based on the trained classification sub-module; training the reconstruction sub-module using a second preset function, and removing redundant information from the category-discriminating features based on the trained reconstruction sub-module; combining the first and second preset functions to collaboratively update the classification and reconstruction sub-modules to determine the final classification and reconstruction sub-modules; determining the final feature purification module based on the final classification and reconstruction sub-modules; and inputting the category-discriminating features into the final feature purification module to obtain cleaned features that retain key discriminative features and remove redundant information. Foreign object intrusion within the rail transit target clearance is detected based on these cleaned features. This invention has high detection accuracy.

[0043] Example 1

[0044] This invention provides a method for detecting foreign object intrusion within the clearance limits of rail transit systems based on feature extraction. Figure 1This is a flowchart of a foreign object intrusion detection method based on feature extraction within the rail transit clearance area provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the specific steps include the following:

[0045] S1. Obtain the original image information within the target clearance of the rail transit and preprocess the original image information.

[0046] The rail transit target clearance refers to the safety boundary range of rail transit facilities, including critical infrastructure such as tracks, overhead contact lines, and signaling equipment. It is the core monitoring area for foreign object intrusion detection. Raw image information refers to images of the scene within the rail transit clearance, collected by image acquisition equipment such as cameras and LiDAR installed at fixed monitoring points, without any technical processing. Preprocessing the acquired raw image information involves standardizing the images, including pixel value normalization, resizing, and noise reduction. This preprocessing unifies the data format, reduces interference factors, and provides high-quality data input for subsequent feature extraction.

[0047] S2. Extract features from the preprocessed image information to obtain a multi-level image feature set.

[0048] Specifically, this includes using a convolutional neural network to extract features from the preprocessed image information, resulting in a multi-level image feature set.

[0049] Deep convolutional neural networks (CNNs) are selected as the core extraction tool. These CNNs can include Residual Neural Networks (ResNet) and Densely Connected Convolutional Networks (DenseNet), among others. These networks possess deep structures and strong feature representation capabilities, making them suitable for rail transit scenarios. Preprocessed image information is input in batches into the selected CNN. The CNN processes the image progressively through multi-stage convolutional blocks, outputting multi-level image feature sets at different scales and levels.

[0050] Among them, the multi-level feature set contains three key types of information: texture, structure, and semantics. It can capture the subtle details of foreign objects through shallow features and identify the essence of categories through deep features, thus avoiding the information limitations of single-level features.

[0051] S3. Perform feature unwrapping processing on the multi-level image feature set to separate the category recognition features and interference features, and remove the interference features.

[0052] Feature unwrapping refers to the process of dividing features in a multi-level image feature set into category-discriminating features and interference features. Category-discriminating features are those that can accurately distinguish between normal conditions and abnormalities within the rail transit clearance. Examples include the outline of the object and the standard structure of the track. Category-discriminating features are crucial for subsequent detection. Interference features are image style features that are irrelevant to the object detection task and affect the accuracy of the discrimination. Examples include appearance differences caused by changes in lighting, surface contamination, texture differences, and camera angle.

[0053] By separating category recognition features and interference features, we can effectively eliminate style interference that is irrelevant to the task, such as changes in lighting, surface contamination, and texture differences. This prevents these factors from masking the true nature of the anomalies, making the expression of category recognition features more stable, adapting to complex rail transit scenarios, suppressing style interference, and improving feature stability.

[0054] For example, a 50-layer residual neural network, ResNet-50, is used as the core extraction tool. This network consists of five main stages: an initial convolutional layer (conv1), four residual block groups (conv2_x, conv3_x, conv4_x, conv5_x), a global average pooling layer, and a fully connected classification layer. To balance local texture, structural patterns, and global semantic features, feature unwrapping is performed after the outputs of the conv2_x, conv3_x, and conv4_x stages, respectively, to simultaneously process low-level details, mid-level textures, and high-level semantic information at different scales. For shallow features, feature unwrapping focuses more on suppressing high-frequency noise and background noise; for deep features, feature unwrapping focuses more on preserving semantic structure and removing overall stylistic features.

[0055] Specifically, the multi-level image feature set is subjected to feature unwrapping processing, including: normalizing the multi-level image feature set; decomposing the normalized multi-level image feature set to obtain amplitude components and phase components; separating category recognition features and interference features by retaining the phase components and suppressing the amplitude components, and removing interference features.

[0056] The normalization process for the multi-level image feature set includes: performing batch normalization and instance normalization for each group of multi-level image features in the multi-level image feature set; and performing learnable weighted fusion of the results after batch normalization and the results after instance normalization to obtain the normalized multi-level image feature set.

[0057] A feature unwrapping module is set after the multi-stage convolutional block of the convolutional neural network. The feature unwrapping module removes image style information and retains class recognition features by learningable weighted fusion of batch normalization (BN) and instance normalization (IN).

[0058] Applying batch normalization (BN) and instance normalization (IN) to the intermediate features of each stage's output in a convolutional neural network (CNN) can help adjust the statistical properties of each sample, making them closer to the statistical properties of the target style. However, BN may reduce the inter-class variance of the image, leading to a decrease in the network's discriminative ability. Batch normalization (BN) can preserve task-related content features between different samples, but it is sensitive to changes in image appearance. Since the two are complementary, a better approach is to fully integrate the two normalization strategies. Preferably, the learnable weighting coefficients can be channel-level vectors or scalars, and are adaptively updated by backpropagation during training. In practice, this fusion can be represented as a weighted linear combination of the BN and IN outputs, thereby minimizing style interference in the spatial domain while preserving class-discriminative features.

[0059] Based on the above embodiments, the normalized multi-level image feature set is decomposed to obtain amplitude components and phase components, including: performing a discrete Fourier transform on the normalized multi-level image feature set to decompose it into amplitude components and phase components.

[0060] The fusion features in the spatial domain are mapped to the frequency domain using a two-dimensional discrete Fourier transform (DFT). The resulting complex spectrum is then decomposed into amplitude and phase components. The phase component primarily reflects the structural and geometric information of the image, while the amplitude component mainly reflects stylistic information such as lighting and texture. The process of using the DFT to unravel the features in the frequency domain includes:

[0061] ;

[0062] In the formula, f norm The normalized features are the fused features, and γ is a learnable parameter. f BN and f IN Calculated from BN and IN respectively. H, W, and C refer to the height, width, and number of channels of the feature map, respectively. For single-channel normalized features... f norm Its Fourier transform F(x) is expressed as:

[0063] ;

[0064] In the formula, w and h are spatial coordinates, u and v are Fourier spatial coordinates, and i is the imaginary unit. Composed of complex numbers, it can be represented as:

[0065] ;

[0066] in, and They are The real and imaginary parts. In Fourier space, Consists of an amplitude component and a phase component The composition is as follows:

[0067] ;

[0068] Phase generally represents the content information of an image, while amplitude represents style information. Although style information does not have semantic information, completely removing style information will obviously affect performance. Therefore, a learnable parameter λ is used to control the degree of style information removal, and the specific control process is as follows:

[0069]

[0070] This is the inverse Fourier transform process. This refers to spatial features that have had irrelevant style information removed after being processed by the feature unwrapping module.

[0071] To reduce the interference of style on the discrimination, a learnable scaling factor (i.e., parameter λ) is introduced into the amplitude component in the frequency domain and suppressed. The scaling factor can be a scalar, a channel vector, or a band-related mask, so as to adaptively adjust the suppression strength for different channels or frequency bands during training. The spectrum is reconstructed using the suppressed amplitude component and the original phase component, and the spectrum is mapped back to the spatial domain through inverse Fourier transform to obtain a destylized feature representation relative to the original features.

[0072] S4. Refine the category recognition features to obtain purified features. Figure 2 This is a flowchart of feature purification provided in an embodiment of the present invention, such as... Figure 2 As shown, it specifically includes:

[0073] S41. Input the category recognition features into the feature purification module, which includes a classification sub-module and a reconstruction sub-module.

[0074] S42. The classification sub-module is trained using the first preset function, and the key discriminant features of the category recognition feature are extracted based on the trained classification sub-module.

[0075] S43. The second preset function is used to train the reconstruction submodule, and redundant information in the category recognition features is removed based on the trained reconstruction submodule.

[0076] S44. Combining the first preset function and the second preset function, the classification submodule and the reconstruction submodule are updated collaboratively to determine the final classification submodule and the final reconstruction submodule.

[0077] S45. Determine the final feature purification module based on the final classification submodule and the final reconstruction submodule.

[0078] S46. Input the category recognition features into the final feature purification module to obtain purified features that retain key discriminative features and remove redundant information.

[0079] Specifically, the classification sub-module is trained using a first preset function, including: obtaining core information related to category discrimination from the category discrimination features based on the first preset function; extracting key discrimination features based on the core information; the key discrimination features include the predicted category probability distribution; labeling the original image information, using the labeled information as label data, and training the classification sub-model based on the predicted category probability distribution and the label data.

[0080] For example, the first preset function is the cross-entropy loss function. Based on the predicted class probability distribution and label data, a classification sub-model is trained, including: inputting the predicted class probability distribution and label data into the cross-entropy loss function to calculate the classification loss value; and adjusting the classification sub-model parameters based on the classification loss value.

[0081] The category discrimination features, after feature unwrapping and style interference removal, are input into the classification submodule. Under the supervision of the cross-entropy loss function, core information strongly correlated with category discrimination, such as normal state within rail transit clearance and foreign object intrusion, is selected from the category discrimination features. Based on this core information, key discrimination features are extracted through global average pooling and fully connected layer processing. These key discrimination features are specifically represented by the predicted category probability distribution. The collected raw image information is manually labeled, and the labeled category identifiers are used as label data. The predicted category probability distribution and the label data are then substituted into the cross-entropy loss function to obtain the classification loss value L. cls Based on the backpropagation algorithm, the internal parameters of the classification sub-module, such as the weights of fully connected layers and the parameters of convolutional kernels, are adaptively adjusted according to the classification loss value. The process of feature input, core information filtering, prediction probability output, loss calculation, and parameter adjustment is continuously iterated until the classification loss value stabilizes at a low level, thus completing the training of the classification sub-model. The trained classification sub-model has the ability to accurately distinguish between normal conditions and foreign object intrusion within the rail transit boundary.

[0082] Specifically, the classification submodule and reconstruction submodule are collaboratively updated by combining the first and second preset functions. This includes: reconstructing the category recognition features based on the second preset function and outputting the reconstructed image corresponding to the category recognition features; inputting the reconstructed image and preprocessed image information into the second preset function to calculate the reconstruction loss value; deriving the adversarial loss value based on the reconstruction loss value; combining the adversarial loss value with the classification loss value to construct the total loss function; and collaboratively updating the classification submodule and reconstruction submodule based on the total loss function. For example, the second preset function may include a mean squared error function or an L1 loss function.

[0083] The high-level features processed by the unwrapping module are input into the parallel classification and reconstruction sub-modules. The classification branch outputs the predicted class probability distribution after global average pooling and several fully connected layers, which is used to maintain the ability to distinguish different types of foreign objects in the scene within the rail transit boundary. The reconstruction sub-module mainly consists of an upsampling layer, residual blocks, and a 1×1 convolutional layer. To train the reconstruction module, the reconstructed image and the preprocessed image information are input into a second preset function to calculate the reconstruction loss value. For example, the second preset function uses the mean squared error function, and the calculation formula is as follows:

[0084] ;

[0085] in, This represents the reconstruction loss value, which is a value mapped to the [0,1] interval after normalization. xi represents the preprocessed image information. This represents the reconstructed image. Since fine-grained recognition relies primarily on subtle differences between high-level semantic features, the main task of the reconstruction submodule is to reconstruct low-level details irrelevant to the core task. To prevent the detection process from focusing on these task-irrelevant features and instead concentrate on extracting task-relevant features, an adversarial reconstruction method is employed to prevent the reconstruction of such feature information. The goal of the feature purification module is defined as:

[0086] ;

[0087] in, This represents the loss value against the loss.

[0088] By combining the adversarial loss value and the classification loss value, a total loss function is constructed, which is expressed as:

[0089] ;

[0090] Where α is a weighting coefficient used to balance classification accuracy and feature purification effect.

[0091] Based on the constructed total loss function, the internal parameters of the classification sub-module and the reconstruction sub-module are simultaneously and collaboratively updated through the backpropagation algorithm. For example, the weights of the fully connected layer of the classification sub-module, the convolution kernel parameters of the reconstruction sub-module, etc. are updated. During the training process, a staged training or alternating training strategy can be adopted to ensure that while the classification sub-module maintains feature discriminability, the reconstruction sub-module can effectively remove low-level redundant information, ultimately achieving parameter adaptation and performance optimization of the two sub-modules.

[0092] This embodiment adopts a joint optimization strategy of classification loss and adversarial loss. Among them, the classification loss measures the difference between the predicted label and the true label through the cross-entropy function. The reconstruction loss measures the difference between the original image and the reconstructed image through the mean square error. The adversarial loss is defined as the reverse measure of the reconstruction loss and is used to suppress the retention of irrelevant information in the features. The total loss function is the sum of the classification loss and the weighted adversarial loss, where the weighting coefficient α is used to balance the classification accuracy and the feature purification effect. For example, when α takes 0.2, the best performance is obtained through verification set testing.

[0093] S5. Detect foreign object intrusion within the target clearance of rail transit based on the purified features.

[0094] Foreign object intrusion refers to various abnormal situations that invade the target clearance of rail transit, including: foreign objects dropped on the track, bird nests attached to the catenary poles, tools left over from construction, and out-of-distribution (OOD) anomalies not covered in the training set, such as kite entanglement, cable damage, etc.

[0095] Specifically, the prediction confidence score S is obtained through the output layer of the final classification sub-module. If S ≥ T (T is a preset threshold), it is determined as a normal clearance image; if S < T, it is determined that there is foreign object intrusion within the target clearance. Since the feature purification process eliminates redundant information and suppresses non-target styles, when the input is an out-of-distribution foreign object image that has not been seen before, the prediction confidence score will drop significantly, thus achieving effective identification of unknown foreign objects.

[0096] Calculate the prediction confidence score using the purified features. Among them, the scoring method can choose the maximum softmax probability, energy function, or feature space distance metric, etc. Compare the calculated prediction confidence score with the preset threshold. If the prediction confidence score is lower than the preset threshold, it means that the image corresponding to this feature does not conform to the normal scene feature distribution and is determined as foreign object intrusion. If the confidence score is higher than the threshold, it is determined as a normal scene. Through abnormal alarm, the identification and early warning of foreign objects dropped on the track, catenary bird nests, and other out-of-distribution anomalies are completed. It effectively solves the problem of easy omission of unknown foreign objects under the interference of complex environments and significantly improves the robustness and reliability of clearance detection.

[0097] This embodiment acquires original image information within the target boundary of rail transit, preprocesses the original image information, and extracts features from the preprocessed image information to obtain a multi-level image feature set. The multi-level image feature set undergoes feature unwrapping processing to separate category-discriminating features and interference features, and removes interference features. The category-discriminating features are then purified to obtain clean features. Specifically, this includes: inputting the category-discriminating features into a feature purification module, which includes a classification sub-module and a reconstruction sub-module; training the classification sub-module using a first preset function, and extracting key discriminative features based on the trained classification sub-module; training the reconstruction sub-module using a second preset function, and removing redundant information from the category-discriminating features based on the trained reconstruction sub-module; combining the first and second preset functions to collaboratively update the classification sub-module and the reconstruction sub-module to determine the final classification sub-module and the final reconstruction sub-module; determining the final feature purification module based on the final classification sub-module and the final reconstruction sub-module; and inputting the category-discriminating features into the final feature purification module to obtain clean features that retain key discriminative features and remove redundant information. Based on the aforementioned purification features, foreign object intrusion within the target clearance of rail transit is detected. Through feature unwrapping processing, category identification features and interference features are separated and eliminated, directly reducing the impact of irrelevant factors on detection. This allows subsequent feature processing and detection stages to focus only on the core features related to foreign object intrusion, avoiding interference features from masking the anomaly and facilitating accurate detection.

[0098] Meanwhile, the feature purification stage employs a dual-path design with a classification submodule and a reconstruction submodule, achieving dual optimization by preserving key features and eliminating redundant information. Through training with a first preset function, key discriminative features are specifically extracted from the category identification features, enhancing the feature purification module's ability to distinguish between normal and abnormal categories. Through training with a second preset function, redundant information in the category identification features is selectively eliminated, reducing the interference of useless details on the detection results. Combining the first and second preset functions, the classification and reconstruction submodules are updated collaboratively, resulting in more coordinated parameter adaptation between the two submodules. The final feature purification module outputs purified features that possess both strong discriminative power and sufficient purity, improving detection accuracy from the source.

[0099] Example 2

[0100] This invention also provides a foreign object intrusion detection system within rail transit clearance based on feature extraction. Figure 3 This is a schematic diagram of the structure of the foreign object intrusion detection system within the rail transit clearance area based on feature extraction provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the system includes:

[0101] The information acquisition module is used to acquire image information of the rail transit target clearance and to preprocess the image information.

[0102] The feature extraction module is used to extract features from the preprocessed image information to obtain a multi-level image feature set.

[0103] The feature unwrapping module is used to perform feature unwrapping processing on multi-level image feature sets to separate category recognition features and interference features, and remove interference features;

[0104] The feature purification module is used to purify the category discrimination features to obtain purified features. Specifically, it includes: inputting the category discrimination features into the feature purification module, which comprises a classification sub-module and a reconstruction sub-module; training the classification sub-module using a first preset function, and extracting key discriminative features of the category discrimination features based on the trained classification sub-module; training the reconstruction sub-module using a second preset function, and removing redundant information from the category discrimination features based on the trained reconstruction sub-module; combining the first and second preset functions to collaboratively update the classification sub-module and the reconstruction sub-module to determine the final classification sub-module and the final reconstruction sub-module; determining the final feature purification module based on the final classification sub-module and the final reconstruction sub-module; and inputting the category discrimination features into the final feature purification module to obtain purified features that retain key discriminative features and remove redundant information.

[0105] The detection module is used to detect foreign object intrusion within the target clearance of rail transit based on the purification features; specifically, it includes: inputting the purification features into the final classification submodule to obtain a predicted confidence score that the image within the target clearance is in a normal state; determining whether the predicted confidence score is less than a preset threshold; if so, determining that there is foreign object intrusion within the target clearance, and that the foreign object is an unknown foreign object outside the distribution.

[0106] The feature extraction-based foreign object intrusion detection system for rail transit clearance provided in this embodiment is used to execute the feature extraction-based foreign object intrusion detection method for rail transit clearance in any of the above embodiments, and has the beneficial effects of any of the above embodiments, which will not be repeated here.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction, characterized in that, include: S1. Obtain the original image information within the target clearance of the rail transit, and preprocess the original image information; S2. Extract features from the preprocessed image information to obtain a multi-level image feature set; S3. Perform feature unwrapping processing on the multi-level image feature set to separate the category recognition features and interference features, and remove the interference features; S4. Refine the category identification features to obtain purified features; Specifically, it includes: The category discrimination features are input into the feature purification module, which includes a classification sub-module and a reconstruction sub-module. The classification submodule is trained using a first preset function, and key discriminant features of category recognition characteristics are extracted based on the trained classification submodule. The reconstruction submodule is trained using the second preset function, and redundant information in the category recognition features is removed based on the trained reconstruction submodule. By combining the first and second preset functions, the classification submodule and the reconstruction submodule are updated collaboratively to determine the final classification submodule and the final reconstruction submodule; The final feature purification module is determined based on the final classification submodule and the final reconstruction submodule; The category identification features are input into the final feature purification module to obtain purified features that retain key discriminative features and remove redundant information. S5. Detect foreign object intrusion within the target clearance of rail transit based on the purification characteristics; Specifically, this includes: inputting the purification features into the final classification submodule to obtain a predicted confidence score that the image within the target boundary is in a normal state; determining whether the predicted confidence score is less than a preset threshold; if so, determining that there is a foreign object intruding into the target boundary, and that the foreign object is an unknown foreign object outside the distribution range.

2. The method for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction according to claim 1, characterized in that, In step S3, feature unwrapping processing is performed on the multi-level image feature set, including: Normalize the multi-level image feature set; The normalized multi-level image feature set is decomposed to obtain the amplitude component and the phase component; By preserving the phase component and suppressing the amplitude component, category-discriminating features and interfering features are separated, and interfering features are eliminated.

3. The method for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction according to claim 2, characterized in that, The normalized multi-level image feature set is decomposed to obtain amplitude and phase components, including: A discrete Fourier transform is performed on the normalized multi-level image feature set to decompose it into amplitude and phase components.

4. The method for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction according to claim 2, characterized in that, Normalization processing of multi-level image feature sets includes: For each set of multi-level image features in the multi-level image feature set, batch normalization and instance normalization are performed. The results of batch normalization and instance normalization are fused using a learnable weighted fusion method to obtain a multi-level image feature set after normalization.

5. The method for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction according to claim 1, characterized in that, In step S2, feature extraction is performed on the preprocessed image information to obtain a multi-level image feature set, including: A convolutional neural network is used to extract features from the preprocessed image information to obtain a multi-level image feature set.

6. The method for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction according to claim 1, characterized in that, In step S4, the classification submodule is trained using a first preset function, including: Based on the first preset function, core information related to category discrimination is obtained from the category recognition features; Key discriminative features are extracted based on core information, including the probability distribution of predicted categories; The original image information is labeled, and the labeled information is used as label data. Based on the predicted category probability distribution and the label data, a classification sub-model is trained.

7. The method for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction according to claim 6, characterized in that, The first preset function is the cross-entropy loss function. Based on the predicted class probability distribution and label data, a classification sub-model is trained, including: The predicted category probability distribution and label data are input into the cross-entropy loss function to calculate the classification loss value; Adjust the parameters of the classification sub-model based on the classification loss value.

8. The method for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction according to claim 7, characterized in that, In step S4, the classification submodule and the reconstruction submodule are updated collaboratively by combining the first preset function and the second preset function, including: The category identification features are reconstructed based on the second preset function, and the reconstructed image corresponding to the category identification features is output. The reconstructed image and the preprocessed image information are input into the second preset function to calculate the reconstruction loss value; The adversarial loss value is derived from the reconstruction loss value; The adversarial loss value is combined with the classification loss value to construct the total loss function; Based on the total loss function, the classification submodule and the reconstruction submodule are updated collaboratively.

9. The method for detecting foreign object intrusion within the clearance limits of rail transit based on feature extraction according to claim 8, characterized in that, The second preset function includes the mean squared error function or the L1 loss function.

10. A foreign object intrusion detection system within the clearance limits of rail transit based on feature extraction, characterized in that, The system is used to execute the feature extraction-based foreign object intrusion detection method within the rail transit clearance according to any one of claims 1-9, and the system includes: The information acquisition module is used to acquire image information of the rail transit target clearance and to preprocess the image information. The feature extraction module is used to extract features from the preprocessed image information to obtain a multi-level image feature set. The feature unwrapping module is used to perform feature unwrapping processing on multi-level image feature sets to separate category recognition features and interference features, and remove interference features; The feature purification module is used to purify the category discrimination features to obtain purified features. Specifically, it includes: inputting the category discrimination features into the feature purification module, which comprises a classification sub-module and a reconstruction sub-module; training the classification sub-module using a first preset function, and extracting key discriminative features of the category discrimination features based on the trained classification sub-module; training the reconstruction sub-module using a second preset function, and removing redundant information from the category discrimination features based on the trained reconstruction sub-module; combining the first and second preset functions to collaboratively update the classification sub-module and the reconstruction sub-module to determine the final classification sub-module and the final reconstruction sub-module; determining the final feature purification module based on the final classification sub-module and the final reconstruction sub-module; and inputting the category discrimination features into the final feature purification module to obtain purified features that retain key discriminative features and remove redundant information. The detection module is used to detect foreign object intrusion within the target clearance of rail transit based on the purification features; specifically, it includes: inputting the purification features into the final classification submodule to obtain a predicted confidence score that the image within the target clearance is in a normal state; determining whether the predicted confidence score is less than a preset threshold; if so, determining that there is foreign object intrusion within the target clearance, and that the foreign object is an unknown foreign object outside the distribution.

Images