Railway foreign-object intrusion detection method and system based on deep learning

Abstract

The present invention relates to the technical field of railway inspection, and relates in particular to a railway foreign-object intrusion detection method and system based on deep learning. The method comprises: S10, acquiring image data to undergo detection; and S20, inputting the image data into a trained attention semantic segmentation network to obtain a foreign-object detection result. The attention semantic segmentation network is obtained by first training a preset attention semantic segmentation network architecture using a preconfigured data set, and then re-training the attention semantic segmentation network on specified image data using an adaptive correction algorithm. A backbone network architecture is obtained by inserting a specified attention mechanism at a specified position within a pre-selected residual neural network, and using three parallel dilated convolutions as an initial convolutional layer in the residual neural network. A dual-branch decoder combines an edge recognition branch and a semantic segmentation branch. The method is applicable to various scenarios, achieves improved detection accuracy and maintains a lightweight design.

Description

A railway foreign object intrusion detection method and system based on deep learning Technical Field

[0001] The present invention relates to the field of railway detection technology, and in particular to a railway foreign object intrusion detection method and system based on deep learning. Background Art

[0002] Railways play a vital role in transportation, offering advantages such as punctuality, comfort, and energy efficiency. However, given the nature of rail transportation, any person or object that intrudes within the railway's boundaries can cause risks such as derailment and emergency stops. This risk is known as foreign object intrusion, and timely detection and handling of foreign object intrusion is essential for ensuring safe and efficient train operations. However, with the continuous expansion of the railway network and the rapid increase in transportation scale, addressing the randomness of foreign object intrusion and improving detection reliability has become a more challenging issue.

[0003] Currently, foreign object intrusion detection technologies in railway systems can be roughly divided into three types: the first is passive protection, including fence-type systems such as electronic fences and radiation detection. This method uses sensing technology to determine whether foreign objects have come into contact with the sensor. This physical contact detection method is significantly affected by external factors, is costly, requires a lot of maintenance, and is not suitable for large-scale use. The second type relies on patrol vehicles, but due to operating hours, they cannot guarantee round-the-clock detection. The last type is manual inspection of data collected by cameras to provide reliable and continuous monitoring, but this is time-consuming and labor-intensive.

[0004] To overcome these limitations, AI-driven active detection methods have been widely studied. Research on foreign object intrusion can be divided into traditional machine learning algorithms and deep learning-based object detection algorithms. Traditional methods typically involve extracting tracks to identify regions of interest (ROIs), followed by object detection and recognition. The inherent geometric properties of the tracks, including track spacing and track parallelism, are often key features for track extraction. Using the extracted ROIs, background subtraction and optical flow are used to detect foreign objects based on pixel variations. Subsequently, machine learning algorithms (such as support vector machines) use manually designed features to identify detected objects. However, these algorithms face challenges in handling complex dynamic scenes, and distributed processing is not conducive to optimizing storage space. Deep learning methods primarily utilize object detection algorithms, which require learning objects of interest from a training set. Therefore, current research focuses on detecting common foreign objects such as trains, pedestrians, and animals. To detect different foreign objects, corresponding datasets must be established. However, foreign objects occur relatively infrequently, and to improve the generalization ability of the model, it is necessary to cover different lighting conditions and scenarios. This makes the construction of datasets for unknown foreign object images more complex, making it difficult to apply to unknown foreign object detection. In addition, most methods are mainly verified using test sets with ideal distributions, and general methods for practical deployment in different scenarios are still lacking.

[0005] Given the complex operation scenarios, diverse and unpredictable foreign objects in trains, building a comprehensive integrated prevention and control system on board and on the ground is the basis for achieving reliable foreign object intrusion detection. Therefore, lightweight intelligent detection methods suitable for different scenarios and unknown foreign objects are indispensable. Summary of the Invention

[0006] (1) Technical issues to be resolved

[0007] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a railway foreign object intrusion detection method and system based on deep learning, which solves the following technical problems of the prior art: the existing foreign object intrusion detection method is basically a target detection architecture. In order to detect different foreign objects, it is necessary to establish a corresponding data set. However, the frequency of occurrence of foreign objects is relatively low, and in order to improve the generalization ability of the model, different lighting conditions and scenes need to be covered, which makes the process of constructing a data set for unknown foreign object images more complicated and difficult to apply to unknown foreign object detection.

[0008] (2) Technical solution

[0009] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0010] In a first aspect, an embodiment of the present invention provides a railway foreign object intrusion detection method based on deep learning, comprising:

[0011] S10, obtaining image data to be detected;

[0012] S20, inputting the image data to be detected into the trained second attention semantic segmentation network to obtain a foreign body detection result;

[0013] The trained second attention semantic segmentation network is obtained by retraining the first attention semantic segmentation network using specified image data captured by a visual sensing device and an adaptive correction algorithm;

[0014] The first attention semantic segmentation network is obtained by training a preset attention semantic segmentation network architecture using a prefabricated dataset;

[0015] The prefabricated dataset is obtained by extracting data associated with railway scenes from the COCO-Stuff dataset;

[0016] The attention semantic segmentation network architecture includes a backbone network architecture and a dual-branch decoder;

[0017] Among them, the backbone network architecture is obtained by inserting a specified attention mechanism at a specified position in a preselected residual neural network, and using three parallel dilated convolutions as the initial convolution layer in the residual neural network; the dual-branch decoder is a dual-branch decoder that combines an edge recognition branch and a semantic segmentation branch.

[0018] Optionally, before S10, the method further includes:

[0019] S01. Extract data related to railway scenes from the COCO-Stuff dataset to obtain a prefabricated dataset.

[0020] S02. Using the prefabricated dataset, training the attention semantic segmentation network architecture to obtain a first attention semantic segmentation network;

[0021] S03. Use the specified image data captured by the visual sensing device and adopt an adaptive correction algorithm to retrain the first attention semantic segmentation network to obtain a second attention semantic segmentation network.

[0022] Optionally, the designated position is after each of the five stages in the residual neural network;

[0023] The attention mechanism is a dual attention mechanism of space and channel.

[0024] Optionally, the kernel sizes of the three parallel extended convolutions are 3, 5 and 7 respectively.

[0025] Optionally, there are five stages in the architecture of the backbone network, and the resolution of the data output by each stage is 1 / 2, 1 / 4, 1 / 8, 1 / 8, and 1 / 8 respectively relative to the input data.

[0026] Optionally, S02 includes:

[0027] S021. Input the data in the prefabricated dataset into the attention semantic segmentation network architecture to obtain output results of the semantic segmentation branch and the edge recognition branch respectively;

[0028] S022. Based on the true label of the input data and the output result, use the two-branch total loss expression to calculate the total loss of the two branches of the attention semantic segmentation network architecture;

[0029] The total loss expression of the two branches is: L Z =L v(x) +L ζ(x) +L b(x) ,

[0030] Among them, L Z represents the total loss of the two branches, L v(x) represents the cross entropy loss of the category balance weight and the positive and negative sample balance weight, L ζ(x) is the edge auxiliary loss, L b(x) Branch loss for edge identification;

[0031] The edge-assisted loss is calculated using an edge-assisted loss expression;

[0032] The edge-assisted loss expression is:

[0033] Among them, L ζ(x) is the edge identification branch loss, A(x n ) d 、A(x n ) b Represent the output results of the semantic segmentation branch and the edge branch respectively, Represent semantic labels and edge labels respectively; Same as L v(x) The class balance weights are the same, N represents the number of samples, and n represents the sample index;

[0034] S023. Repeat S021 and S022 until the total loss value of the two branches converges to obtain the trained first attention semantic segmentation network.

[0035] Optionally, the adaptive correction algorithm is a domain adaptive correction algorithm based on multi-kernel maximum mean square deviation, and is connected after each attention mechanism;

[0036] The adaptive correction algorithm includes an unsupervised domain adaptive algorithm and a post-correction algorithm.

[0037] Optionally, S03 includes:

[0038] S031. Input the source domain and the target domain into the first attention semantic segmentation network to obtain a generated result;

[0039] The source domain is the prefabricated dataset, and the target domain is the specified image data captured by the visual sensing device;

[0040] S032. Use the distribution difference calculation expression to calculate the distribution difference between the source domain and the target domain, that is, to guide the network to adapt to the loss of the target domain;

[0041] The distribution difference calculation expression is:

[0042] Among them, L Φ(s，t) Represents the distribution difference between the source domain and the target domain, s and t represent the source domain and target domain samples respectively, using different kernel functions G k (·) is mapped to the Hilbert space to calculate the expected gap; K represents the number of kernel functions used, k represents the kernel function index; P represents the number of source domain samples, p represents the source domain sample index; Q represents the number of target domain samples, q represents the target domain sample index; P and Q are equal;

[0043] S033. Correct the generated result using a post-correction algorithm based on the calculated distribution difference to obtain a corrected result, and compare the corrected result with the corresponding target domain sample to obtain a correction loss;

[0044] S034. Repeat S031-S033 until the correction loss is less than a preset threshold, and obtain a trained second attention semantic segmentation network.

[0045] Optionally, the method further includes:

[0046] S30. If the foreign object detection result is a foreign object, an alarm is sent to the driver, and the alarm is fed back to the back-end management personnel for verification.

[0047] In a second aspect, an embodiment of the present invention provides a railway foreign object intrusion detection system based on deep learning, comprising:

[0048] At least one processor and at least one memory communicatively connected to the processor; wherein the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the railway foreign object intrusion detection method based on deep learning as described in any one of the first aspects.

[0049] (3) Beneficial effects

[0050] The present invention's deep learning-based railway foreign object intrusion detection method and system uses a pre-fabricated dataset to pre-train an attention semantic segmentation network. This system indirectly detects unknown foreign objects by recognizing the inherent content of railway scenes. Compared to existing technologies, this method does not require the construction of an image dataset for unknown foreign objects, making the process of obtaining a training dataset much simpler.

[0051] The use of an adaptive correction algorithm enables the attention semantic segmentation network to be applied to a variety of different scenarios, effectively avoiding the problem of insufficient generalization of the network model caused by limited data sets;

[0052] The use of the attention mechanism can avoid the storage difficulties caused by excessively large parameters and ensure high detection accuracy;

[0053] Due to the unsupervised approach, it allows the target domain to utilize raw sensor data without any preprocessing, thus avoiding the need for manual labeling of data used in retraining. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] FIG1 is a schematic diagram of a flow chart of a railway foreign object intrusion detection method based on deep learning provided by one embodiment of the present invention;

[0055] FIG2 is a schematic diagram of the structure of a boundary-assisted dual-branch attention semantic segmentation network provided by one embodiment of the present invention;

[0056] FIG3 is a schematic diagram of the structure of an attention mechanism provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.

[0058] The present invention, as presented in the embodiments of the present invention, proposes a deep learning-based method for detecting foreign object intrusion in railways. Using a prefabricated dataset, an attention semantic segmentation network is trained to identify inherent elements of railway scenes, thereby indirectly detecting unknown foreign objects. The network incorporates an attention mechanism to capture global contextual information and extract representative features, achieving a lightweight architecture without compromising accuracy. Furthermore, an adaptive correction algorithm is employed to improve the network's generalization capabilities across different scenarios, enabling easy deployment in a variety of scenarios.

[0059] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. Instead, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0060] As shown in FIG1 , this embodiment provides a railway foreign object intrusion detection method based on deep learning, the method comprising:

[0061] S10: Acquire image data to be detected.

[0062] Optionally, it can be obtained using a vision sensor, which can be installed on the track or deployed next to the track where foreign objects often appear.

[0063] S20: Input the image data to be detected into the trained second attention semantic segmentation network to obtain a foreign body detection result.

[0064] The trained second attention semantic segmentation network is obtained by retraining the first attention semantic segmentation network using the specified image data captured by the visual sensing device and adopting the adaptive correction algorithm.

[0065] The first attention semantic segmentation network is obtained by training a preset attention semantic segmentation network architecture using a prefabricated dataset.

[0066] The prefabricated dataset is obtained by extracting data related to railway scenes from the COCO-Stuff dataset, and is recorded as Coco-stuff-Rail.

[0067] As shown in Figure 2, the attention semantic segmentation network architecture consists of a backbone network and a two-branch decoder. To incorporate sufficient contextual information for semantic segmentation, a GLAM (Global-local attention module) is inserted after each stage of the ResNet (residual neural network). This allows the network to extract global context from the original image from the outset, avoiding the storage difficulties caused by excessive parameters while maintaining high detection accuracy. Specifically, a global attention mechanism, called a GAM (global attention module), is introduced after each ResNet stage. This allows the network to extract global context from the original image at an early stage and fuse local and global features through residual connections. Three parallel dilated convolutions with kernel sizes of 3, 5, and 7 are used as the initial convolution. The output of this layer is obtained by element-wise summing the results of these dilated convolutions. The final backbone network has five stages, with resolutions of 1 / 2, 1 / 4, 1 / 8, 1 / 8, and 1 / 8 relative to the input, respectively, to minimize spatial information loss caused by downsampling.

[0068] Segmentation accuracy relies heavily on accurate edge recognition. Considering the simple composition and fixed elements of railway scenes, this embodiment uses a dual-branch decoder that combines an edge recognition branch and a semantic segmentation branch. The learned edge information is used to assist the segmentation task through a multi-task architecture. Edge labels can be easily obtained based on known semantic labels. This embodiment designates boundary points as pixels with different labels from the surrounding pixels, which allows the boundary information of each material to be obtained, rather than being limited to the boundary between two materials.

[0069] The structure of the attention mechanism is shown in Figure 3. GAM plays a vital role in capturing any dependencies between pixels, effectively expanding the perception field. Specifically, a dual attention mechanism of space and channel is introduced, and the position and channel attention obtain Q, K and V vectors through convolution. This process involves square-level calculations of the input dimension and the convolution channel, resulting in significant computational and storage overhead, making it challenging to directly process the original image. This embodiment improves the calculation of the spatial attention matrix. By selecting the top M largest correlation values ​​in vector K and their corresponding positions in vector V, the weight matrix can be converted from Convert to This effectively reduces parameters. For channel attention, it can be observed that the channel output of ResNet above 50 layers increases four times compared to the channel output below 50 layers. Therefore, the convolution blocks connected between the two levels are modified to keep the channel output of each level to 64, 128, 256, and 512. This modification reduces the computation and storage consumption by 16 times, enabling the use of deeper ResNet to extract higher quality features.

[0070] The adaptive correction algorithm consists of an unsupervised domain adaptation module and a post-correction algorithm. The unsupervised domain adaptation module adopts MK-MMD because it is simple and easy to use and does not require more parameters or new models.

[0071] In the practical application of this embodiment, before S1, the following steps are also included:

[0072] S01. Extract data related to railway scenes from the COCO-Stuff dataset to obtain a prefabricated dataset, denoted as Coco-Stuff-Rail.

[0073] S02. Use the prefabricated dataset to train the attention semantic segmentation network architecture to obtain a first attention semantic segmentation network.

[0074] Specifically, the data in the prefabricated dataset is first input into the attention semantic segmentation network architecture to obtain the output results of the semantic segmentation branch and the edge recognition branch respectively;

[0075] Then, based on the true label of the input data and the output result, the total loss of the two branches of the attention semantic segmentation network architecture is calculated using the two-branch total loss expression;

[0076] The total loss expression of the two branches is: L Z =L v(x) +L ζ(x) +L b(x) ,

[0077] Among them, L Z represents the total loss of the two branches, L v(x ) represents the cross entropy loss of category balance weight and positive and negative sample balance weight, L ζ(x) is the edge auxiliary loss, L b(x) Branch loss for edge identification;

[0078] The edge-assisted loss is calculated using an edge-assisted loss expression;

[0079] The edge-assisted loss expression is:

[0080] Among them, L ζ(x) is the edge identification branch loss, A(xn ) d 、A(x n ) b Represent the output results of the semantic segmentation branch and the edge branch respectively, Represent semantic labels and edge labels respectively; Same as L v(x) The class balance weights are the same, N represents the number of samples, and n represents the sample index;

[0081] Repeat S021 and S022 until the total loss value of the two branches converges to obtain the trained first attention semantic segmentation network.

[0082] S03. Use the specified image data captured by the visual sensing device and adopt an adaptive correction algorithm to retrain the first attention semantic segmentation network to obtain a second attention semantic segmentation network.

[0083] Specifically, the source domain and the target domain are first input into the first attention semantic segmentation network to obtain the generated result;

[0084] The source domain is the prefabricated dataset, and the target domain is the specified image data captured by the visual sensing device;

[0085] Then, the distribution difference calculation expression is used to calculate the distribution difference between the source domain and the target domain, which is the loss of guiding the network to adapt to the target domain;

[0086] The distribution difference calculation expression is:

[0087] Among them, L Φ(s,t) Represents the distribution difference between the source domain and the target domain, s and t represent the source domain and target domain samples respectively, using different kernel functions G k (·) is mapped to the Hilbert space to calculate the expected gap; K represents the number of kernel functions used, k represents the kernel function index; P represents the number of source domain samples, p represents the source domain sample index; Q represents the number of target domain samples, q represents the target domain sample index; P and Q are equal;

[0088] Then, according to the calculated distribution difference, a post-correction algorithm is used to correct the generated result to obtain a corrected result, and the corrected result is compared with the corresponding target domain sample to obtain a correction loss;

[0089] Repeat S031-S033 until the correction loss is less than a preset threshold, and obtain a trained second attention semantic segmentation network.

[0090] On the other hand, this embodiment also provides a railway foreign object intrusion detection system based on deep learning, characterized in that the system includes:

[0091] At least one processor and at least one memory communicatively connected to the processor; wherein the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute any of the railway foreign object intrusion detection methods based on deep learning as described above in the embodiments.

[0092] Wherein, the memory is used to store computer programs;

[0093] The processor is configured to execute the program stored in the memory by performing the following steps:

[0094] Acquire image data to be detected;

[0095] Inputting the image data to be detected into the trained second attention semantic segmentation network to obtain a foreign object detection result;

[0096] The trained second attention semantic segmentation network is obtained by retraining the first attention semantic segmentation network using the specified image data captured by the visual sensing device and an adaptive correction algorithm;

[0097] The first attention semantic segmentation network is obtained by training the preset attention semantic segmentation network architecture using a pre-made dataset;

[0098] The prefabricated dataset is obtained by extracting data related to railway scenes from the COCO-Stuff dataset;

[0099] The attention semantic segmentation network architecture includes a backbone network architecture and a dual-branch decoder;

[0100] Among them, the backbone network architecture is obtained by inserting a specified attention mechanism at a specified position in a pre-selected residual neural network, and using three parallel dilated convolutions as the initial convolution layer in the residual neural network; the dual-branch decoder is a dual-branch decoder that combines an edge recognition branch and a semantic segmentation branch.

[0101] The aforementioned memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk storage. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0102] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0103] Furthermore, the processor utilizes various interfaces and lines to connect various parts of the system, and performs various functions and processes data by running or executing software programs and / or modules stored in the memory, and calling data stored in the memory.

[0104] Optionally, the processor may include one or more processing cores; the processor may integrate an application processor and a modem processor, wherein the application processor primarily handles operating storage media, user interfaces, and application programs, and the modem processor primarily handles wireless communications. It is understood that the modem processor may not be integrated into the processor.

[0105] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0106] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; and internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0107] In the present invention, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Furthermore, when a first feature is "above," "above," or "above" a second feature, it may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is at a higher level than the second feature. When a first feature is "below," "below," or "below" a second feature, it may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is at a lower level than the second feature.

[0108] In the description of this specification, the terms "one embodiment", "some embodiments", "embodiments", "examples", "specific examples" or "some examples" refer to the specific features, structures, materials or characteristics described in conjunction with the embodiment or example and included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and features of different embodiments or examples, unless they are mutually inconsistent.

[0109] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may alter, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A railway foreign body intrusion detection method based on deep learning, characterized in that: The method comprises: S10, obtaining image data to be detected; S20, inputting the image data to be detected into the trained second attention semantic segmentation network to obtain a foreign body detection result; The trained second attention semantic segmentation network is obtained by retraining the first attention semantic segmentation network using specified image data captured by a visual sensing device and an adaptive correction algorithm; The first attention semantic segmentation network is obtained by training a preset attention semantic segmentation network architecture using a prefabricated dataset; The prefabricated dataset is obtained by extracting data associated with railway scenes from the COCO-Stuff dataset; The attention semantic segmentation network architecture includes a backbone network architecture and a dual-branch decoder; Among them, the backbone network architecture is obtained by inserting a specified attention mechanism at a specified position in a preselected residual neural network, and using three parallel dilated convolutions as the initial convolution layer in the residual neural network; the dual-branch decoder is a dual-branch decoder that combines an edge recognition branch and a semantic segmentation branch.

2. The railway foreign object intrusion detection method based on deep learning according to claim 1 is characterized in that: Before S10, the method further includes: S01. Extract data related to railway scenes from the COCO-Stuff dataset to obtain a prefabricated dataset. S02. Using the prefabricated dataset, training the attention semantic segmentation network architecture to obtain a first attention semantic segmentation network; S03. Use the specified image data captured by the visual sensing device and adopt an adaptive correction algorithm to retrain the first attention semantic segmentation network to obtain a second attention semantic segmentation network.

3. The railway foreign object intrusion detection method based on deep learning according to claim 1 is characterized in that: The designated position is after each of the five stages in the residual neural network; The attention mechanism is a global and local fusion attention mechanism.

4. The railway foreign object intrusion detection method based on deep learning according to claim 1 is characterized in that: The kernel sizes of the three parallel dilated convolutions are 3, 5, and 7, respectively.

5. The railway foreign object intrusion detection method based on deep learning according to claim 4 is characterized in that: There are five stages in the backbone network architecture, and the data resolution output by each stage is 1 / 2, 1 / 4, 1 / 8, 1 / 8, and 1 / 8 respectively relative to the input data.

6. The railway foreign body intrusion detection method based on deep learning according to claim 2 is characterized in that: S02 includes: S021. Input the data in the prefabricated dataset into the attention semantic segmentation network architecture to obtain output results of the semantic segmentation branch and the edge recognition branch respectively; S022. Based on the true label of the input data and the output result, use the two-branch total loss expression to calculate the total loss of the two branches of the attention semantic segmentation network architecture; The total loss expression of the two branches is: L Z =L v(x) +L ζ(x) +L b(x) , Among them, L Z represents the total loss of the two branches, L v(x) represents the cross entropy loss of the category balance weight and the positive and negative sample balance weight, L ζ(x) is the edge auxiliary loss, L b(x) Branch loss for edge identification; The edge-assisted loss is calculated using an edge-assisted loss expression; The edge-assisted loss expression is: Among them, L ζ(x) is the edge identification branch loss, A(x n ) d 、A(x n ) b Represent the output results of the semantic segmentation branch and the edge branch respectively, Represent semantic labels and edge labels respectively; Same as L v(x) The class balance weights are the same, N represents the number of samples, and n represents the sample index; S023. Repeat S021 and S022 until the total loss value of the two branches converges to obtain the trained first attention semantic segmentation network.

7. The railway foreign object intrusion detection method based on deep learning according to claim 2 is characterized in that: The adaptive correction algorithm is a domain adaptive correction algorithm based on multi-core maximum mean square deviation and is connected after each attention mechanism; The adaptive correction algorithm includes an unsupervised domain adaptive algorithm and a post-correction algorithm.

8. The railway foreign body intrusion detection method based on deep learning according to claim 7 is characterized in that: S03 includes: S031. Input the source domain and the target domain into the first attention semantic segmentation network to obtain a generated result; The source domain is the prefabricated dataset, and the target domain is the specified image data captured by the visual sensing device; S032. Use the distribution difference calculation expression to calculate the distribution difference between the source domain and the target domain, that is, to guide the network to adapt to the loss of the target domain; The distribution difference calculation expression is: Among them, L Φ(s，t) Represents the distribution difference between the source domain and the target domain, s and t represent the source domain and target domain samples respectively, using different kernel functions G k (·) is mapped to the Hilbert space to calculate the expected gap; K represents the number of kernel functions used, k represents the kernel function index; P represents the number of source domain samples, p represents the source domain sample index; Q represents the number of target domain samples, q represents the target domain sample index; P and Q are equal; S033. Correct the generated result using a post-correction algorithm based on the calculated distribution difference to obtain a corrected result, and compare the corrected result with the corresponding target domain sample to obtain a correction loss; S034. Repeat S031-S033 until the correction loss is less than a preset threshold, and obtain a trained second attention semantic segmentation network.

9. The railway foreign object intrusion detection method based on deep learning according to claim 1 is characterized in that: The method further comprises: S30. If the foreign object detection result is a foreign object, an alarm is sent to the driver, and the alarm is fed back to the back-end management personnel for verification.

10. A railway foreign object intrusion detection system based on deep learning, characterized in that: The system comprises: At least one processor and at least one memory communicatively connected to the processor; wherein the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the railway foreign object intrusion detection method based on deep learning as described in any one of claims 1 to 9.

Images