Urban traffic rail invading foreign matter detection method based on improved YOLOv11

Abstract

The invention relates to an urban traffic rail invading foreign matter detection method based on improved YOLOv11, and aims to solve the problems of poor environmental adaptability, unbalanced precision and real-time performance and high deployment cost in the prior art. The method comprises the following steps: firstly, acquiring an urban traffic rail area image, and inputting the image into an improved detection model to complete invasion foreign matter identification; according to the model, a Laplacian-Gaussian module is arranged at the input end of a YOLOv11 backbone network to enhance image edge and texture features, a C3k2-MSEPA module integrating multiple scales to enhance parallel attention is used for replacing an original C3k2 module, a GSConv module with a gating mechanism is added to the joint of a backbone network and a neck network, and a boundary-sensitive Shape-NWD loss function is matched for optimization training. According to the scheme, the detection stability under the background of complex illumination, severe weather and weak texture is greatly improved, the features of foreign matters with different scales are efficiently captured, the detection precision and the real-time performance are both considered, the parameter quantity is controllable, high-performance hardware support is not needed, the deployment cost is reduced, invasion foreign matters such as pedestrians, vehicles and abandoned articles in a track area can be accurately identified, and the method is suitable for popularization and application. And reliable guarantee is provided for the operation safety of urban rail transit.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of target detection, and particularly relates to a city traffic track intrusion foreign matter detection method based on an improved YOLOv11. BACKGROUND

[0002] As the core backbone of urban public transportation, the operation safety of urban rail transit is directly related to the life safety of millions of passengers and the stable operation of the urban transportation network. If there are intruding foreign matters such as pedestrians, vehicles and abandoned objects in the track area, it is easy to cause major safety accidents such as train emergency braking and derailment, causing serious personnel casualties and economic losses. Therefore, developing efficient and accurate urban rail transit intrusion foreign matter detection technology to realize real-time monitoring and abnormal early warning of the track area is a key link to ensure the safe operation of urban rail transit, and has important practical significance and application value.

[0003] Current urban rail transit intrusion foreign matter detection technologies mainly fall into two categories: contact type and non-contact type. The contact type scheme requires large-scale deployment of pressure sensors, vibration sensors or optical fiber sensors along the track, as well as a complex trackside cabinet and power supply system. Not only is the construction cost high and the construction period long, but after the alarm, the type of intruding foreign matter needs to be confirmed by manual on-site confirmation, and the disposal link is long, which is easy to delay the emergency handling opportunity. In the non-contact type scheme, computer vision technology has been widely concerned due to its flexible deployment and intuitive results. Early, traditional vision algorithms (such as background difference method and inter-frame difference method) were used for foreign matter identification, but these algorithms had high false alarm rates under sudden changes in light, bad weather such as rain, snow and fog, and complex background interference. In recent years, deep learning target detection algorithms have emerged, which have improved detection accuracy, but have performance bottlenecks. Single-stage detection algorithms (such as early YOLO series) have fast detection speed, but based on anchor mechanism, they easily produce a large number of redundant candidate boxes and have insufficient small target detection accuracy. Two-stage detection algorithms have high accuracy, but they need to repeatedly select candidate regions and adjust the bounding box, which is slow in detection speed and difficult to meet the real-time monitoring demand. Moreover, most algorithm models have large parameter quantity and high calculation cost, which require high-performance hardware support and are difficult to adapt to the actual demand of distributed deployment of multiple cameras in the urban rail transit scene.

[0004] Current technologies still cannot effectively address the core pain points of intrusion detection in urban rail transit scenarios: First, they have poor adaptability to complex environments, struggling to reliably extract foreign object features in the face of changing lighting, inclement weather, and weak track texture backgrounds, resulting in high false alarm and false negative rates. Second, it is difficult to balance accuracy and speed; either prioritizing speed sacrifices accuracy for small targets, or increasing accuracy leads to detection delays, failing to meet real-time early warning requirements. Third, model deployment costs are high; complex deep learning models have stringent hardware performance requirements, making them difficult to widely apply in existing rail transit monitoring hardware systems. Therefore, there is an urgent need for an intrusion foreign object detection solution that can adapt to the complex scenarios of urban rail transit, balance detection accuracy and real-time performance, and has controllable deployment costs to compensate for the shortcomings of existing technologies. Summary of the Invention

[0005] To address the problems existing in the background art, one aspect of the present invention provides a method for detecting foreign objects intruding into urban traffic tracks based on an improved YOLOv11, comprising: acquiring track images of urban traffic track areas; inputting the track images into a track intrusion detection model designed based on the improved YOLOv11 to identify track intrusions and obtain track intrusion identification results; wherein, the track intrusion detection model comprises: setting a Laplacian-Gaussian module at the input end of the YOLOv11 backbone network to enhance the edges and textures of the input track images; replacing the C3k2 module in the YOLOv11 backbone network and neck network with a C3k2-MSEPA module integrating multi-scale enhanced parallel attention to improve the network's ability to process features at different scales of the image; and simultaneously adding GSConv modules with gating mechanisms at the three connections of the YOLOv11 backbone network and neck network to improve the model's ability to capture details of track intrusions; thus obtaining the foreign object detection model.

[0006] Another aspect of the present invention provides an urban traffic track intrusion foreign object detection system based on an improved YOLOv11, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned urban traffic track intrusion foreign object detection method based on an improved YOLOv11.

[0007] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for detecting intrusive foreign objects in urban traffic tracks based on an improved YOLOv11.

[0008] The present invention has at least the following beneficial effects

[0009] This invention effectively enhances the edge and texture features of track images by setting a Laplacian-Gaussian module at the input of the YOLOv11 backbone network, improving the stability of foreign object feature extraction under complex lighting, severe weather, and weak texture backgrounds, thus solving the problem of poor environmental adaptability of existing technologies. It replaces the original C3k2 module with a C3k2-MSEPA module that integrates multi-scale enhanced parallel attention, and adds a GSConv module with a gating mechanism at the key connection between the backbone and neck networks. Combined with the boundary-sensitive Shape-NWD loss function, it significantly improves the detection accuracy and detail capture capability of intrusive foreign objects at different scales, achieving a balance between detection accuracy and real-time performance. This overcomes the shortcomings of existing technologies, which are either insufficient in speed or have a high false negative rate for small targets. Simultaneously, the model reasonably controls the number of parameters and computational load during performance optimization, eliminating the need for high-performance hardware, reducing deployment costs, and adapting to the actual needs of multi-camera distributed monitoring in urban rail transit. This provides an efficient, accurate, and economical guarantee for intrusive foreign object detection in rail operation safety. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0011] Figure 2 This is a schematic diagram of the LoGStem module of the present invention;

[0012] Figure 3 This is a schematic diagram of the structure of the C3k2-MSEPA module of the present invention;

[0013] Figure 4 This is a schematic diagram of the attention module of the EPA module of the present invention;

[0014] Figure 5 This is a schematic diagram of the GSConv module of the present invention. Detailed Implementation

[0015] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0016] Please see Figure 1One aspect of the present invention provides a method for detecting foreign objects intruding into urban traffic tracks based on an improved YOLOv11, comprising: acquiring track images of urban traffic track areas; inputting the track images into a track intrusion detection model designed based on the improved YOLOv11 to identify track intrusions and obtain track intrusion identification results; wherein, the track intrusion detection model comprises: setting a Laplacian-Gaussian module at the input end of the YOLOv11 backbone network to enhance the edges and textures of the input track images; replacing the C3k2 module in the YOLOv11 backbone network and neck network with a C3k2-MSEPA module integrating multi-scale enhanced parallel attention to improve the network's ability to process features at different scales of the image; and simultaneously adding GSConv modules with gating mechanisms at the three connections of the YOLOv11 backbone network and neck network to improve the model's ability to capture details of track intrusions; thus obtaining the foreign object detection model.

[0017] Please see Figure 2 Preferably, the Laplacian-Gaussian module comprises: a first convolutional layer, a LoG filter, a first BN+SiLU activation function layer, a second convolutional layer, a first deconvolutional layer, a first Gaussian filter, a first BN layer, a second Gaussian filter, and a DRFD module, cascaded sequentially; wherein the input features of the first convolutional layer and the output features of the first BN+SiLU activation function layer are summed to obtain the input features of the second convolutional layer; the input features of the first Gaussian filter and the output features of the first Gaussian filter are summed to obtain the input features of the first BN layer.

[0018] Preferably, the DRFD module includes: a third convolutional layer, a second BN layer, a fourth convolutional layer, a first Maxpool layer, a second BN+SiLU activation function layer, a third BN layer, a Concat concatenation function, and a fifth convolutional layer; wherein, the output features of the fifth convolutional layer serve as the output features of the DRFD module; the input features of the third convolutional layer serve as the input features of the DRFD module; the output features of the third convolutional layer serve as the input features of the second BN layer; the output features of the second BN layer serve as the input features of the fourth convolutional layer and the first Maxpool layer, respectively; the output features of the fourth convolutional layer serve as the input features of the second BN+SiLU activation function layer; the output features of the first Maxpool layer serve as the input features of the third BN layer; and the output features of the second BN+SiLU activation function layer and the third BN layer are concatenated using the Concat concatenation function and then used as the input features of the fifth convolutional layer.

[0019] In this embodiment, the Laplace-Gaussian module (LoGStem) utilizes the noise reduction of the Gaussian filter and the edge awareness of the Laplace to improve the robustness of edge information extraction.

[0020] The LoG filter combines Gaussian smoothing and Laplacian sharpening techniques. The Gaussian component reduces noise, while the Laplacian operator highlights areas with rapidly changing intensity. (Input image) First, a 7×7 convolution transformation is performed to extract initial features, then a kernel size of 7×7 is applied and... The LoG filter enables the network to learn feature representations and subsequently enhance its edge features. At the feature pixel location... The LoG filter is defined as follows:

[0021]

[0022] in, and Represents pixel coordinates; This represents the standard deviation of Gaussian. Indicates the kernel size. Represents pi (π). Represents the natural base; The standard deviation is expressed as The kernel size is LoG filter;

[0023] At feature pixel location Gaussian filter The calculation formula is:

[0024]

[0025] in, Indicates the kernel size. Standard deviation; Standard deviation is The kernel size is Gaussian filter.

[0026] The output of the LoG filter is passed through an activation function (SiLU) and batch normalization (BacthNorm), and then combined with the original image I:

[0027]

[0028] This residual-like structure ensures smooth gradient flow and preserves original image details and edge information. Subsequently... Downsampling is performed through two convolutional layers to generate :

[0029]

[0030] in, It is a 3×3 convolution with a stride of 2.

[0031] To further refine the feature maps and integrate multi-scale information After processing with two Gaussian filters, the results are summed, normalized, and then fed into a DRFD module with global and local sensing capabilities for downsampling. The processing procedure is as follows:

[0032]

[0033]

[0034]

[0035] in, and These are Gaussian filters with kernel sizes of 9×9 and 5×5, and a standard deviation of 0.5, respectively. , It is a 1×1 convolution kernel with a stride of 1 and a 3×3 convolution with a stride of 2. MaxPool is the max pooling operation.

[0036] By employing a cascaded design of convolutional layers, LoG filters, Gaussian filters, and DRFD modules using a Laplacian-Gaussian module, combined with a residual structure of feature addition, the Laplacian-Gaussian module can reduce noise interference in orbital images through Gaussian filtering and highlight the edge features of foreign objects through the Laplacian operator. The DRFD module then integrates multi-scale local and global features, effectively improving the edge and texture representation accuracy of orbital images under extreme weather, occlusion, and weak texture backgrounds. This provides a clearer and more robust foundation for subsequent foreign object feature extraction, helping to reduce the risk of missing small target foreign objects.

[0037] Please see Figure 3 Preferably, the C3k2-MSEPA module includes a C3k2 module, an MSPC module, and an EPA module cascaded in sequence.

[0038] In this embodiment, MSEPA consists of a multi-scale parallel convolution module (MSPC) and an enhanced parallel attention module (EPA). This module significantly improves the accuracy of multi-scale target detection through the synergistic effect of multi-scale feature fusion and multiple attention mechanisms.

[0039] Preferably, the MSPC module comprises: a fourth BN layer, a first PW convolutional layer, a sixth convolutional layer, a first GELU activation function, a depth dilation layer, a second PW convolutional layer, a second GELU activation function, and a third PW convolutional layer, cascaded sequentially; wherein the input features of the fourth BN layer and the output features of the third PW convolutional layer are summed to obtain the output features of the MSPC module; the depth dilation layer comprises multiple parallel DWDConv layers of different scales; the output features of the first GELU activation function are respectively used as the input features of multiple DWDConv layers of different scales; the output features of the multiple DWDConv layers of different scales are concatenated and used as the input features of the second PW convolutional layer.

[0040] In this embodiment, the MSPC module has multi-scale characteristics and a large receptive field. First, assuming x is the original feature map, it is normalized using batch normalization (BatchNorm), i.e. Batch normalization can accelerate network convergence, improve generalization ability, and prevent overfitting. The MSPC process is as follows:

[0041]

[0042]

[0043]

[0044] Here, PWConv represents pointwise convolution. Gonv represents a convolution with a kernel size of 5, and GELU is the activation function. DWDConv7, DWDConv13, and DWDConv19 represent depthwise separable dilated convolutions of 7×7, 13×13, and 19×19 with dilation rates of 3, respectively. DWDConv5 represents a dilated convolution with a kernel size of 5, which is a 5×5 depthwise separable dilated convolution with a dilation rate of 2. Finally, Concat represents concatenating features along the channel dimension.

[0045] Four parallel convolutions with different kernel sizes can extract multi-scale features. Large and medium-sized dilated convolutions possess long-range modeling capabilities and large receptive fields, effectively capturing features of interference areas and large targets in a wide-ranging railway environment. Small-sized dilated convolutions focus on local railway interference areas and small targets, restoring the texture details of the targets. We connect multi-scale information along the channel dimension, enabling… Feature dimensions become Four times that.

[0046]

[0047] Then, The input is fed into the multilayer perceptron, and Feature dimension adjustment and The same. This multilayer perceptron contains two pointwise convolutions and uses GELU as the activation function. Finally, the output of the multilayer perceptron is the same as... Adding to obtain multi-scale perceptual feature maps .

[0048] Preferably, the EPA module includes: a fifth BN layer, an attention module, a fourth PW convolutional layer, a third GELU activation function, and a fifth PW convolutional layer cascaded in sequence; the input features of the fifth BN layer and the output features of the fifth PW convolutional layer are summed to obtain the output features of the EPA module; the attention module includes a parallel scattering attention layer, a simple pixel attention layer, a channel attention layer, and a pixel attention layer; the output features of the fifth BN layer are used as the input features of the scattering attention layer, the simple pixel attention layer, the channel attention layer, and the pixel attention layer, respectively; the output features of the scattering attention layer, the simple pixel attention layer, the channel attention layer, and the pixel attention layer are concatenated to obtain the input features of the fourth PW convolutional layer.

[0049] In this embodiment, the EPA module integrates different types of attention mechanisms, including descattering attention, simple pixel attention, channel attention, and pixel attention, to enhance the perception of multi-scale features from different perspectives, such as... Figure 3 and Figure 4 As shown. Let x be a feature map, and standardize it through batch normalization to obtain... .

[0050] Descattering attention primarily suppresses scattering noise in extreme weather conditions and enhances the feature representation of sharp areas. Branches such as Figure 4 As shown in (a), this branch generates attention weights. The formula is:

[0051]

[0052]

[0053] Where PWConv represents pointwise convolution, Conv represents convolution with a kernel size of 3, GELU and Sigmoid are activation functions, and IN is instance normalization; Represents the tensor product.

[0054] The simple pixel attention module consists of two branches: and like Figure 4 As shown in (b). It is the feature extraction branch. It's a pixel-gated branch used to refine local features. As The gating signal. Its calculation process can be expressed as:

[0055]

[0056]

[0057]

[0058] Channel attention can efficiently extract global information and change the channel dimension of features. Channel attention has... This branch focuses on salient features within the global perspective, such as... Figure 4 As shown in (c). The formula is:

[0059]

[0060]

[0061] Global channel-gated features are extracted using PWConv-GELU-PWConv, the sigmoid function, and Global Average Pooling (GAP). As The global channel gating signal.

[0062] Pixel attention can effectively extract location-related information features, such as Figure 4 As shown in (d), it consists of pointwise convolution and GELU, focusing on fine-grained perception and improving attention to small-scale objects. The specific formula is:

[0063]

[0064]

[0065] The PWConv-GELU-PWConv structure is used to fit the features. The Sigmoid function is used to extract global pixel-gated features, and then... As The global pixel gating signal.

[0066] Four different attention outcomes are connected along the channel dimension. Then, a multilayer perceptron structure is applied to reduce the channel dimension of the connected features to the same dimension as the input x. Finally, the output of the multilayer perceptron is added to x. The formula is as follows:

[0067]

[0068]

[0069] The enhanced perception attention module improves object detection at multiple field-of-view granularities by processing multi-scale features through attention mechanisms with different emphases, thereby enhancing robustness to object distribution.

[0070] The MSPC module leverages multi-scale DWDConv to extract features across different ranges in parallel, balancing large-scale environmental interference with the capture of local foreign object details. The EPA module utilizes multiple attention mechanisms, such as scattering attention and channel attention, to enhance feature perception from different dimensions. Together, they achieve efficient extraction and optimization of multi-scale and multi-dimensional features of foreign objects intruding into the orbit, significantly improving the network's ability to characterize foreign objects of different sizes and shapes. This effectively alleviates the detection challenges caused by large differences in the scale and mixed features of foreign objects in orbital scenarios, laying a solid foundation for accurate identification.

[0071] Please see Figure 5 Preferably, the GSConv module includes two branches; the first branch includes a first BottConv layer, a first GN layer, a first ReLU activation function, a second BottConv layer, a second GN layer, and a second ReLU activation function, all cascaded in sequence; the second branch includes a third BottConv layer, a third GN layer, a third ReLU activation function, a seventh convolutional layer, a sixth BN layer, and a first Sigmoid activation function, all cascaded in sequence; the output features of the first and second branches are multiplied together, then passed through a fourth BottConv layer, a fourth GN layer, and a fourth ReLU activation function, and finally added to the input features of the GSConv module to obtain the output features of the GSConv module.

[0072] In this embodiment, a GSConv module utilizing a gating mechanism is designed. The gating mechanism enables dynamic features for each spatial location and channel, enhancing the model's ability to capture target details in complex urban rail environments. Due to challenges such as light attenuation, scattering, and interference from suspended particles in urban rail environments, a bottleneck convolution (BottConv) with a low-rank approximation is embedded to map feature representations from a high-dimensional space to a low-dimensional space. The advantages of BottConv in urban rail detection are: point-wise convolution (PointConv) can effectively extract multi-scale features of the target, while depth convolution (DepthConv) performs spatial information extraction in a low-dimensional subspace, which is beneficial for handling problems such as blurred target boundaries and unclear texture details in urban rail environments.

[0073] like Figure 5 As shown, input features Reserved as To form residual connections to facilitate information flow.

[0074] Subsequently, features The BottConv layer is followed by group normalization and ReLU activation to generate features. In the BottConv layer, pointwise convolution is used to reduce dimensionality and highlight the main features. Then, depthwise separable convolution is used to capture background texture information, and pointwise convolution is used to restore dimensionality and integrate information, thus achieving dynamic adjustment. The specific calculation is shown in formula (13):

[0075]

[0076]

[0077]

[0078] For Group Normalization (GN), Batch Normalization (BN) can also be used, prioritizing the GN layer and adding a gate mechanism. The Gate and Gate are combined through multiplication to highlight important features. Suppress irrelevant information:

[0079]

[0080]

[0081] Where Conv represents a convolution with a kernel size of 1, and BN stands for BatchNorm, representing batch normalization. The gated features are further refined using BottConv. After residual connections, the final output is:

[0082]

[0083]

[0084] Finally, this design enables the model to dynamically enhance the distinguishability of targets from complex urban environments while preserving the characteristics of urban rail intrusions, achieving better detection accuracy in the face of challenges such as changes in lighting conditions and abnormal weather.

[0085] Preferably, the orbital intrusion foreign object detection model is trained using a boundary-sensitive Shape-NWD loss function, including:

[0086]

[0087]

[0088]

[0089]

[0090]

[0091] in, Represents the loss function; This represents a constant that is set. Represents the natural base; Indicates the scale factor; and These are the weighting coefficients for the horizontal and vertical directions, respectively; and Indicates the center coordinates of the actual annotation box; and Indicates the center coordinates of the prediction box; and Indicates the width and height of the prediction box; and This indicates the width and height of the actual annotation box; This represents a defined constant.

[0092] In this embodiment, the traditional IoU boundary regression method mainly considers the geometric relationship between the predicted bounding box and the ground truth bounding box when calculating the loss, but ignores the influence of the bounding box's shape and scale on the regression results. In actual railway target detection scenarios, small targets occupy fewer pixels in the image, and their features are relatively inconspicuous, making accurate detection of small targets more difficult. Therefore, this embodiment incorporates the idea of ​​Shape-IoU into the NWD loss function. Shape-NWD adaptively adjusts the regression weights in different directions according to the shape features of the intruder and considers the influence of the bounding box scale on the regression, improving the detection accuracy of targets with different regular shapes. In extreme weather conditions with poor image quality, shape prior information guides the bounding box regression, reducing localization errors caused by environmental interference.

[0093] Another aspect of the present invention provides an urban traffic track intrusion foreign object detection system based on an improved YOLOv11, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned urban traffic track intrusion foreign object detection method based on an improved YOLOv11.

[0094] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for detecting intrusive foreign objects in urban traffic tracks based on an improved YOLOv11.

[0095] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0096] Experiment and Results Analysis

[0097] Dataset:

[0098] Currently, there is a lack of publicly available datasets specifically designed for foreign object intrusion in urban rail scenarios. Existing general object detection data struggles to reflect the complex background structure of rail tracks, the large differences in target scale, and the variable imaging conditions of real-world monitoring. To support the effective modeling of these key factors by deep learning models, the Crail dataset, covering typical intrusion categories, was constructed. Ten common intrusion objects on urban rail tracks—pedestrians, bicycles, ebikes, three-wheeled vehicles, bottles, bags, cars, buses, trucks, and books—were used as experimental subjects. The Crail dataset consists of 4000 images, divided into training, validation, and test sets in an 8:1:1 ratio. 3200 images are used for training, 400 for validation, and 400 for the test. The image size is consistently maintained at approximately 600×800 pixels.

[0099] Experimental environment

[0100]

[0101] Evaluation indicators

[0102] This paper uses the following metrics to evaluate the model's performance: mean accuracy (mAP), detection speed (FPS), and computational cost (GLOPs). The calculation formulas are as follows:

[0103]

[0104]

[0105] Where TP is the correctly predicted positive sample class, FP is the incorrectly predicted positive sample class, and FN is the incorrectly predicted negative sample class.

[0106] Average accuracy (AP) is the ratio of coordinate axes and The boundary region of the precision-recall curve, with the precision formula for each category as follows:

[0107]

[0108] Mean accuracy (mAP) is an evaluation metric for the accuracy of all detection categories, and the formula is as follows:

[0109]

[0110] in, It represents the number of categories.

[0111] Detection speed is primarily evaluated using floating-point operations (FLOPs), frames per second (FPS), and network parameters. Parameters are the total number of parameters the model needs to learn. FLOPs represent the number of floating-point operations performed during the learning process. More floating-point operations result in higher computational complexity and higher computational costs. FPS represents the number of images detected and recognized per second. A higher value indicates faster detection. The formulas for these metrics are as follows:

[0112]

[0113]

[0114]

[0115] in, Represents the size of the convolution kernel. and These represent the number of input and output channels, respectively. and These represent the height and width of the input feature map, respectively. The number of frames represented in the statistics. This represents the total time (in seconds) spent.

[0116] Experimental Results and Analysis

[0117] To demonstrate the effectiveness of this invention, simulations were performed on the same dataset as existing classical algorithms. The simulation results are shown in Table 6.

[0118] Table 6 Comparison results between the classic algorithm and the algorithm of this application

[0119]

[0120] Among all the comparison algorithms, the LGM-YOLOv11 proposed in this paper stands out, with an mAP50 of 89.3% and an mAP50-95 of 62.1%, which is superior to other comparison algorithms.

[0121] In terms of accuracy, LGM-YOLOv11 achieves a 4.9 percentage point improvement in mAP50 compared to the second-best performing YOLOv11n algorithm. This improvement is mainly attributed to the LoGStem module's effective enhancement of edge features in extreme weather images and the GSConv gated activation convolution module's accurate capture of target details.

[0122] From an efficiency perspective, YOLOv5n exhibits the best lightweight characteristics, with only 3.19M parameters and 7.7G of computation, but its detection accuracy is relatively low. YOLOv11n performs well in terms of parameter control, with only 3.58M parameters, but its mAP50 is 84.4%, which is 4.9 percentage points lower than LGM-YOLOv11.

[0123] RetinaNet and VFNet are similar in terms of parameter count, but their detection performance is inferior to LGM-YOLO11. It's worth noting that LGM-YOLOv11 achieves the best detection accuracy while having 5.52M parameters and 18.1G of computation. Although this is an increase compared to ultra-lightweight models, it remains within a reasonable range.

[0124] Overall, LGM-YOLOv11 achieved the best performance across all key performance metrics, significantly outperforming other algorithms in mAP, while maintaining a reasonable model complexity and achieving a better balance between speed and accuracy. In contrast, other models fall short in detection accuracy and efficiency, making them unsuitable for complex scenarios.

[0125] This paper proposes an LGM-YOLOv11 detection framework for urban railway tracks. Through improvements in key stages such as feature extraction, cross-scale modeling, and boundary regression, the framework effectively mitigates the impact of extreme weather, lighting variations, and complex backgrounds on detection performance. The model sequentially introduces LoGStem to enhance edge and texture information in degraded images, utilizes MSEPA for detail enhancement under multi-scale parallel attention, improves feature discrimination capabilities in complex urban environments through the GSConv gating mechanism, and achieves more stable boundary regression performance through Shape-NWD design. Experimental results demonstrate that the proposed method achieves significant performance improvements over baseline models in urban railway track intrusion detection, validating the effectiveness of the overall structural design and key modules.

[0126] In summary, this invention effectively enhances the edge and texture features of track images by setting a Laplacian-Gaussian module at the input of the YOLOv11 backbone network, improving the stability of foreign object feature extraction under complex lighting, severe weather, and weak texture backgrounds, thus solving the problem of poor environmental adaptability in existing technologies. It replaces the original C3k2 module with a C3k2-MSEPA module integrating multi-scale enhanced parallel attention, and adds a GSConv module with a gating mechanism at the key connection between the backbone and neck networks, coupled with a boundary-sensitive Shape-NWD loss function, significantly improving the detection accuracy and detail capture capability of intrusive foreign objects at different scales. This achieves a balance between detection accuracy and real-time performance, overcoming the shortcomings of existing technologies that are either too slow or have a high false negative rate for small targets. Simultaneously, the model reasonably controls the number of parameters and computational load during performance optimization, eliminating the need for high-performance hardware, reducing deployment costs, and adapting to the actual needs of multi-camera distributed monitoring in urban rail transit, providing efficient, accurate, and economical intrusive foreign object detection assurance for track operation safety.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting intrusive foreign objects on urban transportation tracks based on an improved YOLOv11, characterized in that, include: Track images of urban traffic rail areas are acquired, and these images are input into a track intrusion detection model based on an improved YOLOv11 design to identify intrusions and obtain intrusion detection results. The track intrusion detection model includes: adding a Laplacian-Gaussian module at the input of the YOLOv11 backbone network to enhance the edges and textures of the input track images; replacing the C3k2 module in the YOLOv11 backbone and neck network with a C3k2-MSEPA module integrating multi-scale enhanced parallel attention to improve the network's ability to process features at different image scales; and adding GSConv modules with gating mechanisms at all three connections in the YOLOv11 backbone and neck network to improve the model's ability to capture details of track intrusions; thus obtaining the intrusion detection model.

2. The method for detecting intrusive foreign objects on urban transportation tracks based on the improved YOLOv11 according to claim 1, characterized in that, The Laplacian-Gaussian module includes: a first convolutional layer, a LoG filter, a first BN+SiLU activation function layer, a second convolutional layer, a first deconvolutional layer, a first Gaussian filter, a first BN layer, a second Gaussian filter, and a DRFD module, all cascaded sequentially. The input features of the first convolutional layer and the output features of the first BN+SiLU activation function layer are summed to obtain the input features of the second convolutional layer; the input features of the first Gaussian filter and the output features of the first Gaussian filter are summed to obtain the input features of the first BN layer.

3. The method for detecting intrusive foreign objects on urban transportation tracks based on the improved YOLOv11 according to claim 2, characterized in that, The DRFD module includes: a third convolutional layer, a second batch normalization (BN) layer, a fourth convolutional layer, a first maxpool layer, a second BN+SiLU activation function layer, a third BN layer, a concatenation function (Concat), and a fifth convolutional layer. The output features of the fifth convolutional layer serve as the output features of the DRFD module; the input features of the third convolutional layer serve as the input features of the DRFD module; the output features of the third convolutional layer serve as the input features of the second BN layer; the output features of the second BN layer serve as the input features of both the fourth convolutional layer and the first maxpool layer; the output features of the fourth convolutional layer serve as the input features of the second BN+SiLU activation function layer; the output features of the first maxpool layer serve as the input features of the third BN layer; and the output features of the second BN+SiLU activation function layer and the third BN layer are concatenated using the Concat function and then used as the input features of the fifth convolutional layer.

4. The method for detecting intrusive foreign objects on urban transportation tracks based on the improved YOLOv11 according to claim 1, characterized in that, The C3k2-MSEPA module includes a C3k2 module, an MSPC module, and an EPA module cascaded together.

5. The method for detecting intrusive foreign objects on urban transportation tracks based on the improved YOLOv11 according to claim 4, characterized in that, The MSPC module includes: a fourth BN layer, a first PW convolutional layer, a sixth convolutional layer, a first GELU activation function, a depth dilation layer, a second PW convolutional layer, a second GELU activation function, and a third PW convolutional layer, all cascaded sequentially. The input features of the fourth BN layer and the output features of the third PW convolutional layer are summed to obtain the output features of the MSPC module. The depth dilation layer includes multiple parallel DWDConv layers of different scales. The output features of the first GELU activation function serve as the input features of the multiple DWDConv layers of different scales. The output features of the multiple DWDConv layers of different scales are concatenated and then used as the input features of the second PW convolutional layer.

6. The method for detecting intrusive foreign objects on urban transportation tracks based on the improved YOLOv11 according to claim 4, characterized in that, The EPA module includes: a fifth BN layer, an attention module, a fourth PW convolutional layer, a third GELU activation function, and a fifth PW convolutional layer, cascaded sequentially; the input features of the fifth BN layer and the output features of the fifth PW convolutional layer are summed to obtain the output features of the EPA module; the attention module includes a parallel scattering attention layer, a simple pixel attention layer, a channel attention layer, and a pixel attention layer; the output features of the fifth BN layer are used as the input features of the scattering attention layer, the simple pixel attention layer, the channel attention layer, and the pixel attention layer, respectively; the output features of the scattering attention layer, the simple pixel attention layer, the channel attention layer, and the pixel attention layer are concatenated to obtain the input features of the fourth PW convolutional layer.

7. The method for detecting intrusive foreign objects on urban transportation tracks based on the improved YOLOv11 according to claim 1, characterized in that, The GSConv module includes two branches. The first branch includes a first BottConv layer, a first GN layer, a first ReLU activation function, a second BottConv layer, a second GN layer, and a second ReLU activation function, all cascaded in sequence. The second branch includes a third BottConv layer, a third GN layer, a third ReLU activation function, a seventh convolutional layer, a sixth BN layer, and a first Sigmoid activation function, all cascaded in sequence. The output features of the first and second branches are multiplied together, then passed through a fourth BottConv layer, a fourth GN layer, and a fourth ReLU activation function, and finally added to the input features of the GSConv module to obtain the output features of the GSConv module.

8. The method for detecting intrusive foreign objects on urban transportation tracks based on the improved YOLOv11 according to claim 1, characterized in that, The intrusion detection model is trained using a boundary-sensitive Shape-NWD loss function, including: in, Represents the loss function; This represents a constant that is set. Represents the natural base; Indicates the scale factor; and These are the weighting coefficients for the horizontal and vertical directions, respectively; and Indicates the center coordinates of the actual annotation box; and Indicates the center coordinates of the prediction box; and Indicates the width and height of the prediction box; and This indicates the width and height of the actual annotation box; This represents a defined constant.

9. A foreign object intrusion detection system for urban transportation tracks based on an improved YOLOv11, characterized in that, The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the method for detecting foreign objects intruding into urban traffic tracks based on an improved YOLOv11 as described in any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The computer storage medium stores a remote monitoring program, which, when executed by a processor, implements a method for detecting intrusive foreign objects in urban traffic tracks based on an improved YOLOv11 as described in any one of claims 1 to 8.

Images