Point cloud obstacle recognition method for safety monitoring along high-speed rail
By using the attention module combined with N2P and ECA and the point-by-point convolutional segmentation layer in the point cloud obstacle recognition along the high-speed rail line, the problem of inability to efficiently and accurately identify obstacles in the prior art is solved, and higher recognition accuracy and real-time performance are achieved.
Patent Information
- Application Number
- CN202510041662.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-10
- Publication Date
- 2025-05-09
AI Technical Summary
The prior art cannot efficiently and accurately identify obstacles in massive point cloud data along the high-speed rail line, resulting in unsatisfactory identification results.
Using the attention module combined with N2P and ECA, the local to global semantic information of the point cloud is modeled layer by layer through multi-layer feature refinement and point-by-point convolution segmentation layer to achieve accurate identification of obstacles.
It significantly improves the accuracy and real-time nature of obstacle identification, and can accurately identify obstacles along the route in complex road scenarios, reducing the risk of accidents caused by human omissions.
Smart Images
Figure CN119964142A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent recognition technology, and in particular to a point cloud obstacle recognition method for safety monitoring along a high-speed railway line. Background Art
[0002] Traditional obstacle detection technology mainly relies on visual image analysis, but it has limitations in dealing with light changes, object occlusion and image quality limitations, all of which may affect the accuracy of obstacle recognition. In this regard, point cloud data has obvious advantages over traditional visual images: first, it can provide high-precision three-dimensional information of obstacles, which is crucial for accurately identifying and classifying obstacles along the high-speed rail line; second, point cloud data has strong anti-interference ability, is not affected by factors such as light, occlusion and image quality, and can operate stably in complex environments; finally, obstacle recognition based on point cloud data can achieve a high degree of real-time, automation and intelligence, and improve the safety and operation efficiency of high-speed rail lines.
[0003] Domestic researchers have made remarkable achievements in point cloud data processing and application. For example, the image and point cloud fusion technology proposed by Professor Hu Shimin and his team from Changchun University of Science and Technology effectively combines point cloud data with color images, provides new technical means for three-dimensional reconstruction, and improves the accuracy of obstacle recognition. The ground segmentation algorithm based on spatial grid division developed by Liu Sixuan and others from Shijiazhuang Railway University has significantly improved the efficiency and accuracy of ground point cloud processing by accurately segmenting ground point clouds, providing important preliminary data processing support for obstacle recognition. Professor Hou Guojiao used advanced point cloud processing libraries and software to extract key information of railway track structure. These studies not only demonstrate the professional depth of domestic scholars in high-speed rail obstacle recognition technology, but also provide a technical foundation for the intelligent operation and maintenance of railway systems.
[0004] However, domestic researchers have done little research on the real-time application of point cloud data processing, especially along high-speed rail lines, where obstacles cannot be identified efficiently and accurately, resulting in unsatisfactory results. Summary of the invention
[0005] The purpose of the present invention is to solve the above-mentioned problems and provide a point cloud obstacle identification method for safety monitoring along the high-speed railway line, which can efficiently and accurately identify various obstacles along the line from massive point cloud data, provide strong guarantee for the safe operation of the high-speed railway, and minimize the risk of accidents caused by human negligence.
[0006] To achieve the above object, the technical solution of the present invention is: a point cloud obstacle recognition method for safety monitoring along high-speed railways, comprising the following steps:
[0007] S1: Convert point cloud data into processed data and embed it;
[0008] S2: The embedded data is fed to the N2P attention module, so that the model focuses on the abnormal parts of the input data;
[0009] S3: The data processed by the N2P attention module is downsampled (ds1). The downsampled data (ds1) enters the attention module combined with N2P and ECA to extract features layer by layer.
[0010] S4: The data after downsampling (ds1) is transmitted to downsampling (ds2), and the data after downsampling (ds2) enters the attention module combining N2P and ECA again;
[0011] S5: The data after downsampling (ds2) is transmitted to upsampling (us1);
[0012] S6: The data after upsampling (us1) is transmitted to upsampling (us2);
[0013] S7: The upsampled (us2) data enters the point-by-point convolutional segmentation layer;
[0014] S8: Determine obstacles.
[0015] Preferably, in step S1, the point cloud data is projected onto a two-dimensional plane to obtain images of different viewing angles.
[0016] Preferably, the upsampling (us1) and upsampling (us2) in step S5 and step S6 are both passed through an attention module combining N2P and ECA.
[0017] Preferably, it also includes restoring the original image data, combining and comparing the data after the N2P attention module in step S2 with the upsampling (us2), and combining and comparing the data after the downsampling (ds1) in step S3 with the upsampling (us1).
[0018] Preferably, the data processing formula of the N2P attention module is: Suppose there is a point P, Q is the query vector of point P, K is the key vector of the neighborhood point P2 of point P, and V is the feature vector of the neighborhood point P2 after linear layer mapping.
[0019] Preferably, the ECA attention mechanism firstly focuses on the input feature map Perform a global average pooling operation, the feature shape is H*W*C, where H is the height of the feature map, W is the width of the feature map, and C is the number of channels; then the two-dimensional feature map (H*W) of each channel is compressed into a real number. After global average pooling, the output feature shape becomes a one-dimensional feature vector z=F pool(X); Then, after obtaining the feature vector after global average pooling, the ECA attention mechanism module generates channel attention weights through one-dimensional convolution (1D Conv), and the one-dimensional convolution outputs a weight vector s∈R°, which represents the importance of each channel s=F conv (z); Finally, the obtained weight vector s is converted to a value between 0 and 1 through the sigmoid activation function σ as the attention coefficient of each channel, and then the generated channel attention weight vector is multiplied by the feature map of the original input channel by channel. For each channel of the original feature map, its elements are multiplied by the corresponding attention weight to achieve the re-weighting of the feature map X
[0020] Preferably, the kernel size of the one-dimensional convolution is determined by adaptive calculation formula: Where γ and β are hyperparameters, and k is the kernel size of the one-dimensional convolution.
[0021] Compared with the existing technology, the point cloud obstacle recognition method for safety monitoring along the high-speed railway has the following beneficial effects: the combination of N2P and ECA significantly improves the accuracy of obstacle recognition, and shows excellent real-time performance in complex road scenes. It lays a solid foundation for the intelligent driving system to build a more accurate and reliable environmental perception, and is expected to further improve the safety and efficiency of autonomous driving; in addition, by restoring the original image data, that is, combining and comparing the data after the N2P attention module in step S2 with the upsampling (us2), and combining and comparing the data after the downsampling (ds1) in step S3 with the upsampling (us1), it can reduce data judgment errors and improve judgment and recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 The present invention is a flow chart of a point cloud obstacle identification method for safety monitoring along high-speed railway lines. DETAILED DESCRIPTION
[0023] The present invention will now be described in further detail with reference to the accompanying drawings. The accompanying drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and thus only show the components related to the present invention.
[0024] Please refer to Figure 1 , a point cloud obstacle recognition method for safety monitoring along high-speed railways, comprising the following steps:
[0025] S1: converting point cloud data into processed data and embedding it, specifically projecting the point cloud data onto a two-dimensional plane to obtain images from different perspectives;
[0026] S2: The embedded data is fed to the N2P attention module, so that the model focuses on the abnormal parts of the input data;
[0027] S3: The data processed by the N2P attention module is downsampled (ds1). The downsampled data (ds1) enters the attention module combined with N2P and ECA to refine features layer by layer and model the local to global semantic information of the point cloud;
[0028] S4: The data after downsampling (ds1) is transmitted to downsampling (ds2), and the data after downsampling (ds2) enters the attention module combining N2P and ECA again;
[0029] S5: The downsampled (ds2) data is sent to the upsampled (us1). This process processes the coarsest point cloud and refines the global semantic representation by adaptively adjusting the features of different regions. This hierarchical attention mechanism helps the model gradually establish point cloud feature representations from local to global, from low-level to high-level, and comprehensively utilizes semantic information at different scales to better cope with challenges such as sparsity, irregularity, and noise interference of point clouds.
[0030] S6: The data after upsampling (us1) is transmitted to upsampling (us2);
[0031] S7: The upsampled (us2) data enters the point-by-point convolutional segmentation layer;
[0032] S8: Determine obstacles.
[0033] In this embodiment, the upsampling (us1) and upsampling (us2) in step S5 and step S6 are both processed through an attention module combining N2P and ECA.
[0034] That is, the process starts with the embedding step, which converts the point cloud data into two-dimensional plane images with different perspectives. The embedded data enters the N2P attention module, which enables the model to focus on the abnormal parts of the input data, thereby improving processing efficiency and accuracy.
[0035] The data after N2P attention processing is downsampled for the first time (ds1). It is used to reduce the amount of data while retaining important features. The downsampled data enters the attention module that combines N2P and ECA. It refines features layer by layer and models the semantic information of the point cloud from local to global, which helps to extract more meaningful features from the point cloud data.
[0036] The data is downsampled again (ds2) to further reduce the amount of data. The downsampled data enters the attention module that combines N2P and ECA again. Further refining the global semantic representation helps to refine features on coarser-grained data. After the above processing, the data needs to be upsampled twice to restore the original resolution of the data. The upsampled data enters the attention module that combines N2P and ECA again to further process the data. The data that has been processed multiple times finally enters the point-by-point convolution segmentation layer. For each point in the point cloud data, the point-by-point convolution segmentation layer determines which category the point belongs to (obstacle or background) based on the features of its surrounding points and the contextual information of the entire point cloud.
[0037] After going through the above steps, the point cloud data has undergone multiple feature extractions. The point-wise convolutional segmentation layer will receive these processed and refined features. This processing plays a role in feature fusion and refinement. Through the point-wise convolutional segmentation layer, the present invention can comprehensively consider the feature information of the point cloud data at different scales and global semantics, so as to make more accurate segmentation judgments. The point-wise convolutional segmentation layer will eventually output a segmentation result, marking each point in the point cloud data map as belonging to a specific category, and this segmentation result can be used for subsequent applications. It can comprehensively utilize the features extracted by the previous multiple processing steps to accurately segment irregular point cloud data, thereby realizing the recognition of targets such as obstacles.
[0038] In addition, the recognition method also includes restoring the original image data, that is, the data after the N2P attention module in step S2 is combined and compared with the upsampling (us2), and the data after the downsampling (ds1) in step S3 is combined and compared with the upsampling (us1). If the upsampling is performed directly without combining with the original data, errors may occur. Therefore, by restoring the original image data, the data judgment error can be reduced and the judgment and recognition accuracy can be improved.
[0039] In the present invention, the data processing formula of the N2P attention module is: Suppose there is a point P, Q is the query vector of point P, K is the key vector of the neighborhood point P2 of point P, and V is the feature vector of the neighborhood point P2 after linear layer mapping.
[0040] When the ECA attention mechanism is running, first, the input feature map (The shape is H*W*C, where H is the height of the feature map, W is the width of the feature map, and C is the number of channels) to perform a global average pooling operation. It compresses the two-dimensional feature map (H*W) of each channel into a real number. After global average pooling, the output feature shape becomes a one-dimensional feature vector z=Fpool (X).
[0041] After obtaining the feature vector after global average pooling, the ECA module generates channel attention weights through a simple one-dimensional convolution (1DConv). The kernel size of this one-dimensional convolution is adaptively determined, and the calculation formula is Where γ and β are hyperparameters, and k is the kernel size of the one-dimensional convolution. Through this adaptive method, the convolution kernel size can be flexibly adjusted according to the number of channels, avoiding the inadaptability of fixed parameters in different channel number scenarios. The output of this one-dimensional convolution is a weight vector s∈R°, which represents the importance of each channel: s=F conv (z).
[0043] Finally, the obtained weight vector s is converted to a value between 0 and 1 through a sigmoid activation function σ as the attention coefficient of each channel. Then the generated channel attention weight vector is multiplied by the feature map of the original input channel by channel. For each channel of the original feature map, its elements are multiplied by the corresponding attention weight. To achieve the reweighting of the feature map X:
[0044] In this way, channels with higher attention weights are enhanced, while channels with lower attention weights are suppressed.
[0045] Combining the above process, the formula of the ECA attention mechanism can be written as:
[0046]
[0047] The input feature map is The output feature map is i=1,2,…,H; j=1,2,…,W; c=1,2,…,C.
[0048] In the solution of the present invention, N2P is responsible for capturing the relationship between points and neighborhoods and enhancing the expressiveness of local features, while ECA enhances feature discrimination and robustness by adaptively adjusting the weights of different channels. Unlike the attention mechanism used alone, the combination of N2P and ECA introduces stronger nonlinearity and achieves self-calibration of channel attention and local spatial relationships with extremely low computational overhead, thereby extracting more discriminative local features.
[0049] After downsampling (ds1), by recursively applying the combination of N2P and ECA at different scales of the point cloud, features can be refined layer by layer, and the semantic information of the point cloud can be modeled from local to global; after downsampling (ds2), the attention module combining N2P and ECA is used again. At this time, the point cloud has been downsampled to a lower resolution, and each point contains more extensive semantic information. The combination of N2P and ECA adaptively adjusts the feature responses of different regions, which not only retains the details of local features, but also effectively focuses on the key parts of the object, suppresses the influence of background noise, and obtains a more refined global feature representation.
[0050] Compared with the separate N2P or ECA mechanism, the combination of N2P and ECA has significant advantages in computational efficiency and parameter quantity, and is particularly suitable for scenarios with high real-time requirements, such as obstacle recognition in autonomous driving. In addition, the combination of N2P and ECA also enables the network to adaptively focus on different areas of the point cloud and model its internal and external semantic associations, greatly improving the accuracy of obstacle recognition; the fusion of N2P and ECA attention mechanisms can efficiently and accurately identify various obstacles along the line from massive point cloud data, such as trees, buildings, mechanical equipment, etc., providing strong guarantees for the safe operation of high-speed rail. This can not only reduce the cost of manual inspections, but also minimize the risk of accidents caused by human negligence, opening up new ways for intelligent operation and maintenance and safety management of high-speed rail.
[0051] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.
Claims
1. A point cloud obstacle recognition method for safety monitoring along high-speed railways, characterized in that: The steps include: S1: Convert point cloud data into processed data and embed it; S2: The embedded data is fed to the N2P attention module, so that the model focuses on the abnormal parts of the input data; S3: The data processed by the N2P attention module is downsampled (ds1). The downsampled data (ds1) enters the attention module combined with N2P and ECA to extract features layer by layer. S4: The data after downsampling (ds1) is transmitted to downsampling (ds2), and the data after downsampling (ds2) enters the attention module combining N2P and ECA again; S5: The data after downsampling (ds2) is transmitted to upsampling (us1); S6: The data after upsampling (us1) is transmitted to upsampling (us2); S7: The upsampled (us2) data enters the point-by-point convolutional segmentation layer; S8: Determine obstacles.
2. The point cloud obstacle identification method for high-speed railway line safety monitoring according to claim 1 is characterized in that: In the step S1, the point cloud data is projected onto a two-dimensional plane to obtain images of different viewing angles.
3. The point cloud obstacle identification method for high-speed railway line safety monitoring according to claim 1 is characterized in that: In the steps S5 and S6, the upsampling (us1) and upsampling (us2) are both performed through the attention module combining N2P and ECA.
4. The point cloud obstacle identification method for high-speed railway line safety monitoring according to any one of claims 1 to 3, characterized in that: It also includes restoring the original image data, combining and comparing the data after the N2P attention module in step S2 with the upsampling (us2), and combining and comparing the data after the downsampling (ds1) in step S3 with the upsampling (us1).
5. The point cloud obstacle identification method for safety monitoring along high-speed railways according to any one of claims 1 to 3, characterized in that: The data processing formula of the N2P attention module is: Suppose there is a point P, Q is the query vector of point P, K is the key vector of the neighborhood point P2 of point P, and V is the feature vector of the neighborhood point P2 after linear layer mapping.
6. The point cloud obstacle identification method for safety monitoring along high-speed railways according to any one of claims 1 to 3, characterized in that: The ECA attention mechanism firstly focuses on the input feature map Perform a global average pooling operation, the feature shape is H*W*C, where H is the height of the feature map, W is the width of the feature map, and C is the number of channels; then the two-dimensional feature map (H*W) of each channel is compressed into a real number. After global average pooling, the output feature shape becomes a one-dimensional feature vector z=F pool (X); Then, after obtaining the feature vector after global average pooling, the ECA attention mechanism module generates channel attention weights through one-dimensional convolution (1DConv), and the one-dimensional convolution outputs a weight vector s∈R°, which represents the importance of each channel s=F conv (z); Finally, the obtained weight vector s is converted to a value between 0 and 1 through the sigmoid activation function σ as the attention coefficient of each channel, and then the generated channel attention weight vector is multiplied by the feature map of the original input channel by channel. For each channel of the original feature map, its elements are multiplied by the corresponding attention weight to achieve the re-weighting of the feature map X 7. The point cloud obstacle identification method for high-speed railway line safety monitoring according to claim 6 is characterized in that: The kernel size of the one-dimensional convolution is determined by adaptation, and the calculation formula is: Where γ and β are hyperparameters, and k is the kernel size of the one-dimensional convolution.