Train Forward Clearance Target Detection Method and System Based on Point Cloud Feature Aggregation Enhancement

Abstract

The present invention provides a method and system for detecting forward clearance targets of a train based on point cloud feature aggregation enhancement, belonging to the technical field of forward clearance target detection of trains. Vehicle-mounted point cloud data is acquired to construct a scene point cloud data set; the sparse point cloud is divided into regular voxel grids, and voxel feature encoding converts the point cloud information into a compact feature representation, and 3D sparse convolution is used to downsample layer by layer to extract multi-scale voxel features; set sampling is used to sample, group, and extract features from the extracted multi-level voxel features, and the multi-scale voxel features are aggregated to the voxel center points; a fully sparse and highly compressed method is used to map the extracted hierarchical multi-scale voxel features into a BEV feature map, and through 2D sparse convolution processing and skip connections, BEV feature maps with different resolutions are fused; the centroid offset loss of the 3D regression box is introduced to optimize the multi-task loss, and the detection head is guided to realize the classification of forward clearance targets of the train and the regression of three-dimensional bounding boxes.

Description

Technical Field

[0001] The present invention relates to the technical field of train forward clearance target detection, and particularly to a train forward clearance target detection method and system based on point cloud feature aggregation enhancement. Background Art

[0002] In modern rail transit systems, train safety is the primary task to ensure operation efficiency and passenger safety. With the rapid development of urban rail transit, potential obstacles such as pedestrians, vehicles, and animals that trains face during operation may cause train collisions, seriously affecting operation safety. Therefore, it is particularly important to carry out the detection and recognition of train forward clearance targets.

[0003] Traditional forward clearance detection methods mainly rely on technologies such as manual monitoring and video surveillance. These methods have insufficient accuracy and real-time performance and are easily affected by environmental factors such as light and weather. With the gradual maturity of lidar (LiDAR) and point cloud technology, these problems have been effectively solved. Lidar can obtain three-dimensional information of the surrounding environment with high precision and generate rich point cloud data. These point cloud data not only contain the position and shape information of objects in the scene but also provide rich reflection intensity and depth information.

[0004] In order to make full use of point cloud data for target detection, in recent years, point cloud processing technologies based on deep learning have developed rapidly. Point cloud feature extraction and aggregation technologies have become the key to improving detection accuracy. Among them, the voxel feature encoding method converts point cloud data into a regular grid, which can effectively reduce the computational complexity while maintaining spatial structure information. Through a 3D sparse convolutional network, efficient feature extraction and downsampling can be achieved to obtain multi-scale feature representations. However, traditional feature aggregation methods often have the problem of information loss when processing multi-level features and are difficult to make full use of feature information at different scales. Therefore, improving three-dimensional target detection ability is of great significance for train forward clearance obstacle detection. Summary of the Invention

[0005] The purpose of the present invention is to provide a train forward clearance target detection method and system based on point cloud feature aggregation enhancement to solve at least one of the technical problems in the above background art.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] In the first aspect, the present invention provides a train forward clearance target detection method based on point cloud feature aggregation enhancement, including:

[0008] Obtain on-vehicle point cloud data and construct a scene point cloud data set;

[0009] Divide the sparse point cloud into regular voxel grids. Voxel feature encoding converts the point cloud information into a compact feature representation, and uses 3D sparse convolution to downsample layer by layer to extract multi-scale voxel features;

[0010] Use set sampling to sample, group, and extract features from the extracted multi-level voxel features, and aggregate the multi-scale voxel features to the voxel center points;

[0011] Adopt a fully sparse and highly compressed method to map the extracted hierarchical multi-scale voxel features to a BEV feature map, and through 2D sparse convolution processing and skip connections, fuse BEV feature maps of different resolutions;

[0012] Introduce the centroid offset loss of the 3D regression box to optimize the multi-task loss and guide the detection head to achieve the classification of the train forward clearance target and the regression of the three-dimensional bounding box.

[0013] As a further limitation of the first aspect of the present invention, the division of the sparse point cloud into regular voxel grids, voxel feature encoding converts the point cloud information into a compact feature representation, and uses 3D sparse convolution to downsample layer by layer to extract multi-scale voxel features, including: Voxel gridification: Divide the input N spatial point clouds P=(p1, p2,..., p n ) region into voxel units with a spatial resolution of L×W×H, where the point cloud feature representation is (x, y, z, r), that is, the 3D coordinates and reflection intensity of the point cloud; Voxel feature encoding: Average the point cloud features within the voxel unit to obtain the initially encoded voxel features, represented as a four-dimensional tensor; Use 3D submanifold sparse convolution to perform layer-by-layer downsampling processing on the voxel features, and encode the features of non-empty voxels into multi-scale three-dimensional sparse feature quantities with downsampling ratios of 1x, 2x, 4x, and 8x respectively; Among them, the voxel spatial coordinates of the k-th layer of downsampling are represented as The set representation of its voxel feature quantity is where N k is the number of non-empty voxels in the k-th layer, and the superscript l k (k = 1, 2, 3, 4) represents the downsampling layers with ratios of 1x, 2x, 4x, and 8x.

[0014] As a further limitation of the first aspect of the present invention, the use of set sampling to sample, group, and extract features from the multi-level voxel features extracted in step S2, and aggregate the multi-scale voxel features to the voxel center points, including: Point query feature grouping: Use the 8x downsampling layer after conventional 3D sparse convolution feature encoding as the aggregation layer, and the center of the voxel unit in this layer is used as the query point Q for set sampling, and the spatial coordinates of the query point are calculated according to the voxel index and the actual size of the voxel; Point query clustering: With each query point as the center, according to the set radius r of the k-th layer kFind all non-empty voxels in the layer area, obtain the local grouped non-empty voxel set, and calculate the voxel feature vector set of the k-th layer; point query feature aggregation: use set sampling to extract the features of the voxels in the k-th layer.

[0015] As a further limitation of the first aspect of the present invention, two r with different scales are used in the k layer. k Increase the local variable receptive field and aggregate the voxel features of the initialized voxel layer and all downsampling layers into the query point q i Multi-scale voxel features.

[0016] As a further limitation of the first aspect of the present invention, the fully sparse high compression method is used to map the hierarchical multi-scale voxel features extracted in step S2 into a BEV feature map, and the BEV feature maps of different resolutions are fused through 2D sparse convolution processing and jump connection, including: spatial multi-rate feature mapping: using a fully sparse high compression operation to map the multi-scale non-empty voxel features to the XY plane grid, and accumulating the features at the same Z position to finally generate BEV sparse feature maps with 1x, 2x, 4x and 8x downsampling rates; BEV multi-level feature extraction: extracting the obtained 1 x-scale BEV sparse feature mapping, perform three-level 2D sparse convolution downsampling operations, and extract BEV sparse convolution features with 2x, 4x and 8x downsampling rates layer by layer; Hierarchical feature fusion mechanism: During the downsampling process, a hierarchical feature fusion mechanism is introduced to fuse multi-rate BEV features layer by layer in the order of 1x→2x→4x→8x, extract high-resolution global semantic information, capture more detailed contextual information, and obtain 8x aggregated features with rich contextual information; the obtained 8x aggregated features are converted into 2D dense features and used as the input of the detection head for target classification and positioning tasks.

[0017] As a further limitation of the first aspect of the present invention, the centroid offset loss of the 3D regression box is introduced to optimize the multi-task loss, guiding the detection head to achieve train forward clearance target classification and three-dimensional bounding box regression, including: a 2D feature extraction network is responsible for further processing 2D dense features and generating the final detection output, including two sub-networks: a top-down sub-network performs a 1x1 convolution and two downsamplings on the 2D dense features, and the other sub-network performs upsampling in the opposite direction to restore the spatial resolution of the input, and finally aggregates the features of all layers and the original input features into the detection head to achieve train forward clearance target classification and three-dimensional bounding box regression.

[0018] In a second aspect, the present invention provides a train forward clearance target detection system based on point cloud feature aggregation enhancement, comprising:

[0019] The acquisition module is used to obtain vehicle-mounted point cloud data and construct a scene point cloud dataset;

[0020] A downsampling extraction module is used to divide the sparse point cloud into regular voxel grids. The voxel feature encoding converts the point cloud information into a compact feature representation, and uses 3D sparse convolution to perform downsampling layer by layer to extract multi-scale voxel features;

[0021] A multi-scale voxel feature aggregation module based on point query is used to sample, group, and extract features from the extracted multi-level voxel features using set sampling, and aggregate the multi-scale voxel features to the voxel center points;

[0022] A BEV multi-level auxiliary branch feature extraction module is used to map the extracted hierarchical multi-scale voxel features into a BEV feature map using a fully sparse height compression method, and fuse the BEV feature maps of different resolutions through 2D sparse convolution processing and skip connections;

[0023] A detection module is used to introduce the centroid offset loss of the 3D regression box to optimize the multi-task loss, and guide the detection head to achieve the classification of the train forward clearance target and the regression of the three-dimensional bounding box.

[0024] In a third aspect, the present invention provides a non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the train forward clearance target detection method based on point cloud feature aggregation enhancement as described in the first aspect is implemented.

[0025] In a fourth aspect, the present invention provides a computer device, including a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the train forward clearance target detection method based on point cloud feature aggregation enhancement as described in the first aspect.

[0026] In a fifth aspect, the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory, so that the electronic device executes the instruction to implement the train forward clearance target detection method based on point cloud feature aggregation enhancement as described in the first aspect.

[0027] Advantages of the present invention: The lidar installed at the front end of the train is used to obtain on-vehicle point cloud data and construct a scene point cloud data set; the sparse point cloud is divided into regular voxel grids and transformed into a compact feature representation through voxel feature encoding, and 3D sparse convolution is used for layer-by-layer downsampling to extract multi-scale voxel features; based on the multi-scale voxel feature aggregation module with point query, set sampling is used to sample, group, and extract features from the extracted multi-level voxel features, and the multi-scale features are aggregated to the voxel center points; the BEV multi-level auxiliary branch feature extraction module uses a fully sparse and highly compressed method to map the multi-scale voxel features into a BEV feature map, and fuses feature maps with different resolutions through 2D sparse convolution processing and skip connections; the centroid offset loss of the 3D regression box is introduced to optimize the multi-task loss to guide the detection head to achieve the classification and three-dimensional bounding box regression of the train forward clearance target; the accuracy and efficiency of the train forward clearance target detection are significantly improved, providing strong technical support for the safe operation of the train and showing good application prospects.

[0028] The advantages of the additional aspects of the present invention will be more clearly given in the following description part, or can be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0030] Figure 1 It is a flowchart of the method for detecting the train forward clearance target based on point cloud feature aggregation enhancement described in the embodiment of the present invention.

[0031] Figure 2 It is a schematic diagram of the experimental platform described in the embodiment of the present invention.

[0032] Figure 3 It is a framework diagram of the multi-scale voxel feature aggregation module with point query described in the embodiment of the present invention.

[0033] Figure 4 It is a framework diagram of the BEV multi-level auxiliary branch feature extraction module described in the embodiment of the present invention.

[0034] Figure 5 It is a detection effect diagram described in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

[0036] Those skilled in the art of this technology can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the art to which the present invention pertains.

[0037] It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with their meaning in the context of the prior art, and will not be interpreted with an idealized or overly formal meaning unless defined as herein.

[0038] Those skilled in the art of this technology can understand that, unless specifically stated, the singular forms "a", "an", "the", and "said" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of the present invention means the presence of the described features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or their groups.

[0039] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Without conflict, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0040] To facilitate the understanding of the present invention, the present invention will be further explained below with reference to the accompanying drawings by means of specific embodiments, and the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0041] Those skilled in the art should understand that the drawings are only schematic diagrams of the embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0042] The present invention relates to a method and system for detecting forward clearance targets of trains based on enhanced point cloud feature aggregation. The method first uses a lidar installed at the front end of the train to obtain on-vehicle point cloud data and constructs a scene point cloud data set. Then, the sparse point cloud is divided into regular voxel grids, and through voxel feature encoding, it is transformed into a compact feature representation. 3D sparse convolution is used for layer-by-layer downsampling to extract multi-scale voxel features. In addition, a multi-scale voxel feature aggregation module based on point query is designed, which uses set sampling to sample, group, and extract features from the extracted multi-level voxel features, and aggregates the multi-scale features to the voxel center points. The present invention also proposes a BEV multi-level auxiliary branch feature extraction module, which uses a completely sparse and highly compressed method to map the multi-scale voxel features into a BEV feature map, and fuses feature maps of different resolutions through 2D sparse convolution processing and skip connections. Finally, the centroid offset loss of the 3D regression box is introduced to optimize the multi-task loss, so as to guide the detection head to achieve the classification of forward clearance targets of trains and the regression of three-dimensional bounding boxes. This method significantly improves the accuracy and efficiency of forward clearance target detection of trains, provides strong technical support for the safe operation of trains, and shows good application prospects.

[0043] Embodiment 1

[0044] In this Embodiment 1, first, a system for detecting forward clearance targets of trains based on enhanced point cloud feature aggregation is provided, including: an acquisition module for acquiring on-vehicle point cloud data and constructing a scene point cloud data set; a downsampling and extraction module for dividing the sparse point cloud into regular voxel grids, and voxel feature encoding to transform the point cloud information into a compact feature representation, and using 3D sparse convolution for layer-by-layer downsampling to extract multi-scale voxel features; a multi-scale voxel feature aggregation module based on point query for using set sampling to sample, group, and extract features from the extracted multi-level voxel features, and aggregating the multi-scale voxel features to the voxel center points; a BEV multi-level auxiliary branch feature extraction module for using a completely sparse and highly compressed method to map the extracted hierarchical multi-scale voxel features into a BEV feature map, and fusing BEV feature maps of different resolutions through 2D sparse convolution processing and skip connections; a detection module for introducing the centroid offset loss of the 3D regression box to optimize the multi-task loss and guiding the detection head to achieve the classification of forward clearance targets of trains and the regression of three-dimensional bounding boxes.

[0045] In this embodiment, using the above system, a method for detecting forward clearance targets of trains based on enhanced point cloud feature aggregation is implemented, including the following steps:

[0046] Step S1: Obtain on-vehicle point cloud data through a lidar installed at the front end of the train and construct a scene point cloud data set;

[0047] Step S2: Divide the sparse point cloud into regular voxel grids. Voxel feature encoding converts the point cloud information into a compact feature representation, and 3D sparse convolution is used for layer-by-layer downsampling to extract multi-scale voxel features;

[0048] Step S3: Design a multi-scale voxel feature aggregation module based on point queries. Use set sampling to sample, group, and extract features from the multi-level voxel features extracted in Step S2, and aggregate the multi-scale voxel features to the voxel center points;

[0049] Step S4: Design a BEV multi-level auxiliary branch feature extraction module. Use a fully sparse height compression method to map the hierarchical multi-scale voxel features extracted in Step S2 into a BEV feature map, and fuse BEV feature maps of different resolutions through 2D sparse convolution processing and skip connections;

[0050] Step S5: Introduce the centroid offset loss of the 3D regression box to optimize the multi-task loss, and guide the detection head to achieve the classification of the train forward clearance target and the regression of the three-dimensional bounding box.

[0051] The division of the sparse point cloud into regular voxel grids as described in Step S2, where voxel feature encoding converts the point cloud information into a compact feature representation, and 3D sparse convolution is used for layer-by-layer downsampling to extract multi-scale voxel features, includes the following steps:

[0052] Step S21: Voxel rasterization: Divide the input N spatial point clouds P=(p1, p2,..., p n ) region into voxel units with a spatial resolution of L×W×H, where the point cloud feature representation is (x, y, z, r), that is, the 3D coordinates and reflection intensity of the point cloud.

[0053] Step S22: Voxel feature encoding: Average the point cloud features within the voxel unit to obtain the initially encoded voxel features, represented as a four-dimensional tensor: The corresponding spatial coordinates are represented as where N0 is the number of non-empty voxels, and the superscript l0 represents the initial voxel layer.

[0054] Step S23: Use 3D submanifold sparse convolution to perform layer-by-layer downsampling on the voxel features in Step S22, and encode the features of non-empty voxels into multi-scale three-dimensional sparse feature quantities with downsampling ratios of 1x, 2x, 4x, and 8x respectively. Among them, the voxel spatial coordinates of the k-th layer of downsampling are represented as The set of its voxel feature quantities is represented as where N k is the number of non-empty voxels in the k-th layer, and the superscript l k (k = 1, 2, 3, 4) represents the downsampling layers with ratios of 1x, 2x, 4x, and 8x.

[0055] The multi-scale voxel feature aggregation module based on point query described in step S3 samples, groups, and extracts features from the multi-level voxel features extracted in step S2 using set sampling, and aggregates the multi-scale voxel features to the voxel center point, including the following steps:

[0056] Step S31: Point query feature grouping: Use the 8x downsampling layer in step S23 after conventional 3D sparse convolution feature encoding as the aggregation layer, and the center of the voxel unit in this layer is used as the query point Q for set sampling (SA). Calculate the spatial coordinates of the query point Q = {q1,...,q m} according to the voxel index and the actual voxel size, where m is The set of voxel feature vectors in this layer is represented as The superscript q represents the query layer.

[0057] Step S32: Point query clustering: With each query point q i (i = 1,...,m) as the center, find all non-empty voxels within its neighborhood in this layer according to the set radius r of the k-th layer k to obtain the set of locally grouped non-empty voxels. The set of voxel feature vectors in the k-th layer is represented as The calculation formula is as follows:

[0058]

[0059] In the formula, represents collecting the 16 nearest neighbor voxels within the neighborhood r k and connecting their local relative coordinates with the feature

[0060] Step S33: Point query feature aggregation: Use set sampling (Set Abstraction, SA) (through MLP multi-layer perceptron and max pooling operations) to extract the features of the voxels in the k-th layer:

[0061]

[0062] In the formula, A(.) represents a three-layer MLP network to encode the features, and aggregates the multi-scale features into the query point through the max pooling operation MAX(.) along the channels. In addition, we use two different scales of r in the k-th layer k to increase the local variable receptive field. Aggregate the voxel features of the initial voxel layer and all downsampling layers into the multi-scale voxel feature f i of the query point q i (qv) , and the aggregation formula is as follows:

[0063]

[0064] where \(i = 1,\cdots,m\) denotes the feature concatenation operation; the generated feature \(f\) i (qv) contains both the multi-scale downsampling features of the 3D CNN and the voxel initialization features (i.e., 3D coordinates and reflection intensity).

[0065] The designed BEV multi-level auxiliary branch feature extraction module described in step S4 uses a fully sparse and highly compressed method to map the hierarchical multi-scale voxel features extracted in step S2 into a BEV feature map, and through 2D sparse convolution processing and skip connections, fuses BEV feature maps with different resolutions, including the following steps:

[0066] Step S41: Spatial multi-magnification feature mapping: Using a fully sparse and highly compressed (Sparse Height Compression, SHC) operation, map the multi-scale non-empty voxel features in step S23 onto the X-Y plane grid, and accumulate the features at the same Z position, finally generating BEV sparse feature maps with downsampling magnifications of 1x, 2x, 4x, and 8x.

[0067] Step S42: BEV multi-level feature extraction: Perform a three-level 2D sparse convolution downsampling operation on the 1x magnification BEV sparse feature map obtained in step S41 to extract BEV sparse convolution features with downsampling magnifications of 2x, 4x, and 8x layer by layer.

[0068] Step S43: Hierarchical feature fusion mechanism: During the downsampling process, introduce a hierarchical feature fusion mechanism to fuse multi-magnification BEV features step by step in the order of 1x→2x→4x→8x Extract high-resolution global semantic information, capture more detailed context information, and obtain 8x aggregated features with rich context information.

[0069] Step S44: Convert the 8x aggregated features obtained in step S43 into 2D dense features and use them as the input of the detection head for object classification and localization tasks.

[0070] Introduce the centroid offset loss of the 3D regression box in step S5 to optimize the multi-task loss and guide the detection head to achieve the classification of the forward clearance target of the train and the regression of the three-dimensional bounding box.

[0071] Step S51: The 2D feature extraction network is responsible for further processing the 2D dense features and generating the final detection output, including two sub-networks: a top-down sub-network performs a 1x1 convolution and two downsamplings (16x, 32x) on the 2D dense features, and another sub-network performs upsampling in the reverse direction to restore the spatial resolution of the input (8x). Finally, the features of all layers and the original input features are aggregated and input into the detection head to achieve the classification of the train forward clearance target and the regression of the 3D bounding box.

[0072] Step S52: During the model convergence process, a multi-task loss function is adopted, including classification loss, localization loss, orientation loss, and the centroid offset loss L of the 3D regression box centroid which is calculated as follows:

[0073]

[0074] where C box and C gt are the average coordinates of the spatial points within the 3D predicted bounding box and the ground truth box respectively, γ and τ are the local densities of each point estimated by the kernel density estimation (KDE) method, and we introduce the L2 norm ‖.‖2 to calculate the offset between the centroid of the 3D predicted regression box and the centroid of the ground truth box.

[0075] The overall loss function is defined as:

[0076]

[0077] where L cls is the classification loss, L loc is the localization loss, L dir is the orientation loss, and N pos represents the number of positive example anchors, where λ1 = 1, λ2 = 2, λ3 = 0.2, and λ4 = 1 are the weight hyperparameters for balancing the losses of each part.

[0078] Example 2

[0079] In this embodiment, a train forward clearance target detection method and system based on point cloud feature aggregation enhancement are proposed. Through a series of steps, the method first uses a lidar installed at the front end of the train to obtain on-vehicle point cloud data, thereby constructing a scene point cloud data set. Then, the sparse point cloud is divided into regular voxel grids, and voxel feature encoding is used to convert the point cloud information into a compact feature representation, and 3D sparse convolution is used for layer-by-layer downsampling to extract multi-scale voxel features. In addition, a multi-scale voxel feature aggregation module based on point query is designed, and the extracted multi-level voxel features are sampled, grouped, and feature extracted through set sampling, and the multi-scale voxel features are aggregated to the voxel center points. In this embodiment, a BEV multi-level auxiliary branch feature extraction module is also proposed, which uses a fully sparse and highly compressed method to map the hierarchical multi-scale voxel features to a BEV feature map, and fuses BEV feature maps with different resolutions through 2D sparse convolution processing and skip connections. Finally, the centroid offset loss of the 3D regression box is introduced to optimize the multi-task loss, thereby guiding the detection head to achieve the classification of the train forward clearance target and the regression of the three-dimensional bounding box. The detection method based on point cloud feature aggregation enhancement in this embodiment makes significant contributions to improving the accuracy and efficiency of train forward clearance target detection. At the same time, the method provides strong technical support for the safe operation of trains and provides an effective solution for realizing efficient train forward clearance detection.

[0080] As Figure 1 shown, the train forward clearance target detection method based on point cloud feature aggregation enhancement described in this embodiment includes the following steps:

[0081] Step S1: Obtain on-vehicle point cloud data through a lidar installed at the front end of the train, and construct a scene point cloud data set;

[0082] Step S2: Divide the sparse point cloud into regular voxel grids, voxel feature encoding converts the point cloud information into a compact feature representation, and 3D sparse convolution is used for layer-by-layer downsampling to extract multi-scale voxel features;

[0083] Step S3: Design a multi-scale voxel feature aggregation module based on point query, use set sampling to sample, group, and feature extract the multi-level voxel features extracted in Step S2, and aggregate the multi-scale voxel features to the voxel center points;

[0084] Step S4: Design a BEV multi-level auxiliary branch feature extraction module, use a fully sparse and highly compressed method to map the hierarchical multi-scale voxel features extracted in Step S2 to a BEV feature map, and fuse BEV feature maps with different resolutions through 2D sparse convolution processing and skip connections;

[0085] Step S5: The centroid offset loss of the 3D regression box is introduced to optimize the multi-task loss to guide the detection head to achieve train-forward clear object classification and 3D bounding box regression.

[0086] In this example, the specific technical solutions are as follows:

[0087] (1) Multi-scale voxel feature aggregation based on point query

[0088] By designing a multi-level point query mechanism, multi-scale voxel features are effectively aggregated to the voxel center point. First, the sparse point cloud is divided into a regular voxel grid, and the voxel feature encoding converts the point cloud information into a compact feature representation. The multi-scale voxel features are extracted by downsampling layer by layer using 3D sparse convolution. Subsequently, the ensemble sampling (SetAbstraction, SA) method is used to select voxel features within a specific radius at each scale and aggregate their features to the center point of the voxel. This process allows the model to capture the feature distribution at different spatial scales and form a high-quality spatial feature representation through aggregation operations to ensure that the information obtained from multiple levels is not lost. The aggregated features are processed by a multi-layer perceptron (MLP) to achieve efficient feature extraction and enhancement.

[0089] (2) BEV multi-level auxiliary branch feature extraction module

[0090] First, a 3D sparse convolutional network extracts multi-level voxel features. Then, sparse height compression techniques are used to convert these voxel features into corresponding BEV feature maps. 2D sparse convolutions are used to further extract features, and skip connections are introduced to enable the effective fusion of low-level and high-level features in the feature map, thereby improving feature expressiveness and information flow.

[0091] (3) Centroid offset loss of 3D regression box

[0092] The 3D regression box centroid offset loss constrains the model's learning process by comparing the centroid of the predicted 3D bounding box with the centroid of the ground-truth box. A multi-task loss combining classification, localization, and orientation losses guides the detection head to more accurately classify objects and regress 3D bounding boxes, improving the model's 3D bounding box localization accuracy and enhancing its adaptability to objects of varying scales and shapes.

[0093] In this example, the specific implementation is as follows:

[0094] Obtain vehicle-mounted point cloud data and construct scene point cloud dataset: Obtain vehicle-mounted point cloud data through the lidar installed at the front of the train, such as Figure 2As shown in the figure, in order to better analyze and process these data, a train forward scene point cloud dataset is constructed. This dataset simulates scenarios where people and obstacles (such as cardboard boxes, foam stones) invade the track, aiming to support subsequent object detection and environmental perception tasks.

[0095] (1) Voxel Feature Encoding and Multi-scale Feature Extraction of Sparse Point Clouds

[0096] In this example, the sparse point cloud is divided into regular voxel grids. Voxel feature encoding converts the point cloud information into a compact feature representation, and 3D sparse convolution is used for downsampling layer by layer to extract multi-scale voxel features.

[0097] A) Voxel Grid Generation

[0098] The input N spatial point clouds P = (p1, p2,..., p n ) are divided into voxel units with a spatial resolution of L×W×H. Among them, the point cloud feature representation (x, y, z, r) represents the 3D coordinates and reflection intensity of the point cloud.

[0099] B) Voxel Feature Encoding

[0100] The point cloud features within the voxel unit are averaged to obtain the initially encoded voxel features, represented as a four-dimensional tensor: The corresponding spatial coordinates are represented as where N0 is the number of non-empty voxels, and the superscript l0 represents the initial voxel layer.

[0101] C) Multi-scale Feature Extraction

[0102] 3D submanifold sparse convolution is used to perform downsampling processing on the voxel features in step S22 layer by layer. The features of non-empty voxels are respectively encoded into multi-scale three-dimensional sparse feature quantities with downsampling ratios of 1x, 2x, 4x, and 8x. Among them, the spatial coordinates of the voxels after the k-th layer of downsampling are represented as The set of its voxel feature quantities is represented as where N k is the number of non-empty voxels in the k-th layer, and the superscript l k (k = 1, 2, 3, 4) represents the downsampling layers with ratios of 1x, 2x, 4x, and 8x.

[0103] (2) Design of a Multi-scale Voxel Feature Aggregation Module Based on Point Query

[0104] In this example, a multi-scale voxel feature aggregation module based on point query is designed. Set sampling is used to sample, group, and extract features from the multi-level voxel features extracted in step S2, and the multi-scale voxel features are aggregated to the voxel center points. The algorithm flow is as Figure 3 shown.

[0105] A) Point query feature grouping

[0106] Take the 8x downsampling layer in step S23 after conventional 3D sparse convolution feature encoding as the aggregation layer. The center of the voxel unit of this layer is used as the query point Q for set abstraction (SA). Calculate the spatial coordinates of the query point Q = {q1,..., q m} according to the voxel index and the actual voxel size, where m is The set of voxel feature vectors of this layer is represented as The superscript q represents the query layer.

[0107] B) Point query clustering

[0108] With each query point q i (i = 1,..., m) as the center, find all non-empty voxels in its neighborhood in this layer according to the set radius r of the k-th layer k to obtain the set of locally grouped non-empty voxels. The set of voxel feature vectors of the k-th layer is represented as The calculation formula is as follows:

[0109]

[0110] In the formula, represents collecting the 16 nearest neighbor voxels in the neighborhood r k and connecting their local relative coordinates with the feature

[0111] C) Point query feature aggregation

[0112] Use set abstraction (SA) (through MLP multi-layer perceptron and max pooling operations) to extract the features of the voxels of the k-th layer:

[0113]

[0114] In the formula, A(.) represents a three-layer MLP network to encode the features, and aggregates the multi-scale features into the query point through the channel-wise max pooling operation MAX(.). In addition, we use two different scales of r in the k-th layer k to increase the local variable receptive field. Aggregate the voxel features of the initial voxel layer and all downsampling layers into the multi-scale voxel feature f i of the query point q i (qv) The aggregation formula is as follows:

[0115]

[0116] where i = 1,..., m, Represents the feature concatenation operation; the generated feature f i (qv) It contains both the multi-scale downsampling features of 3D CNN and the voxel initialization features (i.e. 3D coordinates and reflection intensity).

[0117] (3) Design of BEV multi-level auxiliary branch feature extraction module

[0118] This example designs a BEV multi-level auxiliary branch feature extraction module, which uses a completely sparse and highly compressed method to map the hierarchical multi-scale voxel features extracted in step S2 into a BEV feature map. It then fuses BEV feature maps of different resolutions through 2D sparse convolution processing and jump connections. The algorithm flow is as follows: Figure 4 shown.

[0119] A) Spatial multi-rate feature mapping: A fully sparse height compression (SHC) operation is used to map the multi-scale non-empty voxel features in step S23 onto the XY plane grid, and the features at the same Z position are accumulated to finally generate BEV sparse feature maps with 1x, 2x, 4x and 8x downsampling magnifications.

[0120] B) BEV multi-level feature extraction: BEV sparse feature map of 1x magnification obtained in step S41 Perform three-level 2D sparse convolution downsampling operations to extract BEV sparse convolution features with 2x, 4x, and 8x downsampling ratios layer by layer

[0121] C) Hierarchical feature fusion mechanism: During the downsampling process, a hierarchical feature fusion mechanism is introduced to fuse the multi-rate BEV features step by step in the order of 1x → 2x → 4x → 8x. Extracting high-resolution global semantic information and capturing more detailed contextual information results in 8x aggregated features with rich contextual information. By integrating BEV features of different spatial resolutions, the detection accuracy of small objects (such as pedestrians or obstacles) in complex scenes is improved.

[0122] D) Convert the 8x aggregated features obtained in step S43 into 2D dense features and use them as the input of the detection head for target classification and localization tasks.

[0123] (2) 3D regression box centroid offset loss

[0124] This example introduces the centroid offset loss of the 3D regression box to optimize the multi-task loss, guiding the detection head to achieve the train forward clearance target classification and 3D bounding box regression. The detection effect is as follows Figure 5 shown.

[0125] A) The 2D feature extraction network is responsible for further processing the 2D dense features and generating the final detection output, including two sub-networks: a top-down sub-network performs a 1x1 convolution and two downsamplings (16x, 32x) on the 2D dense features, and another sub-network performs upsampling in the reverse direction to restore the spatial resolution of the input (8x). Finally, the features of all layers and the original input features are aggregated and input into the detection head to achieve the classification of the train's forward clearance target and the regression of the 3D bounding box.

[0126] B) During the model convergence process, a multi-task loss function is adopted, including classification loss, localization loss, orientation loss, and the centroid offset loss L of the 3D regression box centroid which is calculated as follows:

[0127]

[0128] where C box and C gt are the average coordinates of the spatial points within the 3D predicted bounding box and the ground truth box respectively, γ and τ are the local densities of each point estimated by the kernel density estimation (KDE) method, and we introduce the L2 norm ||.||2 to calculate the offset of the centroid of the 3D predicted regression box from the centroid of the ground truth box.

[0129] The overall loss function is defined as:

[0130]

[0131] where L cls is the classification loss, L loc is the localization loss, L dir is the orientation loss, and N pos represents the number of positive example anchors, where λ1 = 1, λ2 = 2, λ3 = 0.2, and λ4 = 1 are the weight hyperparameters for balancing the losses of each part.

[0132] In summary, this example presents a method and system for forward clearance target detection of trains based on enhanced point cloud feature aggregation. Preferably, through a series of steps, the method first uses a lidar installed at the front end of the train to obtain on-vehicle point cloud data, thereby constructing a scene point cloud data set. In this example, the sparse point cloud is divided into regular voxel grids, and voxel feature encoding is used to convert the point cloud information into a compact feature representation, and 3D sparse convolution is used for downsampling layer by layer to extract multi-scale voxel features, so as to realize forward clearance target detection of trains. At the same time, this example uses a designed multi-scale voxel feature aggregation module based on point query to sample, group, and extract features from the extracted multi-level voxel features through set sampling, and further aggregates the multi-scale voxel features to the voxel center points. Then, the designed BEV multi-level auxiliary branch feature extraction module uses a completely sparse and highly compressed method to map the hierarchical multi-scale voxel features into a BEV feature map, and uses 2D sparse convolution processing and skip connections to fuse BEV feature maps of different resolutions, further improving the feature expression ability. In particular, this example optimizes the multi-task loss by introducing the centroid offset loss of the 3D regression box to guide the detection head to realize the classification and three-dimensional bounding box regression of the forward clearance target of the train.

[0133] The method and system for forward clearance target detection of trains based on enhanced point cloud feature aggregation in this example realize accurate detection of forward clearance targets through the acquisition and processing of lidar data, combined with sparse convolution and feature aggregation methods, thereby improving the safety and intelligent level of trains, and providing important technical support for the construction of intelligent railway systems.

[0134] Embodiment 3

[0135] This Embodiment 3 provides a non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the method for forward clearance target detection of trains based on enhanced point cloud feature aggregation as described above is realized. The method includes:

[0136] Obtain on-vehicle point cloud data and construct a scene point cloud data set;

[0137] Divide the sparse point cloud into regular voxel grids, use voxel feature encoding to convert the point cloud information into a compact feature representation, and use 3D sparse convolution for downsampling layer by layer to extract multi-scale voxel features;

[0138] Use set sampling to sample, group, and extract features from the extracted multi-level voxel features, and aggregate the multi-scale voxel features to the voxel center points;

[0139] Use a completely sparse and highly compressed method to map the extracted hierarchical multi-scale voxel features into a BEV feature map, and through 2D sparse convolution processing and skip connections, fuse BEV feature maps of different resolutions;

[0140] Introduce the centroid offset loss of the 3D regression box to optimize the multi-task loss and guide the detection head to achieve the classification of the train forward clearance target and the 3D bounding box regression.

[0141] Example 4

[0142] This Example 4 provides a computer device, including a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the train forward clearance target detection method based on point cloud feature aggregation enhancement as described above. The method includes:

[0143] Obtain on-vehicle point cloud data and construct a scene point cloud data set;

[0144] Divide the sparse point cloud into regular voxel grids, voxel feature encoding converts the point cloud information into a compact feature representation, and use 3D sparse convolution to downsample layer by layer to extract multi-scale voxel features;

[0145] Use set sampling to sample, group and extract features from the extracted multi-level voxel features, and aggregate the multi-scale voxel features to the voxel center point;

[0146] Adopt a fully sparse and highly compressed method to map the extracted hierarchical multi-scale voxel features into a BEV feature map, and through 2D sparse convolution processing and skip connections, fuse BEV feature maps of different resolutions;

[0147] Introduce the centroid offset loss of the 3D regression box to optimize the multi-task loss and guide the detection head to achieve the classification of the train forward clearance target and the 3D bounding box regression.

[0148] Example 5

[0149] This Example 5 provides an electronic device, including: a processor, a memory and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory so that the electronic device executes the instructions to implement the train forward clearance target detection method based on point cloud feature aggregation enhancement as described above. The method includes:

[0150] Obtain on-vehicle point cloud data and construct a scene point cloud data set;

[0151] Divide the sparse point cloud into regular voxel grids, voxel feature encoding converts the point cloud information into a compact feature representation, and use 3D sparse convolution to downsample layer by layer to extract multi-scale voxel features;

[0152] Sampling, grouping, and feature extraction are performed on the extracted multi-level voxel features using set sampling, aggregating the multi-scale voxel features to the voxel center points;

[0153] A fully sparse and highly compressed method is adopted to map the extracted hierarchical multi-scale voxel features into a BEV feature map, and through 2D sparse convolution processing and skip connections, BEV feature maps of different resolutions are fused;

[0154] The centroid offset loss of the 3D regression box is introduced to optimize the multi-task loss, guiding the detection head to achieve the classification of the train forward clearance target and the regression of the three-dimensional bounding box.

[0155] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0156] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the processes and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0157] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0158] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide means for implementing the functions specified in one process Figure 1One process or multiple processes and / or boxes Figure 1 Steps of the functions specified in one box or multiple boxes.

[0159] Although the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, they do not limit the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts should be covered within the protection scope of the present invention.

Claims

1. A train forward clearance target detection method based on enhanced point cloud feature aggregation, characterized in that, Including: Obtain on-vehicle point cloud data and construct a scene point cloud dataset; Divide the sparse point cloud into regular voxel grids. Voxel feature encoding converts the point cloud information into a compact feature representation, and 3D sparse convolution is used for layer-by-layer downsampling to extract multi-scale voxel features; Sampling, grouping, and feature extraction of the extracted multi-level voxel features are performed using set sampling, aggregating the multi-scale voxel features to the voxel center points, including: Point query feature grouping: The 8x downsampling layer after conventional 3D sparse convolution feature encoding is used as the aggregation layer. The center of the voxel unit in this layer is used as the query point Q for set sampling, and the spatial coordinates of the query point are calculated based on the voxel index and the actual voxel size; Point query clustering: With each query point as the center, all non-empty voxels within the set radius r of the k-th layer are found in its neighborhood in this layer, obtaining the set of locally grouped non-empty voxels, and calculating the set of voxel feature vectors of the k-th layer; Point query feature aggregation: Set sampling is used to extract the features of the voxels of the k-th layer; k Finding all non-empty voxels within its neighborhood in this layer, obtaining the set of locally grouped non-empty voxels, and calculating the set of voxel feature vectors of the k-th layer; Point query feature aggregation: Using set sampling to extract the features of the voxels of the k-th layer; Adopt a fully sparse and highly compressed method to map the extracted hierarchical multi-scale voxel features into a BEV feature map, and through 2D sparse convolution processing and skip connections, fuse BEV feature maps of different resolutions; including: Spatial multi-magnification feature mapping: Adopt a fully sparse and highly compressed operation to map the multi-scale non-empty voxel features onto the X-Y plane grid, and accumulate the features at the same Z position. Finally, generate BEV sparse feature maps with downsampling magnifications of 1x, 2x, 4x, and 8x; BEV multi-level feature extraction: Perform three-level 2D sparse convolution downsampling operations on the obtained 1x magnification BEV sparse feature map to extract BEV sparse convolution features with downsampling magnifications of 2x, 4x, and 8x layer by layer; Hierarchical feature fusion mechanism: During the downsampling process, introduce a hierarchical feature fusion mechanism to fuse the multi-magnification BEV features in the order of 1x→2x→4x→8x step by step, extract high-resolution global semantic information, capture more detailed context information, and obtain 8x aggregated features with rich context information; Convert the obtained 8x aggregated features into 2D dense features and use them as the input of the detection head for target classification and localization tasks; Introduce the centroid offset loss of the 3D regression box to optimize the multi-task loss and guide the detection head to achieve train forward clearance target classification and three-dimensional bounding box regression; including: The 2D feature extraction network is responsible for further processing the 2D dense features and generating the final detection output, including two sub-networks: A top-down sub-network performs a 1x1 convolution and two downsamplings on the 2D dense features, and the other sub-network performs upsampling in the reverse direction to restore the spatial resolution of the input. Finally, aggregate the features of all layers and the original input features into the detection head to achieve train forward clearance target classification and three-dimensional bounding box regression.

2. The method for detecting the forward clearance target of a train based on point cloud feature aggregation enhancement according to claim 1, wherein: The method of dividing sparse point clouds into regular voxel grids, encoding voxel features to convert point cloud information into a compact feature representation, and using 3D sparse convolution for layer-by-layer downsampling to extract multi-scale voxel features includes: Voxel gridification: Divide the input N spatial point clouds P = (p1, p2,..., p n ) region into voxel units with a spatial resolution of L×W×H, where the point cloud feature representation is (x, y, z, r), that is, the 3D coordinates and reflection intensity of the point cloud; Voxel feature encoding: Average the point cloud features within the voxel unit to obtain the initially encoded voxel features, represented as a four-dimensional tensor; Use 3D submanifold sparse convolution to perform layer-by-layer downsampling on the voxel features, and encode the features of non-empty voxels into multi-scale three-dimensional sparse feature quantities with downsampling ratios of 1x, 2x, 4x, and 8x respectively; Among them, the voxel spatial coordinates of the k-th layer of downsampling are represented as The set of its voxel feature quantities is represented as where N k is the number of non-empty voxels in the k-th layer, and the superscript l k (k = 1, 2, 3, 4) represents the downsampling layers with ratios of 1x, 2x, 4x, and 8x.

3. The method for detecting the forward clearance target of a train based on point cloud feature aggregation enhancement according to claim 2, wherein: Use two different scales of r in layer k k Increase the local variable receptive field, and aggregate the voxel features of the initial voxel layer and all downsampling layers into the query point q i of multi-scale voxel features.

4. A train forward clearance target detection system based on enhanced point cloud feature aggregation, characterized in that, Including: An acquisition module for obtaining on-vehicle point cloud data and constructing a scene point cloud dataset; A downsampling extraction module for dividing the sparse point cloud into regular voxel grids. Voxel feature encoding converts the point cloud information into a compact feature representation, and 3D sparse convolution is used for layer-by-layer downsampling to extract multi-scale voxel features; Multi-scale voxel feature aggregation module based on point query, which is used to sample, group and extract features from the extracted multi-level voxel features using set sampling, and aggregate the multi-scale voxel features to the voxel center points; it includes: Point query feature grouping: The 8x downsampling layer is used as the aggregation layer after conventional 3D sparse convolution feature encoding. The center of the voxel unit in this layer is used as the query point Q for set sampling, and the spatial coordinates of the query point are calculated according to the voxel index and the actual size of the voxel; Point query clustering: With each query point as the center, all non-empty voxels in its neighborhood in the k-th layer are found according to the set radius r of the k-th layer k to obtain the set of locally grouped non-empty voxels, and calculate the set of voxel feature vectors in the k-th layer; Point query feature aggregation: Use set sampling to extract the features of the voxels in the k-th layer; BEV Multi - level Auxiliary Branch Feature Extraction Module, which is used to map the extracted hierarchical multi - scale voxel features into BEV feature maps by using a fully sparse and highly compressed method, and fuse BEV feature maps with different resolutions through 2D sparse convolution processing and skip connections; it includes: Spatial Multi - magnification Feature Mapping: Adopting a fully sparse and highly compressed operation, mapping multi - scale non - empty voxel features onto the X - Y plane grid, and accumulating features at the same Z position. Finally, BEV sparse feature maps with downsampling magnifications of 1x, 2x, 4x, and 8x are generated; 5. A non-transitory computer-readable storage medium, characterized in that, BEV Multi - level Feature Extraction: For the obtained BEV sparse feature map with a 1x magnification, perform three - level 2D sparse convolution downsampling operations to gradually extract BEV sparse convolution features with downsampling magnifications of 2x, 4x, and 8x; 6. A computer device, characterized in that, Hierarchical Feature Fusion Mechanism: During the downsampling process, introduce a hierarchical feature fusion mechanism to fuse multi - magnification BEV features step by step in the order of 1x→2x→4x→8x, extract high - resolution global semantic information, capture more detailed context information, and obtain 8x aggregated features with rich context information; convert the obtained 8x aggregated features into 2D dense features and use them as the input of the detection head for object classification and localization tasks; 7. An electronic device, characterized in that, Detection Module, which is used to introduce the centroid offset loss of the 3D regression box to optimize the multi - task loss and guide the detection head to achieve the classification of the train's forward clearance target and the regression of the three - dimensional bounding box; it includes: The 2D Feature Extraction Network is responsible for further processing the 2D dense features and generating the final detection output, including two sub - networks: A top - down sub - network performs a 1x1 convolution and two downsamplings on the 2D dense features, and another sub - network performs upsampling in the reverse direction to restore the spatial resolution of the input, and finally aggregates the features of all layers and the original input features into the detection head to achieve the classification of the train's forward clearance target and the regression of the three - dimensional bounding box. The non - transitory computer - readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the method for detecting the train's forward clearance target based on point - cloud feature aggregation enhancement as described in any one of claims 1 - 3 is implemented. It includes a memory and a processor. The processor and the memory communicate with each other. The memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the method for detecting the train's forward clearance target based on point - cloud feature aggregation enhancement as described in any one of claims 1 - 3. It includes: A processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory so that the electronic device executes the instructions for implementing the method for detecting the train's forward clearance target based on point - cloud feature aggregation enhancement as described in any one of claims 1 - 3.

Images