Point cloud column three-dimensional target detection method and device

Abstract

The application provides a point cloud column three-dimensional target detection method and device, which has the following characteristics: a three-dimensional target detection model comprises a column feature coding module used to obtain a pseudo image; a pseudo image feature extraction module used to perform feature extraction on the pseudo image to obtain a feature vector; and a bounding box prediction module used to obtain a target detection result according to the feature vector. The pseudo image feature extraction module comprises: a down-sampling unit used to perform multiple times of down-sampling on the pseudo image to obtain a down-sampled feature map; a first up-sampling unit used to perform multiple times of up-sampling and jump connection on the down-sampled feature map to obtain an up-sampled feature map; a second up-sampling unit used to up-sample the down-sampled feature map with the smallest size and each up-sampled feature map to a preset dimension respectively to obtain corresponding sampling feature maps; and a connection unit used to connect all the sampling feature maps in series to obtain the feature vector. In summary, the method can improve the accuracy of target detection based on point cloud data.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision and autonomous driving perception, and relates to a 3D target detection method, specifically a point cloud column 3D target detection method and device. Background Technology

[0002] Autonomous vehicles need to use various sensors, such as cameras, LiDAR, and millimeter-wave radar, to perceive their surroundings and provide accurate driving environment information. Object detection is a crucial task in perception, helping intelligent vehicles perceive the position and semantic information of surrounding objects in complex road scenarios and predict accurate bounding boxes. This primarily involves predicting the bounding boxes of three common object types in autonomous driving scenarios: people, vehicles, and cyclists.

[0003] Autonomous vehicles rely on multiple sensors, among which LiDAR is crucial for 3D environmental perception in autonomous driving scenarios, possessing rich geometric information lacking in images. With the significant advancements in deep learning methods in computer vision, numerous studies have explored how to apply this technology to target detection using LiDAR point clouds, achieving continuous improvements and progress, and finding widespread application in multi-target 3D target detection in autonomous driving scenarios.

[0004] However, existing technologies for target detection using point cloud data still suffer from low accuracy. Summary of the Invention

[0005] This invention is made to solve the above-mentioned problems, and its purpose is to provide a method and device for detecting three-dimensional targets of point cloud pillars.

[0006] This invention provides a point cloud-based 3D target detection method for pillars, used to obtain corresponding target detection results based on point cloud data of the target. The method includes the following steps: Step S1, preprocessing the point cloud data to obtain a preprocessed point cloud; Step S2, inputting the preprocessed point cloud into a 3D target detection model to obtain the target detection result. The 3D target detection model includes: a pillar feature encoding module for extracting pillar features from the preprocessed point cloud to obtain a pseudo-image; a pseudo-image feature extraction module for extracting features from the pseudo-image to obtain a feature vector; and a bounding box prediction module for obtaining a 3D detection box based on the feature vector. For the target detection results, the pseudo-image feature extraction module includes: a downsampling unit, used to downsample the pseudo-image multiple times to obtain multiple downsampled feature maps of different sizes; a first upsampling unit, used to upsample the smallest downsampled feature map multiple times, and concatenate the feature map obtained from each upsampling with the downsampled feature map of the corresponding size to obtain multiple upsampled feature maps of different sizes; a second upsampling unit, used to upsample the smallest downsampled feature map and each upsampled feature map to a preset dimension respectively to obtain the corresponding sampled feature maps; and a concatenation unit, used to concatenate all sampled feature maps to obtain a feature vector.

[0007] The point cloud column 3D target detection method provided by the present invention may also have the following feature: wherein, preprocessing is used to screen out outliers in the point cloud data.

[0008] The point cloud column 3D target detection method provided by this invention may also have the following features: the column feature encoding module includes: a size tensor generation unit, used to mesh and sample the preprocessed point cloud to obtain a size tensor; a feature extraction unit, used to extract features from the size tensor to obtain a tensor; a max pooling unit, used to perform max pooling on the tensor to obtain an output tensor; and a transformation unit, used to perform coordinate transformation on the output tensor according to the column index to obtain a pseudo image.

[0009] The point cloud column 3D target detection method provided by this invention may also have the following features: the downsampling unit includes a first convolution sub-unit, a second convolution sub-unit, a first SE sub-unit, a third convolution sub-unit, and a second SE sub-unit connected in sequence. The pseudo image is processed by the first convolution sub-unit to obtain a first downsampling feature map. The first downsampling feature map is processed by the second convolution sub-unit and the first SE sub-unit in sequence to obtain a second downsampling feature map. The second downsampling feature map is processed by the third convolution sub-unit and the second SE sub-unit in sequence to obtain a third downsampling feature map. The first downsampling feature map, the second downsampling feature map, and the third downsampling feature map are downsampling feature maps of different sizes, and the third downsampling feature map is the smallest downsampling feature map.

[0010] The point cloud column 3D target detection method provided by the present invention may also have the following features: wherein the first upsampling unit includes a first deconvolution sub-unit and a second deconvolution sub-unit; the feature map obtained by the upsampling processing of the third downsampling feature map by the first deconvolution sub-unit is concatenated with the second downsampling feature map to obtain the first upsampling feature map; the feature map obtained by the upsampling processing of the first upsampling feature map by the second deconvolution sub-unit is concatenated with the first downsampling feature map to obtain the second upsampling feature map; and the first upsampling feature map and the second upsampling feature map are upsampling feature maps of different sizes.

[0011] The point cloud column 3D target detection method provided by the present invention may also have the following features: wherein the bounding box prediction module performs 3D target detection through the SSD detection head and uses the NMS method to eliminate redundant 3D detection boxes.

[0012] The point cloud column 3D target detection method provided by this invention may also have the following features: the training process of the 3D target detection model includes the following steps: Step T1, constructing a training dataset based on existing point cloud data; Step T2, constructing an initial 3D target detection model; Step T3, selecting multiple frames of point cloud data from the training dataset as a batch, inputting them into the initial 3D target detection model, and obtaining the training target detection result; Step T4, calculating the loss error based on the training target detection result; Step T5, updating the parameters of the initial 3D target detection model based on the loss error; Step T6, repeating steps T3 to T5 until the training completion condition is met, then using the trained initial 3D target detection model as the 3D target detection model.

[0013] The point cloud column 3D target detection method provided by the present invention may also have the following features: wherein the loss error includes category loss, position loss and orientation classification loss, the category loss is calculated by focal loss function, and the position loss is calculated by SmoothL1 loss function.

[0014] This invention also provides a point cloud column 3D target detection device for obtaining corresponding target detection results based on point cloud data of a target. It comprises: a preprocessing unit for preprocessing the point cloud data to obtain a preprocessed point cloud; and a detection unit including a 3D target detection model for obtaining target detection results based on the preprocessed point cloud. The 3D target detection model includes: a column feature encoding module for extracting column features from the preprocessed point cloud to obtain a pseudo-image; a pseudo-image feature extraction module for extracting features from the pseudo-image to obtain a feature vector; and a bounding box prediction module for obtaining a 3D detection box based on the feature vector. As the target detection result, the pseudo-image feature extraction module includes: a downsampling unit, used to downsample the pseudo-image multiple times to obtain multiple downsampled feature maps of different sizes; a first upsampling unit, used to upsample the smallest downsampled feature map multiple times, and concatenate the feature map obtained from each upsampling with the downsampled feature map of the corresponding size to obtain multiple upsampled feature maps of different sizes; a second upsampling unit, used to upsample the smallest downsampled feature map and each upsampled feature map to a preset dimension respectively to obtain the corresponding sampled feature maps; and a concatenation unit, used to concatenate all sampled feature maps to obtain a feature vector.

[0015] The role and effect of invention

[0016] According to the point cloud column 3D target detection method and apparatus of the present invention, by preprocessing the point cloud data to remove outliers and noise, the impact of data noise on the detection results is reduced; by using a downsampling unit to adaptively adjust the weights of different channels with channel attention, higher quality features are obtained; and by using a first upsampling unit to stitch the upsampled feature map with the corresponding downsampled feature map for multi-scale feature fusion, the mutual transfer, supplementation, and integration of feature map information at different scales during feature extraction can be achieved, thereby effectively enhancing the model's feature extraction capability and perception capability for small targets, maximizing information utilization, and obtaining more accurate and comprehensive target detection results. Therefore, the point cloud column 3D target detection method and apparatus of the present invention can improve the accuracy of target detection based on point cloud data. Attached Figure Description

[0017] Figure 1 This is a block diagram of the point cloud column three-dimensional target detection device in an embodiment of the present invention;

[0018] Figure 2 This is a block diagram of the three-dimensional target detection model in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the process of training a three-dimensional target detection model in an embodiment of the present invention;

[0020] Figure 4 This is a flowchart illustrating the point cloud column 3D target detection method in an embodiment of the present invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of the present invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, provide a detailed description of the point cloud column three-dimensional target detection method and device of the present invention.

[0022] This embodiment provides a point cloud column 3D target detection device, which is used to obtain the corresponding target detection result based on the point cloud data of the target.

[0023] Figure 1 This is a block diagram of the point cloud column three-dimensional target detection device in an embodiment of the present invention.

[0024] like Figure 1 As shown, the point cloud column 3D target detection device 100 includes a preprocessing unit 10 and a detection unit 20.

[0025] The preprocessing unit 10 is used to preprocess the point cloud data to obtain a preprocessed point cloud.

[0026] Preprocessing is used to remove outliers from the point cloud data. In this embodiment, noise removal and other operations are performed on the point cloud data to remove outliers, thereby reducing the impact of data noise on the detection results.

[0027] The detection unit 20 includes a three-dimensional target detection model, which is used to obtain target detection results based on the preprocessed point cloud.

[0028] Figure 2 This is a block diagram of a three-dimensional target detection model in an embodiment of the present invention.

[0029] like Figure 2 As shown, the three-dimensional target detection model 200 includes a column feature encoding module 30, a pseudo-image feature extraction module 40, and a bounding box prediction module 50.

[0030] The pillar feature encoding module 30 is used to extract pillar features from the preprocessed point cloud to obtain a pseudo image. The pillar feature encoding module 30 includes a size tensor generation unit 301, a feature extraction unit 302, a max pooling unit 303, and a transformation unit 304.

[0031] The size tensor generation unit 301 is used to mesh and sample the preprocessed point cloud to obtain a size tensor.

[0032] In this embodiment, the size tensor generation unit 301 discretizes the point cloud into a uniformly spaced grid on the xy plane, creating a set of pillars, each a voxel of infinite height on the z-axis. Then, P non-empty pillars (pillars containing points) are selected, and N points are taken from each pillar. The dimension of each point is D = 9, thus obtaining a size tensor of size (D, P, N). Here, a voxel is the smallest unit in three-dimensional space, equivalent to a pixel in two-dimensional space.

[0033] If the number of points in a non-empty column is greater than N, then N points are obtained by random sampling; if the number of points in a non-empty column is less than N, then 0 is used to fill the gaps.

[0034] A point with D dimensions can be represented as a vector (x, y, z, r, x). c ,y c ,z c ,x p ,y p (), where x, y, and z represent three-dimensional coordinates, r represents reflectivity, x c y c z c x represents the distance from the point to the arithmetic mean of all points on the column. p y p These represent the offsets of the point from the center of the column along the x-axis and y-axis, respectively.

[0035] The feature extraction unit 302 is used to extract features from the size tensor to obtain the tensor.

[0036] In this embodiment, the feature extraction unit 302 is PointNet, which includes a 1×1 convolutional layer (Linear), the ReLU activation function, and batch normalization (BN). PointNet processes each point in the size tensor to obtain a tensor of size (C, P, N).

[0037] The max pooling unit 303 is used to perform max pooling on the tensor to obtain the output tensor. In this embodiment, the max pooling unit 303 performs max pooling on each column to obtain an output tensor of size (C, P).

[0038] The transformation unit 304 is used to perform coordinate transformation on the output tensor according to the column index to obtain a pseudo-image. In this embodiment, the transformation unit 304 obtains a pseudo-image of size (C, H, W), where C is the number of feature channels.

[0039] In this embodiment, the pillar index is the position index of the voxel in three-dimensional space, i.e. (x, y, z). Since the pillar is a voxel with infinite height on the z-axis, the coordinate of the z-axis is not considered. The pillar coordinate is the coordinate of the pillar on the xy plane, which can be represented as (x, y).

[0040] The pseudo-image feature extraction module 40 is used to extract features from the pseudo-image to obtain a feature vector. The pseudo-image feature extraction module 40 includes a downsampling unit 401, a first upsampling unit 402, a second upsampling unit 403, and a connection unit 404.

[0041] The downsampling unit 401 is used to downsample the pseudo image multiple times to obtain multiple downsampled feature maps of different sizes.

[0042] The downsampling unit 401 includes a first convolutional sub-unit, a second convolutional sub-unit, a first SE sub-unit, a third convolutional unit, and a second SE sub-unit connected in sequence. In this embodiment, the SE sub-unit adaptively adjusts the weights of different channels through channel attention to obtain higher quality features.

[0043] The pseudo-image is processed by the first convolutional subunit to obtain the first downsampled feature map. The first downsampled feature map is then processed by the second convolutional subunit and the first SE subunit to obtain the second downsampled feature map. The second downsampled feature map is then processed by the third convolutional subunit and the second SE subunit to obtain the third downsampled feature map.

[0044] The first, second, and third downsampled feature maps are downsampled feature maps of different sizes. The third downsampled feature map is the smallest downsampled feature map.

[0045] The first upsampling unit 402 is used to perform multiple upsamplings on the smallest downsampled feature map, and to concatenate the feature map obtained from each upsampling with the downsampled feature map of the corresponding size to obtain multiple upsampled feature maps of different sizes.

[0046] The first upsampling unit 402 includes a first deconvolution subunit and a second deconvolution subunit.

[0047] The feature map obtained by upsampling the third downsampled feature map through the first deconvolution subunit is concatenated with the second downsampled feature map to obtain the first upsampled feature map. The feature map obtained by upsampling the first upsampled feature map through the second deconvolution subunit is concatenated with the first downsampled feature map to obtain the second upsampled feature map.

[0048] The first upsampled feature map and the second upsampled feature map are upsampled feature maps of different sizes.

[0049] The second upsampling unit 403 is used to upsample the smallest downsampled feature map and each upsampled feature map to a preset dimension, respectively, to obtain the corresponding sampled feature map.

[0050] The connecting unit 404 is used to concatenate all sampled feature maps to obtain a feature vector.

[0051] The bounding box prediction module 50 is used to obtain a 3D detection box as the target detection result based on the feature vector.

[0052] The bounding box prediction module 50 performs 3D target detection using the SSD detection head and uses the NMS method to eliminate redundant 3D detection boxes.

[0053] In this embodiment, since the features are planar 2D feature maps from a bird's-eye view, 2D IOU is used to match the predicted bounding boxes and ground truth boxes, while the height is not used. The height of the object is an additional regression target, that is, object detection is performed in a 2D grid. The z-axis coordinates and height are obtained through model regression, and then NMS is used to eliminate redundant 3D detection boxes.

[0054] In this embodiment, a 3D target detection model 200 is trained based on existing point cloud data, thereby improving the target detection performance of the 3D target detection model 200.

[0055] Figure 3 This is a schematic diagram of the process of training a three-dimensional target detection model in an embodiment of the present invention.

[0056] like Figure 3 As shown, the training process of the 3D object detection model includes the following steps:

[0057] Step T1: Construct a training dataset based on existing point cloud data. In this embodiment, the existing KITTI dataset is used as the existing point cloud data. This KITTI dataset contains images and point cloud data collected from real-world scenes such as urban areas, rural areas, and highways, and mainly annotates three types of objects in the scene: pedestrians, cars, and cyclists.

[0058] Step T2: Construct the initial 3D target detection model.

[0059] Step T3: Select multiple frames of point clouds from the training dataset as a batch, input them into the initial 3D object detection model, and obtain the training object detection results.

[0060] In this embodiment, 16 frames of point clouds are treated as a batch. The size tensor generation unit 301 first encodes the batch into a size of B*P*N*D. The feature extraction unit 302 then generates a tensor of size (B,C,P,N) for each point. The max pooling unit 303 then outputs an output tensor of size (B,C,P) based on the tensor. Finally, the transformation unit 304 performs coordinate transformation on the output tensor to obtain a pseudo-image of size (B,C,H,W).

[0061] After the above process, each channel stores H*W space containing data for one feature dimension of all pillars within a frame, for a total of 64 feature dimensions, i.e., B*64*H*W. Here, H and W are related to the preset pillar size; in this embodiment, H = 432 and W = 496.

[0062] The pseudo-images continue to be processed by subsequent modules until the bounding box prediction module 50 predicts the target boxes based on the SSD detection head, with an output data size of B*42*H*W. Here, 42 represents the 6 anchors in each region, and the predicted values ​​for each box are 7 dimensions, specifically xyzwhlr, where xyz represents the xyz coordinates of the center point, hwl represents the length, width, and height, and r represents the rotation angle.

[0063] Step T4: Calculate the loss error based on the training target detection results.

[0064] The loss error includes category loss, location loss, and orientation classification loss. Category loss is calculated using the focal loss function to address the imbalance of class samples in existing point cloud data. Location loss is calculated using the SmoothL1 loss function, which calculates the difference between the target box's 3D coordinates (length, width, and height) and the ground truth values. Orientation classification loss is used to prevent errors in orientation loss, as the original orientation loss penalizes both forward and backward directions equally; therefore, it's necessary to additionally predict the orientation of the bounding box.

[0065] Step T5: Update the parameters of the initial 3D target detection model based on the loss error.

[0066] Step T6: Repeat steps T3 to T5 until the training completion condition is met. Then, use the trained initial 3D object detection model as the 3D object detection model 200.

[0067] The training completion condition is that the number of training iterations reaches a preset value. In this embodiment, the preset value is 80.

[0068] The process of performing point cloud column three-dimensional target detection using the point cloud column three-dimensional target detection device 100 of the present invention will be described below with reference to the accompanying drawings.

[0069] Figure 4 This is a flowchart illustrating the point cloud column 3D target detection method in an embodiment of the present invention.

[0070] like Figure 4 As shown, the point cloud column 3D target detection method includes the following steps:

[0071] Step S1: The point cloud data is preprocessed by the preprocessing unit 10 to obtain a preprocessed point cloud.

[0072] Step S2: The preprocessed point cloud is input into the three-dimensional target detection model 200 using the detection unit 20 to obtain the target detection result.

[0073] In this embodiment, to verify the target detection performance of the point cloud column 3D target detection method (i.e., the proposed method), it was tested against the existing PointPillars method on the KITTI dataset. The results showed that the proposed method achieved an mAP of 60.1 for the three classes of people, vehicles, and cyclists, while the PointPillars method achieved an mAP of 59.2 for the same three classes. Therefore, it is evident that the proposed method has better target detection accuracy compared to the existing PointPillars method.

[0074] The role and effect of the embodiments

[0075] According to the point cloud column 3D target detection method and apparatus involved in this embodiment, the impact of data noise on the detection results is reduced by preprocessing the point cloud data to remove outliers and noise points; the downsampling unit adaptively adjusts the weights of different channels using channel attention to obtain higher quality features; and the first upsampling unit concatenates the upsampled feature map with the corresponding downsampled feature map for multi-scale feature fusion, enabling the mutual transfer, supplementation, and integration of information from different scale feature maps during feature extraction. This effectively enhances the model's feature extraction capability and its ability to perceive small targets, maximizing information utilization and obtaining more accurate and comprehensive target detection results. In summary, this method and apparatus can improve the accuracy of target detection based on point cloud data.

[0076] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A point cloud-based three-dimensional target detection method for pillars, used to obtain corresponding target detection results based on point cloud data of the target, characterized in that, Includes the following steps: Step S1: Preprocess the point cloud data to obtain a preprocessed point cloud; Step S2: Input the preprocessed point cloud into the 3D target detection model to obtain the target detection result. The three-dimensional target detection model includes: The column feature encoding module is used to extract column features from the preprocessed point cloud to obtain a pseudo image; The pseudo-image feature extraction module is used to extract features from the pseudo-image to obtain a feature vector; The bounding box prediction module is used to obtain a 3D detection box as the target detection result based on the feature vector. The pseudo-image feature extraction module includes: The downsampling unit is used to downsample the pseudo-image multiple times to obtain multiple downsampled feature maps of different sizes; The first upsampling unit is used to perform multiple upsamplings on the downsampled feature map of the smallest size, and to concatenate the feature map obtained by each upsampling with the downsampled feature map of the corresponding size to obtain multiple upsampled feature maps of different sizes. The second upsampling unit is used to upsample the smallest downsampled feature map and each of the upsampled feature maps to a preset dimension, respectively, to obtain the corresponding sampled feature maps; The connection unit is used to concatenate all the sampled feature maps to obtain the feature vector. The downsampling unit includes a first convolution sub-unit, a second convolution sub-unit, a first SE sub-unit, a third convolution sub-unit, and a second SE sub-unit connected in sequence. The pseudo-image is processed by the first convolutional subunit to obtain a first downsampled feature map. The first downsampled feature map is processed sequentially by the second convolutional subunit and the first SE subunit to obtain the second downsampled feature map. The second downsampled feature map is processed sequentially by the third convolutional subunit and the second SE subunit to obtain the third downsampled feature map. The first downsampled feature map, the second downsampled feature map, and the third downsampled feature map are downsampled feature maps of different sizes. The third downsampled feature map is the downsampled feature map of the smallest size. The first upsampling unit includes a first deconvolution subunit and a second deconvolution subunit. The feature map obtained by upsampling the third downsampled feature map through the first deconvolution subunit is concatenated with the second downsampled feature map to obtain the first upsampled feature map. The first upsampled feature map, after being upsampled by the second deconvolution subunit, is concatenated with the first downsampled feature map to obtain the second upsampled feature map. The first upsampled feature map and the second upsampled feature map are upsampled feature maps of different sizes.

2. The point cloud column three-dimensional target detection method according to claim 1, characterized in that: in, The preprocessing is used to remove outliers and noise from the point cloud data.

3. The point cloud column 3D target detection method according to claim 1, Its features are: The column feature encoding module includes: A size tensor generation unit is used to mesh and sample the preprocessed point cloud to obtain a size tensor. A feature extraction unit is used to extract features from the size tensor to obtain the tensor. The max pooling unit is used to perform max pooling operation on the tensor to obtain the output tensor. The transformation unit is used to perform coordinate transformation on the output tensor according to the column index to obtain the pseudo image.

4. The point cloud column three-dimensional target detection method according to claim 1, characterized in that: in, The bounding box prediction module performs 3D target detection using an SSD detection head and employs the NMS method to eliminate redundant 3D detection boxes.

5. The point cloud column 3D target detection method according to claim 1, Its features are: The training process of the three-dimensional target detection model includes the following steps: Step T1: Construct a training dataset based on the existing point cloud data; Step T2: Construct the initial 3D target detection model; Step T3: Select multiple frames of point clouds from the training dataset as a batch, input them into the initial 3D object detection model, and obtain the training object detection results; Step T4: Calculate the loss error based on the training target detection results; Step T5: Update the parameters of the initial 3D target detection model based on the loss error; Step T6: Repeat steps T3 to T5 until the training completion condition is met, then use the trained initial 3D object detection model as the 3D object detection model.

6. The point cloud column three-dimensional target detection method according to claim 5, characterized in that: in, The loss error includes category loss, position loss, and orientation classification loss. The category loss is calculated using the focal loss function. The position loss is calculated using the SmoothL1 loss function.

7. A point cloud column 3D target detection device, used to obtain corresponding target detection results based on the point cloud data of the target, characterized in that, include: The preprocessing unit is used to preprocess the point cloud data to obtain a preprocessed point cloud; The detection unit includes a three-dimensional target detection model, used to obtain the target detection result based on the preprocessed point cloud. The three-dimensional target detection model includes: The column feature encoding module is used to extract column features from the preprocessed point cloud to obtain a pseudo image; The pseudo-image feature extraction module is used to extract features from the pseudo-image to obtain a feature vector; The bounding box prediction module is used to obtain a 3D detection box as the target detection result based on the feature vector. The pseudo-image feature extraction module includes: The downsampling unit is used to downsample the pseudo-image multiple times to obtain multiple downsampled feature maps of different sizes; The first upsampling unit is used to perform multiple upsamplings on the downsampled feature map of the smallest size, and to concatenate the feature map obtained by each upsampling with the downsampled feature map of the corresponding size to obtain multiple upsampled feature maps of different sizes. The second upsampling unit is used to upsample the smallest downsampled feature map and each of the upsampled feature maps to a preset dimension, respectively, to obtain the corresponding sampled feature maps; The connection unit is used to concatenate all the sampled feature maps to obtain the feature vector. The downsampling unit includes a first convolution sub-unit, a second convolution sub-unit, a first SE sub-unit, a third convolution sub-unit, and a second SE sub-unit connected in sequence. The pseudo-image is processed by the first convolutional subunit to obtain a first downsampled feature map. The first downsampled feature map is processed sequentially by the second convolutional subunit and the first SE subunit to obtain the second downsampled feature map. The second downsampled feature map is processed sequentially by the third convolutional subunit and the second SE subunit to obtain the third downsampled feature map. The first downsampled feature map, the second downsampled feature map, and the third downsampled feature map are downsampled feature maps of different sizes. The third downsampled feature map is the downsampled feature map of the smallest size. The first upsampling unit includes a first deconvolution subunit and a second deconvolution subunit. The feature map obtained by upsampling the third downsampled feature map through the first deconvolution subunit is concatenated with the second downsampled feature map to obtain the first upsampled feature map. The first upsampled feature map, after being upsampled by the second deconvolution subunit, is concatenated with the first downsampled feature map to obtain the second upsampled feature map. The first upsampled feature map and the second upsampled feature map are upsampled feature maps of different sizes.

Images