Event stream segmentation method for small moving targets based on multi-scale point cloud feature fusion

Abstract

The present invention discloses a method for segmenting a small moving target event stream based on multi-scale point cloud feature fusion, which is applied to dynamic visual sensor data processing, including: dividing event stream data into event point clouds; sampling the farthest point, feature learning and splicing the event point clouds to obtain a first local feature; using a point cloud self-attention mechanism to enhance the first local feature to form a second local feature; downsampling the second local feature based on fusion weights, splicing after feature learning in a multi-layer perceptron, and then performing integrated feature learning to form a global feature; after continuously interpolating and upsampling the above features and event point clouds, obtaining multi-layer mixed features through jump links, performing a maximum pooling operation on the multi-layer mixed features to obtain the final one-dimensional credibility, and then performing point-by-point segmentation to obtain the target event point. The present invention realizes the accurate segmentation of small moving targets without relying on artificial prior information.

Description

Technical Field

[0001] The present invention relates to the field of dynamic vision sensor data processing, and in particular to a method for segmenting a small moving target event stream based on multi-scale point cloud feature fusion. Background Art

[0002] Dynamic vision sensor is a new type of optical sensor that simulates the working mechanism of biological vision. The traditional visual acquisition method is based on the "image frame" acquired at a fixed frequency, while all pixels of the dynamic vision sensor work autonomously and asynchronously. Only pixels with changed light intensity are output, and other pixels remain silent. Low redundancy, low latency, and high dynamics are achieved from the source, but the output data is a sparse, discrete, and random event stream, which is difficult to directly adopt for traditional target detection methods based on image frames.

[0003] The event stream of a dynamic vision sensor can be modeled as an event point cloud, which consists of several event points, each of which is three-dimensional in space and time. Target detection on the event stream can be regarded as the segmentation of the event point cloud. The event points belonging to the target are segmented according to the temporal and spatial correlation and significance between the event points. Existing target segmentation methods are aimed at morphological general targets, and the texture, color, shape and other information of the point cloud are used to achieve target segmentation in the three-dimensional point cloud. In systems such as long-distance optical monitoring, the target scale is generally small, with only a few pixels. Such small targets have neither fixed shapes nor color information in the event point cloud of a dynamic vision sensor, and traditional methods that rely on prior information and manually designed features are difficult to segment well. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a moving small target event stream segmentation method based on multi-scale point cloud feature fusion, which solves the technical problem of poor temporal point cloud segmentation effect of dynamic vision sensors in the prior art.

[0005] In order to solve the above technical problems, the present invention provides a moving small target event stream segmentation method based on multi-scale point cloud feature fusion, comprising:

[0006] Acquire event stream data output by a dynamic vision sensor, and divide the event stream data into a plurality of event point clouds at preset time intervals;

[0007] Select any one of the event point clouds for farthest point sampling to form three types of first sub-event point clouds with different scales, use a multi-layer perceptron to perform feature learning on the first sub-event point cloud, and then perform multi-scale feature splicing to form a first local feature;

[0008] The point cloud self-attention mechanism is used to enhance the spatiotemporal features of the first local feature to form a second local feature;

[0009] Downsampling the second local feature based on the fusion weight to form two types of second sub-event point clouds with different scales, and after the second sub-event point clouds complete feature learning in the multi-layer perceptron, performing the multi-scale feature splicing to obtain a third local feature;

[0010] Performing integrated feature learning on the third local feature to form a global feature;

[0011] After continuous interpolation and upsampling according to the global feature, the third local feature, the second local feature and the event point cloud, multi-layer mixed features are obtained through jump links, and a maximum pooling operation is performed on the multi-layer mixed features to obtain a final one-dimensional credibility, and the final one-dimensional credibility is segmented point by point to obtain the target event point.

[0012] Optionally, the selecting any one of the event point clouds for farthest point sampling to form three types of first sub-event point clouds of different scales includes:

[0013] Any one of the event point clouds is selected to perform the farthest point sampling to form a plurality of the first sub-event point clouds at 32 scales, 64 scales and 128 scales.

[0014] Optionally, the step of using a point cloud self-attention mechanism to enhance the spatiotemporal features of the first local feature to form a second local feature includes:

[0015] Obtain W through point cloud neural network training q , W k , W v Three weight matrices, through the formula Q = D Loc1 ×W q , K=D Loc1 ×W k 、V=D Loc1 ×W v , obtain the self-attention matrices Q, K and V, and pass The second local feature is obtained by calculation; wherein, D Loc1 is the first local feature, softmax is a normalized exponential function, and c2 is the dimension of the hidden layer in the point cloud neural network training.

[0016] Optionally, downsampling the second local feature based on a fusion weight to form two types of second sub-event point clouds of different scales includes:

[0017] The second local feature is downsampled based on the fusion weight to form the second sub-event point cloud including a 64-scale and a 128-scale.

[0018] Optionally, the performing the downsampling based on the fusion weight on the second local feature includes:

[0019] The second local feature is downsampled based on the sum of the Euclidean distance and the feature distance. The sampling metric adopts a fusion metric of the distance metric and the feature metric. The calculation formula of the fusion metric is:

[0020] The second local feature is downsampled based on the sum of the Euclidean distance and the feature distance. The sampling metric adopts a fusion metric of the distance metric and the feature metric. The calculation formula of the fusion metric is: E(Z a , Z b )=λLd(Z a , Z b )+L f (Z a , Z b ), where Z a For the currently selected point cloud, Z b is the point cloud to be sampled, L d (Z a , Z b ) represents the distance metric between two point clouds, L f (Z a , Z b ) represents the feature measure between two point clouds, and λ is the weighting coefficient, which is 1.

[0021] Optionally, after continuously interpolating and up-sampling the global feature and the first local feature, the second local feature, and the third local feature, a multi-layer mixed feature is obtained through a skip link, and a final one-dimensional credibility is obtained through a maximum pooling operation, and finally a target event point is obtained by point-by-point segmentation of the final one-dimensional credibility, including:

[0022] After performing point cloud upsampling on the global feature, the number of points of the global feature is made the same as the number of points of the third local feature, and the two are jump-linked, and the multi-layer perceptron learning is called to obtain a first mixed feature;

[0023] After the first mixed feature is upsampled on the point cloud, the number of points of the first mixed feature is made the same as the number of points of the second local feature, and the two are jump-linked, and the multi-layer perceptron learning is called to obtain the second mixed feature;

[0024] After the second mixed feature is upsampled on the point cloud, the number of the second mixed feature points is made the same as the number of the event point cloud points, and the two are jump-linked, and the multi-layer perceptron learning is called to obtain the third mixed feature.

[0025] The third mixed feature performs the maximum pooling operation to obtain the final one-dimensional credibility of the event point cloud;

[0026] The final one-dimensional credibility is segmented point by point to obtain the target event point.

[0027] The present invention also provides a moving small target event stream segmentation device based on multi-scale point cloud feature fusion, comprising:

[0028] Memory for storing computer programs;

[0029] A processor is used to implement the steps of the above-mentioned method for segmenting small moving target event stream based on multi-scale point cloud feature fusion when executing the computer program.

[0030] The present invention also provides a computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps of the above-mentioned method for segmenting small moving target event streams based on multi-scale point cloud feature fusion are implemented.

[0031] The present invention obtains event stream data output by a dynamic visual sensor, and divides the event stream data into a plurality of event point clouds at preset time intervals; selects any one of the event point clouds for farthest point sampling to form three types of first sub-event point clouds with different scales; performs feature learning on the first sub-event point cloud using a multi-layer perceptron, and then performs multi-scale feature splicing to form a first local feature; adopts a point cloud self-attention mechanism to strengthen the spatiotemporal features of the first local feature to form a second local feature; performs down-sampling based on a fusion weight on the second local feature to form two types of second sub-event point clouds with different scales; after the second sub-event point cloud completes feature learning in the multi-layer perceptron, the multi-scale feature splicing is performed to obtain a third local feature; performs integrated feature learning on the third local feature to form a global feature; performs continuous interpolation up-sampling according to the global feature, the third local feature, the second local feature and the event point cloud, obtains a multi-layer mixed feature through a jump link, performs a maximum pooling operation on the multi-layer mixed feature to obtain a final one-dimensional credibility, and performs point-by-point segmentation on the final one-dimensional credibility to obtain a target event point.

[0032] It can be seen that compared with the prior art that relies on prior information, first, the present invention does not rely on artificially designed texture, shape, color and other features. Compared with existing methods, it is more suitable for small target processing and achieves high-reliability segmentation of small moving targets in dynamic visual sensors that does not rely on prior information. Second, the present invention adopts multi-scale feature and multi-level feature fusion learning methods to establish an event point cloud deep neural network, deeply mine the local and global multi-dimensional spatiotemporal features between event points, and can effectively reduce sensor noise, background clutter and other interference. Third, the present invention is suitable for automatic target segmentation under a moving platform, and does not require registration preprocessing. The event point cloud network automatically learns the motion difference between the target and the background under the moving platform, which simplifies the processing flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0034] Figure 1 A schematic diagram illustrating the steps of a method for segmenting a small moving target event stream based on multi-scale point cloud feature fusion provided by an embodiment of the present invention;

[0035] Figure 2 A schematic diagram of the operation flow of a method for segmenting a small moving target event stream based on multi-scale point cloud feature fusion provided by an embodiment of the present invention;

[0036] Figure 3 A schematic diagram of data flow of a method for segmenting a small moving target event stream based on multi-scale point cloud feature fusion provided by an embodiment of the present invention;

[0037] Figure 4 Event point cloud data diagrams of 5 scenes provided in the embodiments of the present invention;

[0038] Figure 5 A segmentation result diagram of the deep learning method provided by an embodiment of the present invention;

[0039] Figure 6 A segmentation result diagram of a moving small target event stream segmentation method based on multi-scale point cloud feature fusion provided by an embodiment of the present invention;

[0040] Figure 7 A schematic diagram of the structure of a small moving target event stream segmentation device based on multi-scale point cloud feature fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0042] Please refer to Figure 1 and Figure 2 , Figure 1 A schematic diagram illustrating the steps of a method for segmenting a small moving target event stream based on multi-scale point cloud feature fusion provided by an embodiment of the present invention, Figure 2 A schematic diagram of the operation flow of a method for segmenting a small moving target event stream based on multi-scale point cloud feature fusion provided in an embodiment of the present invention.

[0043] according to Figure 2 The specific operation process of the embodiment of the present invention is as follows:

[0044] The first step is to obtain event stream data containing small targets from the dynamic visual sensor, and divide the event stream data into several event point clouds at a time interval (T0 = 3ms). It should be noted that this embodiment does not limit the method of dividing the event stream data. For example, the event stream data can be divided at a time interval of 3ms; or the event stream data can be divided at a time interval of 4ms; or the event stream data can be divided at a time interval of 5ms. The value of the time interval can be selected according to the actual situation.

[0045] The second step is to select one of the event point clouds and save it in matrix form, which is called the original point cloud and is recorded as matrix D Org ∈R N×3 , N is the number of points in the original point cloud, and 3 is the dimensional features of the original point cloud, namely the row coordinates, column coordinates and timestamp of the points.

[0046] The third step is to analyze the original point cloud D Org ∈R N×3 According to the farthest point sampling, multi-scale segmentation is performed to obtain three subsets of the original event point cloud at different scales, collectively referred to as the first sub-event point cloud. The three different scales are 32 points, 64 points, and 128 points respectively.

[0047] The farthest point sampling is a downsampling method. The farthest point sampling operation randomly selects a point in the original point cloud as the initial sampling point, calculates the distance between each point in the remaining unsampled points and all the points in the sampled point set, selects the point with the largest sum of distances to add to the sampling point set, and iterates until a set number of points are collected. In this step, three downsampling scales of 32 points, 64 points, and 128 points are set to divide the original point cloud into several original event point cloud subsets (i.e., the first sub-event point cloud) of the three scales of 32 points, 64 points, and 128 points.

[0048] In the fourth step, for the first sub-event point cloud, a multi-layer perceptron is constructed according to the corresponding scale for feature learning, and the learned features of different scales are spliced ​​to obtain the first local feature, which is recorded as the matrix

[0049] Multilayer Perceptron (MLP) is a type of artificial neural network (ANN). In addition to the input and output layers, it can have multiple hidden layers. The simplest MLP contains only one hidden layer, that is, a three-layer structure. It can be understood that this step establishes a multilayer perceptron for the first sub-event point cloud of three scales: 32 points, 64 points, and 128 points. It can learn the features of the first sub-event point cloud to obtain three different scale features, and then splice them to obtain the first local feature. N1 is the number of points in the first local feature, and C1 is the point cloud dimensional feature in the first local feature, which is 320 in this embodiment.

[0050] In the fifth step, the point cloud self-attention mechanism is called to perform position-sensitive feature learning on the first-layer local features, further extract the deep spatiotemporal structural features, and obtain the second local features, which are recorded as the matrix

[0051] The specific process is to obtain the weight matrix through point cloud neural network training According to Q×D Loc1 ×W q , K=D Loc1 ×W k 、V=D Loc1 ×W v , get the self-attention matrix and And finally calculate the feature matrix That is, the second local feature. This step can effectively improve the representation of foreground target points and background event points in the event point cloud, and has a positive effect on retaining target event points in the segmentation task. Among them, N2 is the number of points in the second local feature, N2 = N1, and the point cloud self-attention mechanism does not change the number of points and dimensional characteristics of the second local feature. c2 is the dimension of the hidden layer in the point cloud neural network training.

[0052] In the sixth step, the event points of the second local feature are downsampled based on the fusion weight to form two local feature event point subsets of different scales, which are called the second sub-event point clouds. Among them, the event point downsampling based on the fusion weight is a downsampling based on the sum of the Euclidean distance and the feature distance. The sampling metric uses a fusion metric of the distance metric and the feature metric.

[0053] The weighted event point downsampling randomly selects a point in the second local feature as the initial sampling point, calculates the fusion metric between each point in the remaining unsampled points and all points in the sampled point set, selects the point with the largest sum of fusion metrics and adds it to the sampling point set, and iterates until the set number of points is collected. In this step, two downsampling scales of 64 points and 128 points are set. The second local feature is divided into the second event point cloud of 64 points and 128 points by weighted event point downsampling. The calculation formula of the fusion metric is: E(Z a , Z b )=λL d (Z a , Z b )+L f (Z a , Z b ), where Z a For the currently selected point cloud, Z b is the point cloud to be sampled, L d (Z a , Z b ) represents the distance metric between two point clouds, L f (Z a , Z b ) represents the feature measure between two point clouds, and λ is the weighting coefficient, which is 1.

[0054] In the seventh step, for the second sub-event point cloud, a multi-layer perceptron is constructed according to the corresponding scale for feature learning, and the learned features of different scales are spliced ​​to obtain the third local feature, which is recorded as the matrix N3 is the number of points in the third local feature, and C3 is the point cloud dimensional feature in the third local feature, which is 512 in this embodiment.

[0055] In this step, a multi-layer perceptron for the second sub-event point cloud with two scales of 64 points and 128 points is established. After feature learning of the second sub-event point cloud, two features of different scales are obtained, and then the third local feature is obtained by splicing.

[0056] The eighth step is to perform integrated feature learning on the third local feature, that is, to call the multi-layer perceptron to learn the features of all event points and obtain the global feature.

[0057] The ninth step is to upsample the global features and record them as matrix After upsampling, the number of global feature points is the same as the number of the third local feature points N3, C top is the point cloud dimension feature in the global feature, which is 1024 in this embodiment; a jump link is used to concatenate the global feature with the third local feature, and after calling the multi-layer perceptron for feature learning, the first mixed feature is obtained, which is recorded as a matrix C4 is the point cloud dimension feature in the first mixed feature, which is 256 in this embodiment.

[0058] In the tenth step, the first mixed feature is upsampled. After upsampling, the number of the first mixed feature points is the same as the number of the second local feature event points N2, and the two are spliced ​​by using jump links. Then, the multi-layer perceptron is called for feature learning to obtain the second mixed feature, which is recorded as the matrix C5 is the point cloud dimension feature of the second mixed feature, which is 128 in this embodiment.

[0059] In the eleventh step, the second mixed feature is upsampled. After upsampling, the number of the second mixed feature points is the same as the number of original point cloud points N, and the two are concatenated using skip links. Then, the multilayer perceptron is called for feature learning to obtain the third mixed feature, which is recorded as the matrix C6 is the point cloud dimension feature of the third mixed feature, which is 128 in this embodiment.

[0060] In the twelfth step, the third mixed feature is subjected to multi-dimensional feature maximum pooling to obtain the final one-dimensional credibility, which is recorded as matrix D final ∈R N×1 .

[0061] In the thirteenth step, point-by-point segmentation is performed based on the final one-dimensional credibility.

[0062] All event point clouds segmented in the first step are processed from the second step to the thirteenth step to obtain several point-by-point segmentation results. Figure 3 As shown, Figure 3 A schematic diagram of the data flow of the method for segmenting a small moving target event stream based on multi-scale point cloud feature fusion provided by an embodiment of the present invention, that is, a schematic diagram of the data flow from the second step to the thirteenth step, Figure 3The MLP is a multi-layer perceptron. stands for feature concatenation, and SA stands for point cloud self-attention mechanism.

[0063] In the fourteenth step, several point-by-point segmentation results are combined to obtain the segmentation results of small targets in the event stream data.

[0064] In order to verify the effect of this embodiment, this embodiment collects event stream data of small targets in 4 real scenes, and also collects a set of simulation data, a total of 5 sections of event stream data. The first section of event stream data is the indoor scene of the moving platform, the second section of event stream data is the outdoor scene of the moving platform, the third section of event stream data is the scene of normal lighting on the stationary platform, the fourth section of event stream data is the scene of insufficient lighting on the stationary platform, and the fifth section of event stream data is simulation data. The original point cloud data is as follows: Figure 4 As shown, Figure 4 The event point cloud data diagrams of the five scenes provided in the embodiment of the present invention are shown in Figure 1. The five event stream data are processed by the existing deep learning method and the method of this embodiment, and the small target segmentation result diagrams obtained are shown in Figure 2. Figure 5 and Figure 6 As shown, Figure 5 A segmentation result diagram of the deep learning method provided by an embodiment of the present invention, Figure 6 This is a diagram of the segmentation results of the small moving target event stream segmentation method based on multi-scale point cloud feature fusion provided in an embodiment of the present invention.

[0065] Figure 4 In the figure, a1 is the indoor scene of the moving platform, a2 is the outdoor scene of the moving platform, a3 is the normal lighting scene of the stationary platform, a4 is the insufficient lighting scene of the stationary platform, and a5 is the simulation data. Figure 5 b1~b5 are the segmentation results of the existing deep learning method. Figure 6 c1~c4, d5 are the segmentation results of the method of this embodiment. Figure 5 It can be seen that although the existing deep learning methods can perform segmentation, the segmentation results all contain a large number of false alarm points, especially in scenes with insufficient lighting, where the number of targets accounts for a small proportion. Figure 6 It is obvious that there are fewer false alarm points, and the suppression effect on background event points and noise event points is better. The segmentation effect in each scene is optimal.

[0066] The following is an introduction to a moving small target event stream segmentation device based on multi-scale point cloud feature fusion provided in an embodiment of the present invention. The device and the moving small target event stream segmentation method based on multi-scale point cloud feature fusion described above can correspond to each other.

[0067] Please refer to Figure 7 , Figure 7A schematic diagram of a structure of a moving small target event stream segmentation device based on multi-scale point cloud feature fusion provided by an embodiment of the present invention may include:

[0068] A memory 10, used for storing computer programs;

[0069] The processor 20 is used to execute a computer program to implement the above-mentioned small moving target event stream segmentation method based on multi-scale point cloud feature fusion.

[0070] The memory 10 , the processor 20 , and the communication interface 31 all communicate with each other via the communication bus 32 .

[0071] In the embodiment of the present invention, the memory 10 is used to store one or more programs, and the program may include program code, and the program code includes computer operation instructions.

[0072] In a possible implementation, the memory 10 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function, etc.; the data storage area may store data created during use.

[0073] In addition, the memory 10 may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include an NVRAM. The memory stores an operating system and operating instructions, executable modules or data structures, or a subset thereof, or an extended set thereof, wherein the operating instructions may include various operating instructions for implementing various operations. The operating system may include various system programs for implementing various basic tasks and processing hardware-based tasks.

[0074] The processor 20 may be a central processing unit (CPU), an application specific integrated circuit, a digital signal processor, a field programmable gate array or other programmable logic device, a microprocessor or any conventional processor, etc. The processor 20 may call a program stored in the memory 10 .

[0075] The communication interface 31 may be an interface of a communication module, and is used to connect to other devices or systems.

[0076] Of course, it should be noted that Figure 7 The structure shown does not constitute a limitation on the moving small target event stream segmentation device based on multi-scale point cloud feature fusion in the embodiment of the present application. In actual applications, the moving small target event stream segmentation device based on multi-scale point cloud feature fusion may include: Figure 5 More or fewer components than shown, or combinations of certain components.

[0077] The computer-readable storage medium provided in an embodiment of the present invention is introduced below. The computer-readable storage medium described below and the method for segmenting a small moving target event stream based on multi-scale point cloud feature fusion described above can be referred to each other.

[0078] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the above-mentioned method for segmenting small moving target event streams based on multi-scale point cloud feature fusion are implemented.

[0079] The computer-readable storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0080] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0081] Finally, it should be noted that, in this article, relationships such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0082] The above is a detailed introduction to the moving small target event stream segmentation based on multi-scale point cloud feature fusion provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A moving small target event stream segmentation method based on multi-scale point cloud feature fusion, characterized in that: include: Acquire event stream data output by a dynamic vision sensor, and divide the event stream data into a plurality of event point clouds at preset time intervals; Select any one of the event point clouds for farthest point sampling to form three types of first sub-event point clouds with different scales, use a multi-layer perceptron to perform feature learning on the first sub-event point cloud, and then perform multi-scale feature splicing to form a first local feature; Obtained through point cloud neural network training , , Three weight matrices, through the formula , , , obtain the self-attention matrices Q, K and V, and pass The second local feature is calculated; wherein, is the first local feature, is the normalized exponential function, is the dimension of the hidden layer in the point cloud neural network training; Downsampling the second local feature based on the fusion weight to form two types of second sub-event point clouds with different scales, and after the second sub-event point clouds complete feature learning in the multi-layer perceptron, performing the multi-scale feature splicing to obtain a third local feature; Performing integrated feature learning on the third local feature to form a global feature; After continuous interpolation upsampling according to the global feature, the third local feature, the second local feature and the event point cloud, a multi-layer mixed feature is obtained through a skip link, and a maximum pooling operation is performed on the multi-layer mixed feature to obtain a final one-dimensional credibility, and the final one-dimensional credibility is segmented point by point to obtain a target event point; Among them, after continuously interpolating and up-sampling the global feature and the first local feature, the second local feature and the third local feature, a multi-layer mixed feature is obtained through a skip link, and a final one-dimensional credibility is obtained through a maximum pooling operation, and finally the target event point is obtained by point-by-point segmentation of the final one-dimensional credibility, including: After performing point cloud upsampling on the global feature, the number of points of the global feature is made the same as the number of points of the third local feature, and the two are jump-linked, and the multi-layer perceptron learning is called to obtain a first mixed feature; After the first mixed feature is upsampled on the point cloud, the number of points of the first mixed feature is made the same as the number of points of the second local feature, and the two are jump-linked, and the multi-layer perceptron learning is called to obtain the second mixed feature; After the second mixed feature is upsampled on the point cloud, the number of points of the second mixed feature is made the same as the number of points of the event point cloud, and the two are jump-linked, and the multi-layer perceptron learning is called to obtain a third mixed feature; The third mixed feature performs the maximum pooling operation to obtain the final one-dimensional credibility of the event point cloud; The final one-dimensional credibility is segmented point by point to obtain the target event point.

2. The method for segmenting small moving target event stream based on multi-scale point cloud feature fusion according to claim 1 is characterized in that: The selecting any one of the event point clouds to perform farthest point sampling to form three types of first sub-event point clouds with different scales includes: Any one of the event point clouds is selected to perform the farthest point sampling to form a plurality of the first sub-event point clouds at 32 scales, 64 scales and 128 scales.

3. The method for segmenting small moving target event stream based on multi-scale point cloud feature fusion according to claim 1, characterized in that: The downsampling of the second local feature based on the fusion weight to form two types of second sub-event point clouds with different scales includes: The second local feature is downsampled based on the fusion weight to form the second sub-event point cloud including a 64-scale and a 128-scale.

4. The method for segmenting small moving target event stream based on multi-scale point cloud feature fusion according to claim 3 is characterized in that: The performing the downsampling based on the fusion weight on the second local feature comprises: The second local feature is downsampled based on the sum of the Euclidean distance and the feature distance. The sampling metric adopts a fusion metric of the distance metric and the feature metric. The calculation formula of the fusion metric is: ,in For the currently selected point cloud, is the point cloud to be sampled, Represents the distance metric between two point clouds, Represents the feature metric between two point clouds, is the weighting coefficient, which is 1.

5. A moving small target event stream segmentation device based on multi-scale point cloud feature fusion, characterized in that: include: Memory for storing computer programs; A processor is used to implement the steps of the method for segmenting small moving target event stream based on multi-scale point cloud feature fusion as described in any one of claims 1 to 4 when executing the computer program.

6. A computer-readable storage medium, characterized in that: The computer readable storage A computer program is stored on the medium, and when the computer program is executed by the processor, the steps of the method for segmenting a moving small target event stream based on multi-scale point cloud feature fusion as described in any one of claims 1 to 4 are implemented.

Images