Three-dimensional point cloud scene change detection method and system based on difference perception

Abstract

The present application relates to the technical field of object change detection, and particularly relates to a three-dimensional point cloud scene change detection method and system based on difference perception, three-dimensional point cloud data of target objects at different time phases in the same region is acquired to form a data set; a difference guiding module is designed to extract local features of different scales of the three-dimensional point cloud data at the two time phases, then the difference features between the local features of the three-dimensional point cloud at the two time phases are calculated, and then the difference features are used to enhance the local features of different scales of the three-dimensional point cloud at the two time phases; a difference fusion module is designed, the local features of the difference guiding module are transmitted to the corresponding levels of the difference fusion module through a skip connection, the difference fusion module learns the difference information of different levels from multiple scales and performs difference feature fusion, and the fused difference features are input into two layers of MLP to obtain the change result of the three-dimensional point cloud at the two time phases, and the result is no change and target object removal / increase. The present application improves the three-dimensional target object change detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of object change detection technology, and in particular to a method and system for detecting changes in a three-dimensional point cloud scene based on difference perception. Background Technology

[0002] Buildings, road infrastructure, and advertising signs are essential facilities in both urban and rural areas, and their construction, demolition, and alteration are closely related to socio-economic development. Detecting changes in these structures is crucial for urbanization analysis, identifying illegal buildings, and disaster assessment.

[0003] Deep learning has achieved tremendous success in computer vision, becoming a hot research topic. Many deep learning-based methods for detecting changes in 2D object images have yielded promising results. However, deep learning-based point cloud change detection is still in its early stages. Currently, most deep learning-based laser point cloud change detection methods do not directly process the original point cloud but focus on digital surface models, often leading to information loss. Ku et al. proposed a deep learning-based twin graph convolutional network to handle point cloud change detection. Its approach involves transforming point cloud data into structured graph data and then embedding it into a graph convolutional network to detect changes. This method does not fully utilize the positional information of the point cloud data and is more like a change classification method. Therefore, Wang et al. proposed a 3D Change Detection Network (3DCDNet) to directly extract features from the original point cloud data and output change results. However, this method only uses high-level features of the two-dimensional point cloud as feature difference discrimination information, neglecting the detailed information contained in low-dimensional features, underutilizing multi-scale object difference features, and failing to adequately consider the problem of imbalanced sample classes. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies that only use high-level features of target objects in dual-time 3D point clouds as feature difference discrimination information, neglecting the detailed information contained in low-dimensional features, failing to fully utilize multi-scale difference information, and not fully considering the problem of sample class imbalance. This invention proposes a 3D point cloud scene change detection method based on difference perception to improve the accuracy of 3D target object change detection.

[0005] To achieve the above objectives, the technical solution adopted is:

[0006] A method for detecting changes in a 3D point cloud scene based on difference perception, comprising:

[0007] Acquire 3D point cloud data of target objects in the same region at different times to form a dataset;

[0008] The design of the difference guidance module extracts local features at different scales from dual-temporal 3D point cloud data, then calculates the difference features between the local features of dual-temporal 3D point cloud, and then uses these difference features to enhance the local features at different scales of dual-temporal 3D point cloud.

[0009] The differential fusion module is designed to pass the local features of the differential guidance module to the corresponding level of the differential fusion module through skip connections. The differential fusion module learns the differential information of different levels from multiple scales and performs differential feature fusion. The fused differential features are input into two MLP layers to obtain the dual-temporal 3D point cloud change results, which are unchanged and target object removal / addition.

[0010] Construct a multi-scale loss function and train the model by minimizing the multi-scale loss function using the dataset.

[0011] According to the differential perception-based 3D point cloud scene change detection method of the present invention, before the dual-temporal 3D point cloud data is input into the differential guidance module, the original dual-temporal 3D point cloud data X0 and X1 are first preprocessed by removing ground points and random sampling operations to obtain processed point clouds Y0 and Y1. Then, Y0 and Y1 are respectively input into a single layer MLP to obtain high-dimensional features. and

[0012] According to the 3D point cloud scene change detection method based on difference perception of the present invention, the difference guidance module is further defined as a four-layer difference guidance module (DGM). and The inputs are fed into four DGM layers to obtain differential feature-guided change features. Each DGM layer includes a Local Feature Aggregation Module (LFA), a Nearest Detection Module (NFD), a Multilayer Perceptron (MLP), an activation function sigmoid, matrix multiplication, matrix addition, and random downsampling operations. The output of the previous DGM layer is used as the input of the next DGM layer.

[0013] According to the 3D point cloud scene change detection method based on difference perception of the present invention, the structure of the first layer DGM is further as follows:

[0014] Will and The input is fed into the LFA module to obtain the features of each point and the aggregated features of its 16 neighboring points, resulting in the aggregated features processed by the LFA module. and

[0015] Next, and The NFD module is used to extract differential features and calculate the difference information between each point in the point cloud of each time phase and the nearest point in another time phase. and

[0016] Then and Inputting into the MLP yields the variation characteristics of each time phase. and Secondly, using sigmoid to and Convert to values ​​in the range of 0-1 to obtain weight information. and

[0017] Next, and respectively with and Multiply, then add to each separately and Differential-guided aggregation features were obtained. and

[0018] Finally, and Perform random downsampling to obtain features and

[0019] The output of the first layer DGM serves as the input to the second layer DGM, and so on, with the output feature of the fourth layer DGM being... and

[0020] According to the 3D point cloud scene change detection method based on difference perception of the present invention, the difference guidance module outputs features. and Before being fed into the difference fusion module, the data is first fed into an MLP layer to obtain high-level features. and

[0021] According to the 3D point cloud scene change detection method based on difference perception of the present invention, the difference fusion module is further defined as a four-layer difference fusion module (DFM), which will... and The inputs are fed into four DFM layers to obtain multi-scale difference features. Each DFM layer includes a multilayer perceptron (MLP), a nearest neighbor difference (NFD) module, and an upsampling operation. The output of the previous DFM layer and the output of the corresponding DGM layer are concatenated as the input of that DFM layer.

[0022] According to the difference-aware 3D point cloud scene change detection method of the present invention, the structure of the first layer DFM is further as follows:

[0023] Will and The results obtained by comparing with the corresponding layer DGM and Connect them together, and then input them into an MLP layer to obtain dimensionality-reduced features. and Then to and Upsampling is performed to obtain high-resolution features and

[0024] Next, and Inputting the differential features into the NFD module yields the differential features. and

[0025] Will and Inputting one layer of MLP yields dimensionality-reduced features

[0026] During the decoding stage, for the above branch, four different scales of difference features are obtained through four DFGs. and Then use bilinear interpolation to... and Sampled with If the numbers match, you get and

[0027] Will and Adding them together yields multi-scale difference fusion features.

[0028] Finally, the decoding stage Inputting two layers of MLP yields the change results for each time phase.

[0029] According to the difference-aware 3D point cloud scene change detection method of the present invention, the expression of the multi-scale loss function is further as follows:

[0030] L total =L0+L1+L2+L3+L4

[0031] Among them, L total It is a multi-scale loss function, with L0, L1, L2, L3, and L4 being loss functions at five scales.

[0032] According to the 3D point cloud scene change detection method based on difference perception of the present invention, the calculation process of L0, L1, L2, L3 and L4 is further as follows:

[0033] First, the multi-scale fusion difference features obtained from the upper branch of the decoder are... and four scale differences and Five transformation results were obtained by inputting the two layers of MLP respectively. and Similarly, we obtain the five transformation results of the lower branch of the decoder. and Then, based on the cross-entropy loss function, loss functions L0, L1, L2, L3 and L4 for five scales are obtained.

[0034] Furthermore, the present invention also includes a 3D point cloud scene change detection system based on difference perception, comprising:

[0035] The dataset construction module is used to acquire 3D point cloud data of target objects in the same region at different times to form a dataset.

[0036] The difference guidance module is used to extract local features at different scales of dual-temporal 3D point cloud data, then calculate the difference features between the local features of dual-temporal 3D point cloud, and then use the difference features to enhance the local features at different scales of dual-temporal 3D point cloud.

[0037] The difference fusion module is used to learn difference information at different levels from multiple scales and perform difference feature fusion. The fused difference features are input into a two-layer MLP to obtain the change results of the three-dimensional point cloud in two time phases. The results are unchanged and target object removal / addition.

[0038] The model training module is used to construct a multi-scale loss function and train the model by minimizing the multi-scale loss function using the dataset.

[0039] The beneficial effects achieved by adopting the above technical solution are:

[0040] To address the shortcomings of existing 3D object change detection methods, such as insufficient utilization of multi-scale difference features and low-dimensional features, and sample class imbalance, a difference-aware point cloud scene change detection network (DAPCCDNet) is proposed. It is a Siamese network with weight sharing, similar to the U-Net architecture, and mainly consists of a multi-layer perceptron (MLP), a difference guide module (DGM), and a difference fusion module (DFM). ① The MLP is mainly used to map low-dimensional features to high-dimensional features and obtain change results. ② The DGM acts as an encoder to extract coarse-to-fine local features from the two-temporal 3D point cloud. Specifically, the DGM is used to extract local features at different scales in the two-temporal 3D point cloud, then calculates the difference features between these local features, and uses these difference features to enhance the local features at different scales in the two-temporal 3D point cloud, improving the model's robustness. ③ The DFM acts as a decoder to extract fusion information of multi-scale change features. Specifically, DFM is used to perform four upsampling mappings on the features to obtain feature information at four levels. Then, the difference information between the features at each level is calculated to obtain the fused multi-scale difference features. Finally, the fused features are input into a two-layer MLP to obtain the dual-temporal 3D point cloud change results. Using multi-scale difference fusion features in the decoder stage enhances the model's feature representation ability at multiple scales, improving the target object change detection effect. ④ Using a multi-scale loss function alleviates the sample class imbalance problem, which to some extent improves the "overfitting" situation when the model is trained on a larger number of sample classes, and improves the detection accuracy of a smaller number of sample classes. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0042] Figure 1 This is a flowchart illustrating the method for detecting changes in a 3D point cloud scene based on difference perception, according to an embodiment of the present invention.

[0043] Figure 2 This is a network structure diagram of the 3D point cloud scene change detection method based on difference perception according to an embodiment of the present invention;

[0044] Figure 3 This is a network framework diagram of the differential guidance module according to an embodiment of the present invention;

[0045] Figure 4 This is a network framework diagram of the difference fusion module in an embodiment of the present invention; Detailed Implementation

[0046] The exemplary solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art.

[0047] To address the shortcomings of existing methods that only utilize high-level features of target objects in dual-temporal 3D point clouds as feature difference discrimination information, neglecting the detailed information contained in low-dimensional features, and failing to fully utilize multi-scale difference information and adequately consider the problem of sample class imbalance, this embodiment proposes a difference-aware 3D point cloud scene change detection method. Figure 1 and Figure 2 As shown, it includes the following steps:

[0048] Step S101: Obtain 3D point cloud data of target objects in the same region at different time phases to form a dataset. Labels are assigned to the dual-time phase 3D point cloud data in this dataset, with 0 representing no change and 1 representing change. Change has two meanings: target object removal or target object addition. Specifically, the label type for the earlier time phase 3D point cloud data is "no change" and "target object removal," while the label type for the later time phase 3D point cloud data is "no change" and "target object addition."

[0049] Step S102: Design a difference guidance module to extract local features at different scales of dual-temporal 3D point cloud data, then calculate the difference features between the local features of dual-temporal 3D point cloud, and then use these difference features to enhance the local features at different scales of dual-temporal 3D point cloud.

[0050] Step S103: Design a difference fusion module. Through skip connections, the local features of the difference guidance module are passed to the corresponding level of the difference fusion module. The difference fusion module learns the difference information of different levels from multiple scales and performs difference feature fusion to enhance the multi-scale expressive ability of the model. The fused difference features are input into two layers of MLP to obtain the dual-temporal 3D point cloud change results. The results are unchanged and target object removal / addition.

[0051] Step S104: Construct a multi-scale loss function and train the model by minimizing the multi-scale loss function using the dataset. This alleviates the "overfitting" problem to some extent when the model is trained on a large number of sample categories.

[0052] (1) Difference Guidance Module

[0053] With the rapid development of convolutional neural networks and point cloud change detection technology, more and more point cloud change detection networks are demonstrating powerful feature extraction capabilities. To extract change features of dual-time 3D point clouds of a target object from coarse to fine scale, we employ the Local Feature Aggregation (LFA) and Nearest Feature Difference (NFD) modules of 3DCDNet as the foundation of the DAPCCDNet feature extractor. The LFA module extracts point cloud features and aggregates neighborhood information, while the NFD module identifies the difference features between pairs of point clouds. The DGM guides the multi-scale feature extraction process and uses the difference features as prior knowledge to expand the receptive field between point clouds.

[0054] Given two temporal 3D point clouds Where P and Q represent the number of points. First, to improve computational efficiency, the original dual-temporal 3D point cloud data X0 and X1 are subjected to ground point removal and random sampling operations to obtain the processed point cloud. and Where N represents the number of processed points, and N is set to 8192. Then, to obtain richer and more representative feature representations, Y0 and Y1 are respectively input into a single MLP layer to obtain high-dimensional features. and This process can be represented as:

[0055]

[0056] in, and These are high-dimensional features processed by MLP, with d=64. Next, to fully utilize the differential features, [the following is done / then...] and The inputs are fed into the four-layer DGM to obtain the variation features guided by differential features, such as... Figure 3 As shown, each DGM layer includes LFA, NFD, MLP, Sigmoid, matrix multiplication, matrix addition, and random downsampling operations. The output of the previous DGM layer serves as the input to the next DGM layer.

[0057] For the first layer of DGM, in order to fully utilize the neighborhood information of each point, and The input is fed into the LFA module to obtain the features of each point and the aggregated features of its 16 neighboring points. This process can be represented as:

[0058]

[0059] in, and These are aggregated features after LFA operation, where b equals 128. Next, to obtain the difference features between each time phase and another time phase, the... and The NFD module calculates the difference information between each point in each temporal point cloud and the nearest point in another temporal phase. This process can be represented as follows:

[0060]

[0061] in, and These are the differential features extracted through NFD operations. Then, and Inputting the MLP yields the variation characteristics of each time phase; this process can be represented as:

[0062]

[0063] in, and These are the variation features extracted through MLP operations. Secondly, sigmoid is used to... and Converting values ​​to the range of 0-1 to obtain weight information can be represented as follows:

[0064]

[0065] in, and These are the weights after a Sigmoid transformation. They guide the subsequent feature extraction process during the encoding phase. Next, we will... and respectively with and Multiply, then add to each separately and The difference-guided aggregation features were obtained, and this process can be represented as:

[0066]

[0067] in, and This is a difference-guided aggregation feature. By calculating the difference features between two temporal 3D point clouds as prior information to guide multi-scale difference feature extraction, the expressive power of change features is further improved. Finally, in order to extract local features and enhance the perceptual ability of DGM, the following is used: and The random downsampling operation can be represented as follows:

[0068]

[0069] in, and The features are obtained by random downsampling, where M is the number of points after downsampling, M = 2048. The specific operations and feature input / output dimensions of different layers of the differential guidance module are shown in Table 1.

[0070] Table 1. Structural details of the differential guidance module

[0071]

[0072] (2) Difference Integration Module

[0073] To fully utilize multi-scale differential features, a DFM (Differential Factorization Model) was designed. This module extracts differential information from four scales and fuses them to obtain multi-scale differential fusion features. Then, the fused features are input into a two-layer MLP (Multi-Level Processing) to obtain the change results for each time phase.

[0074] First, in order to further improve the expressive power of the model, the results obtained during the encoding stage will be used to... and The process of inputting a single MLP layer to obtain high-level features can be represented as follows:

[0075]

[0076] in, and These are high-level features obtained through an MLP layer. Then, and The data are input into four layers of DFM to obtain multi-scale differential features. Each DFM layer includes an MLP module, an NFD module, and an upsampling operation. This module aims to learn differential information at different levels from multiple scales and perform differential feature fusion, thereby enhancing the model's feature representation ability at multiple scales and further improving the accuracy of point cloud change detection.

[0077] For the first layer of DFM, in order to alleviate gradient vanishing, accelerate network convergence, and improve the network's expressive power, the following steps are taken: and The results obtained by comparing with the corresponding layer DGM and Connecting them together and then inputting each into an MLP layer to obtain dimensionality-reduced features can be represented as follows:

[0078]

[0079] in, and These are dimensionality-reduced features obtained through an MLP layer. Then, to allow the network to capture more detailed information, [the following steps are taken]. and The upsampling operation can be represented as follows:

[0080]

[0081] in, and These are upsampled, high-resolution features. Next, ... and Inputting into the NFD module yields differential features. and This process can be represented as:

[0082]

[0083] in, and These are the differential features. Finally, to facilitate the subsequent fusion of multi-scale features, the differential features are transformed to a unified dimension. and Inputting a layer of MLP to obtain dimensionality-reduced features can be represented as:

[0084]

[0085] in, These are the difference features after unifying the dimensions. During the decoding stage, for the branch above, four different scales of difference features are obtained after four DFGs. Then, to facilitate subsequent feature fusion, bilinear interpolation is used to... and Sampled with When the points are the same, the process can be represented as:

[0086]

[0087] in, It is a feature of multi-scale differences in a unified dimension and number of points.

[0088] Upsampled multi-scale features contain multi-level change discrimination information, used to simultaneously capture changing targets at different scales. Next, these difference features are added together to obtain multi-scale difference fusion features; this process can be represented as:

[0089]

[0090] in, It is a multi-scale difference fusion feature. Finally, The process of inputting two layers of MLP to obtain the change results for each time phase can be represented as follows:

[0091]

[0092] in, It is a dimensionality reduction feature. This is the result of the change detection.

[0093] The specific operations and feature input / output dimensions of different layers of the difference fusion module are shown in Table 2.

[0094] Table 2 Structural details of the difference fusion module

[0095]

[0096] (3) Loss Function

[0097] As real-world point cloud datasets increase, the problem of class imbalance severely impacts the practical application of point cloud data. Specifically, minority classes (unchanged) constitute the majority of samples, while majority classes (removed and added) constitute a minority. This makes deep learning models prone to overfitting during training, resulting in difficulty improving class detection accuracy with a small number of samples. Therefore, to alleviate the class imbalance problem, the cross-entropy loss function is used as the basis for the loss function. The cross-entropy loss function is defined as:

[0098]

[0099] in, Represents the true label, y i This represents the predicted label, and N represents the number of point clouds.

[0100] To fully account for multi-scale difference information, a multi-scale loss function was designed. First, the multi-scale fused difference features obtained from the upper branch of the decoder are... and four scale differences and The process of inputting data into two MLP layers and obtaining five different results can be represented as follows:

[0101]

[0102] in, and It is a dimensionality reduction feature. and This is the change detection result. Similarly, we can obtain the five change results from the lower branch of the decoder. Next, based on the cross-entropy loss function, loss functions for five scales were obtained, and this process can be expressed as:

[0103]

[0104] in, It is a prediction result. These are the true labels. Finally, the five loss functions are summed to obtain the multi-scale loss function, which can be expressed as:

[0105] L total =L0+L1+L2+L3+L4

[0106] Among them, L total It is a multi-scale loss function, with L0, L1, L2, L3, and L4 being loss functions at five scales.

[0107] Corresponding to the above method, this embodiment also proposes a 3D point cloud scene change detection system based on difference perception, comprising:

[0108] The dataset construction module is used to acquire 3D point cloud data of target objects in the same region at different times to form a dataset.

[0109] The difference guidance module is used to extract local features at different scales of dual-temporal 3D point cloud data, then calculate the difference features between the local features of dual-temporal 3D point cloud, and then use these difference features to enhance the local features at different scales of dual-temporal 3D point cloud.

[0110] The difference fusion module is used to learn difference information at different levels from multiple scales and perform difference feature fusion. The fused difference features are input into a two-layer MLP to obtain the change results of the three-dimensional point cloud in two time phases. The results are unchanged and target object removal / addition.

[0111] The model training module is used to construct a multi-scale loss function and train the model by minimizing the multi-scale loss function using the dataset.

[0112] The invention will be further illustrated by the following experiments to provide a better understanding of it.

[0113] (I) Experiment Implementation Details

[0114] First, to improve computational efficiency and eliminate interference from incomplete ground points, ground points with a height below 0.5m were removed. Then, to ensure consistent size of the original point cloud when inputting into the network, random sampling was performed on the original point cloud, with 8192 points sampled. Next, an adaptive optimizer was used to update the network parameters, with the smoothing factor gamma, learning rate, learning step size, momentum, and weight decay set to 0.96, 0.0015, 1, 0.9, and 0.00035, respectively. The number of nearest neighbors in the NFD during the encoding phase was set to 1, while the number of nearest neighbors in the NFD during the decoding phase was set to 16. The specific input and output dimensions for each layer of DAPCCDNet are shown in Tables 1 and 2. All experiments were conducted using PyTorch 1.8.1 on an NVIDIA GeForce RTX 3090 graphics card with 24GB of VRAM.

[0115] (II) Evaluation Dataset and Evaluation Metrics

[0116] Evaluation Dataset. To compare with other change detection methods, the mainstream SLPCCD dataset was selected. This dataset contains 621 pairs of samples, with the training set, validation set, and test set consisting of 398, 95, and 128 pairs of samples, respectively. For this dataset, changed targets are divided into two categories: those that increase (points added compared to the previous point cloud) and those that decrease (points reduced compared to the subsequent point cloud). The number of points that increase, decrease, and remain unchanged are 1,298,600, 3,263,257, and 80,657,859, respectively.

[0117] Evaluation metrics. To verify the effectiveness of the network, this paper uses three metrics—crossover-union ratio (CUI), average precision (AP), and average CUI—to evaluate the experimental results.

[0118]

[0119] Where A and B represent the set of predicted results and the set of true values, respectively, and N c N represents the number of point clouds that are correctly classified. t The total number of point clouds is represented by s, and the number of categories is represented by s. These categories include those that remain unchanged, those that have been removed, and those that have been added.

[0120] (III) Evaluation results on the SLPCCD dataset

[0121] The change detection results were evaluated using IoU, mIoU, and OA on the SLPCCD dataset. The DAPCCDNet method of this invention was compared with PointNet++-based, DGCNN-based, and 3DCDNet methods. As shown in Table 3, the DAPCCDNet method achieved the best IoU, mIoU, and OA on the SLPCCD dataset. The DAPCCDNet method achieved IoU of 96.04%, 70.72%, and 68.73% for invariance, removal, and addition of categories, respectively, and its OA and mIoU were 96.37% and 78.50%, respectively. The high accuracy of the DAPCCDNet method can be attributed to three factors. First, in the encoding stage, the DAPCCDNet method utilizes multi-scale local features corresponding to bi-temporal point clouds to obtain multi-scale difference features, and then uses these difference features to guide the corresponding local features. These difference features-guided local features better reflect the differences between bi-temporal point clouds and are more robust. Second, in the decoding stage, the DAPCCDNet method obtains and fuses the difference information at four scales. This fused information enhances the model's feature representation capabilities across multiple scales, further improving the accuracy of point cloud change detection. Third, the DAPCCDNet method uses a multi-scale loss function to alleviate the problem of imbalanced sample classes, mitigating overfitting when the model is trained on a larger number of sample classes and improving detection accuracy on a smaller number of sample classes.

[0122] Table 3 Comparison of different methods (%)

[0123]

[0124] (iv) Ablation test

[0125] In this section, to analyze the effectiveness of the differential guidance module, differential fusion module, and multi-scale loss, 3DCDNet was selected as the baseline model. A differential guidance module was added to the baseline model, named DAPCCDNet-1; DAPCCDNet-2 added a differential fusion module to DAPCCDNet-1; and DAPCCDNet integrates the differential guidance module, differential fusion module, and multi-scale loss. Detailed settings for the ablation experiments are shown in Table 4. Comparison of DAPCCDNet-1 with the baseline model validated the effectiveness of the differential guidance module. Comparison of DAPCCDNet-2 with DAPCCDNet-1 validated the effectiveness of the differential fusion module. Comparison of DAPCCDNet with DAPCCDNet-2 validated the effectiveness of the multi-scale loss.

[0126] Table 4 Detailed setup of the ablation experiment

[0127]

[0128] The results of the ablation experiments are shown in Table 5. Comparing DAPCCDNet-1 with the baseline model, it was found that learning local features guided by differential features during the encoding stage can improve the accuracy of point cloud change detection. Comparing DAPCCDNet-2 with DAPCCDNet-1, it was found that learning multi-scale fusion differential information during the decoding stage can enhance the model's expressive ability at multiple scales, thereby improving the accuracy of point cloud change detection. Through a comparison of DAPCCDNet and DAPCCDNet-2, it was found that multi-scale loss can alleviate the problem of sample class imbalance, improve overfitting, and further improve the accuracy of point cloud change detection.

[0129] Table 5. Analysis results of the ablation experiment (%)

[0130]

[0131] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0132] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0133] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0134] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0135] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting changes in a 3D point cloud scene based on difference perception, characterized in that, Include: S1: Acquire 3D point cloud data of target objects in the same region at different times to form a dataset; S2: Design a difference guidance module to extract local features at different scales of dual-temporal 3D point cloud data, then calculate the difference features between the local features of dual-temporal 3D point cloud, and then use these difference features to enhance the local features at different scales of dual-temporal 3D point cloud. Before the dual-temporal 3D point cloud data is input into the difference guidance module, the original dual-temporal 3D point cloud data is first processed. and Preprocessing involves removing ground points and random sampling to obtain the processed point cloud. and Then and The features are input into a single MLP layer to obtain high-dimensional features. and The differential boot module is a four-layer differential boot module (DGM). and The inputs are fed into four DGM layers to obtain differential feature-guided change features. Each DGM layer includes a Local Feature Aggregation (LFA) module, a Nearest Detection (NFD) module, a Multilayer Perceptron (MLP) module, an activation function sigmoid, matrix multiplication, matrix addition, and random downsampling operations. The output of the previous DGM layer is used as the input of the next DGM layer. S3: Design a difference fusion module. Through skip connections, the local features of the difference guidance module are passed to the corresponding level of the difference fusion module. The difference fusion module learns the difference information of different levels from multiple scales and performs difference feature fusion. The fused difference features are input into two MLP layers to obtain the change results of the dual-temporal 3D point cloud. The results are: no change, removal or addition of target objects. Among them, the difference guidance module outputs features and Before being fed into the difference fusion module, the data is first fed into an MLP layer to obtain high-level features. and The difference fusion module is a four-layer difference fusion module (DFM), which will... and The inputs are fed into four DFM layers to obtain multi-scale difference features. Each DFM layer includes a multilayer perceptron (MLP), a nearest neighbor difference (NFD) module, and an upsampling operation. The output of the previous DFM layer and the output of the corresponding DGM layer are concatenated as the input of that DFM layer. S4: Construct a multi-scale loss function and train the model by minimizing the multi-scale loss function using the dataset.

2. The method for detecting changes in a 3D point cloud scene based on difference perception according to claim 1, characterized in that, The structure of the first layer of DGM is as follows: Will and The input is fed into the LFA module to obtain the features of each point and the aggregated features of its 16 neighboring points, resulting in the aggregated features processed by the LFA module. and ; Next, and The NFD module is used to extract differential features and calculate the difference information between each point in the point cloud of each time phase and the nearest point in another time phase. and ; Then and Inputting into the MLP yields the variation characteristics of each time phase. and ; Secondly, using sigmoid to and Convert to values ​​in the range of 0-1 to obtain weight information. and ; Next, and respectively with and Multiply, then add to each separately and Differential-guided aggregation features were obtained. and ; Finally, and Perform random downsampling to obtain features and ; The output of the first layer DGM serves as the input to the second layer DGM, and so on, with the output feature of the fourth layer DGM being... and .

3. The method for detecting changes in a 3D point cloud scene based on difference perception according to claim 1, characterized in that, The structure of the first layer DFM is as follows: Will and The results obtained by comparing with the corresponding layer DGM and Connect them together, and then input them into an MLP layer to obtain dimensionality-reduced features. and Then on and Upsampling is performed to obtain high-resolution features and ; Next, and Inputting the differential features into the NFD module yields the differential features. and ; Will and Inputting one layer of MLP yields dimensionality-reduced features ; During the decoding stage, for the above branch, four different scales of difference features are obtained through four DFGs. , , and Then use bilinear interpolation to... , and Sampled with If the numbers match, you get , and ; Will , , and Adding them together yields multi-scale difference fusion features. ; Finally, the decoding stage Inputting two layers of MLP yields the change results for each time phase. .

4. The method for detecting changes in a 3D point cloud scene based on difference perception according to claim 3, characterized in that, The expression for the multi-scale loss function is: ,in, It is a multi-scale loss function. , , and It is a loss function with five scales.

5. The method for detecting changes in a 3D point cloud scene based on difference perception according to claim 4, characterized in that, , , and The calculation process is as follows: First, the multi-scale fusion difference features obtained from the upper branch of the decoder are... and four scale differences , , and Five transformation results were obtained by inputting the two layers of MLP respectively. , , , Similarly, the five transformation results of the lower branch of the decoder are obtained. , , , and Then, based on the cross-entropy loss function, loss functions at five scales are obtained. , , and .

6. A 3D point cloud scene change detection system based on difference perception, characterized in that, For implementing the difference-aware 3D point cloud scene change detection method as described in any one of claims 1-5, the system comprises: The dataset construction module is used to acquire 3D point cloud data of target objects in the same region at different times to form a dataset. The difference guidance module is used to extract local features at different scales of dual-temporal 3D point cloud data, then calculate the difference features between the local features of dual-temporal 3D point cloud, and then use these difference features to enhance the local features at different scales of dual-temporal 3D point cloud. The difference fusion module is used to learn difference information at different levels from multiple scales and perform difference feature fusion. The fused difference features are input into two layers of MLP to obtain the change results of the two-temporal 3D point cloud, and the results are no change, removal or addition of target objects; The model training module is used to construct a multi-scale loss function and train the model by minimizing the multi-scale loss function using the dataset.

Images