Substation three-dimensional point cloud instance segmentation method
By introducing the element view enhancement module, differential 3D-UNET network, and instance fine module in the three-dimensional point cloud method segmentation method of the substation, solve the problem of weakness and signal-to-noise ratio of vertical dimensions in power scenarios, and the signal-to-noise ratio is too low. Efficient three -dimensional point cloud practical division effect.
Patent Information
- Application Number
- CN202410991353.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-23
- Publication Date
- 2025-01-21
AI Technical Summary
When the prior art deals with the segmentation of three-dimensional point cloud instances in electric power scenarios, the vertical dimension feature expression ability is weak, resulting in limited perception ability of the model in three-dimensional space. At the same time, the target in the real scene is small and the signal-to-noise ratio is too low, resulting in The model recognition segmentation effect is poor.
This paper proposes a 3D point cloud instance segmentation method for substations, including a voxel view enhancement module, a differential 3D-Unet network and an instance refinement module. The voxel view enhancement module enhances voxel features by introducing full view and attention mechanisms; the differential 3D-Unet network improves edge feature extraction capabilities through the cross-scale neighborhood semantic fusion module and semantic query module; the instance fine module uses clustering and Segmentation processing to obtain the final segmentation result.
Effectively enhance the perception ability of the model in the three -dimensional space, improve the edge feature extraction capacity of the power component, and improve the accuracy and efficiency of the instance segmentation. The experimental results show that the average accuracy of the Diffsegnet has increased by 3.03%compared to the mainstream SoftGroup model.
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of deep learning and point cloud instance segmentation, and specifically relates to a substation three-dimensional point cloud instance segmentation method. Background Art
[0002] Due to the large number of substation clouds collected, in order to reduce the amount of calculation, the point cloud encoding part usually adopts overhead columnar encoding. However, the elevation difference of each component in the power scene is large, and the components are often assembled vertically. This encoding method mainly focuses on the feature information in the horizontal direction, resulting in weak vertical dimension feature expression ability, especially when dealing with targets with complex vertical structures in the power scene, which will limit the model's perception ability in three-dimensional space. In addition, in real power scenes, there will be situations where the target is small and the signal-to-noise ratio is too low, and too much background noise will have an adverse effect on model recognition and segmentation.
[0003] In recent years, with the development of deep learning, segmentation methods based on neural networks have shown powerful performance in point cloud processing tasks and have gradually attracted attention from the industry. This type of method usually includes three steps, namely the pre-coding stage, the semantic segmentation stage, and the instance segmentation stage. In the pre-coding stage, in order to improve the signal-to-noise ratio of the point cloud and reduce the computational complexity, multimodal fusion encoding and other methods are adopted. However, the high deployment cost of multimodal cases in real scenarios limits their large-scale promotion. In the segmentation stage, since performing semantic segmentation and instance segmentation step by step may cause information islands and have an adverse effect on the segmentation task, some studies have performed the two simultaneously. Summary of the invention
[0004] In view of the above problems, the present invention provides a substation three-dimensional point cloud instance segmentation method.
[0005] To achieve the above object, the present invention adopts the following technical solutions:
[0006] A substation three-dimensional point cloud instance segmentation method comprises the following steps:
[0007] Step 1, construct a voxel view enhancement module, input the original point cloud, and output the voxel features after view enhancement;
[0008] Step 2: construct a differential 3D-Unet network, input the enhanced voxel features, and output the instance offset vector and semantic confidence score of each voxel feature to obtain preliminary instance segmentation results and semantic categories;
[0009] Step 3: construct an instance refinement module to cluster and segment the instance segmentation results output by the differential 3D-Unet network to obtain visible output results.
[0010] Furthermore, in step 1, a voxel view enhancement module is constructed, specifically:
[0011] Step 1.1, regularization: point cloud encoding based on voxel view enhancement;
[0012] Step 1.2, visualization: introduce a full view based on the observation center;
[0013] Step 1.3, view attention enhancement: The attention mechanism is used to enhance the semantic features of each view, that is, the enhanced voxel features.
[0014] Furthermore, in step 2, a differential 3D-Unet network is constructed, specifically:
[0015] An improved 3D-Unet network is used as the backbone network. A cross-scale neighborhood semantic fusion module and a semantic query module are proposed to enhance the edge resolution ability of semantics and instance masks. The network is divided into two parts: an encoder and a decoder. The encoder is responsible for encoding and pooling features, and the decoder is responsible for decoding and preliminarily segmenting instances. The encoder and decoder each have four layers to fuse features at different scales.
[0016] Furthermore, in step 3, the instance refinement module is constructed, specifically:
[0017] Step 3.1, each voxel is moved according to the offset vector to obtain a series of clustered voxel piles;
[0018] Step 3.2, under each semantics, the obtained voxel piles are clustered using the DBSCAN algorithm to obtain several types of piles;
[0019] Step 3.3, assign the same instance score matrix to all points in each obtained pile;
[0020] Step 3.4, segment the instance according to the instance score matrix to obtain the final segmentation result.
[0021] Compared with the prior art, the present invention has the following advantages:
[0022] 1) To address the problems of large randomness of feature distribution and low feature signal-to-noise ratio in substation scenarios, a voxel view enhancement module is proposed. By establishing a bird's-eye view columnar coding map and a center-radial columnar coding map, and using an attention mechanism to score feature fusion and enhance voxels, the problem of uneven feature distribution is alleviated.
[0023] 2) In order to solve the problem that the edge features of power components are complex and difficult to extract, a differential fusion module (cross-scale neighborhood semantic fusion module) and a semantic query module are designed to extract and fuse the difference features between adjacent abstraction layers and improve the edge feature extraction capability.
[0024] 3) Experiments on self-built power scenario datasets and public data show that the average accuracy of DiffSegNet is 3.03% higher than that of the mainstream SoftGroup model, which illustrates the effectiveness of the proposed model. DETAILED DESCRIPTION
[0025] In order to further illustrate the technical solution of the present invention, the present invention is further described below through embodiments.
[0026] A substation three-dimensional point cloud instance segmentation method comprises the following steps:
[0027] Step 1: Construct a voxel view enhancement module:
[0028] In the traditional point cloud instance segmentation model, due to the large number of collected point clouds, in order to reduce the amount of calculation, the point cloud encoding part usually adopts overhead columnar encoding. However, in the power scene, the elevation difference of each component is large, and the components are often assembled vertically. This encoding method mainly focuses on the feature information in the horizontal direction, resulting in weak vertical dimension feature expression ability, especially when dealing with targets with complex vertical structures in the power scene, which will limit the model's perception ability in three-dimensional space. In addition, in real power scenes, there will be situations where the target is small and the signal-to-noise ratio is too low, and too much background noise will have an adverse effect on model recognition and segmentation.
[0029] In order to solve the above problems, the present invention proposes a point cloud encoding method based on voxel view enhancement, which enriches the feature expression of the model in the vertical direction by introducing a full view based on the observation center. At the same time, in order to highlight the main information, the attention mechanism is used to enhance the semantic features under each view, so that the model can better identify the component information. The voxel view enhancement module inputs the original point cloud and outputs the voxel features after view enhancement. The module can be divided into three steps: regularization, visualization, and view attention enhancement.
[0030] Step 2: Construct a differential 3D-Unet network:
[0031] 3D-Unet networks fuse high- and low-level features by abstracting and connecting concrete features layer by layer, and are often used in instance segmentation models. However, the traditional 3D-Unet network uses simple center cropping as a feature distillation method, which causes the network to ignore some important boundary features and makes it difficult to express the difference between two adjacent layers of features, thus affecting the segmentation ability of the model. Therefore, this paper proposes an improved differential 3D-Unet network for processing segmentation features.
[0032] The differential 3D-Unet network inputs the enhanced voxel features and outputs the instance offset vector and semantic confidence score of each voxel feature to obtain preliminary instance segmentation results and semantic categories. This module uses the improved 3D-Unet network as the backbone network, and proposes a cross-scale neighborhood semantic fusion module and a semantic query module to enhance the edge resolution of semantics and instance masks. The network is divided into two parts: the encoder and decoder. The encoder is responsible for encoding and pooling features, and the decoder is responsible for decoding and preliminary instance segmentation. The encoder and decoder each have four layers to fuse features at different scales.
[0033] Step 3: Construct instance fine module:
[0034] The instance refinement module is responsible for clustering and segmenting the preliminary segmentation results output by the 3D-Unet network to obtain visible output results. First, the instance refinement module moves each voxel according to the offset vector to obtain a series of clustered voxel piles. Then, under each semantic, the DBSCAN algorithm is used to cluster the obtained voxel piles to obtain several types of piles. Then, all points in each obtained pile are assigned the same instance score matrix. Finally, the instance is segmented according to the instance score matrix to obtain the final segmentation result.
Claims
1. A substation three-dimensional point cloud instance segmentation method, characterized in that: The following steps are involved: Step 1, construct a voxel view enhancement module, input the original point cloud, and output the voxel features after view enhancement; Step 2: construct a differential 3D-Unet network, input the enhanced voxel features, and output the instance offset vector and semantic confidence score of each voxel feature to obtain preliminary instance segmentation results and semantic categories; Step 3: construct an instance refinement module to cluster and segment the instance segmentation results output by the differential 3D-Unet network to obtain visible output results.
2. A substation three-dimensional point cloud instance segmentation method according to claim 1, characterized in that: The voxel view enhancement module is constructed in step 1, specifically: Step 1.1, regularization: point cloud encoding based on voxel view enhancement; Step 1.2, visualization: introduce a full view based on the observation center; Step 1.3, view attention enhancement: The attention mechanism is used to enhance the semantic features of each view, that is, the enhanced voxel features.
3. A substation three-dimensional point cloud instance segmentation method according to claim 1, characterized in that: In step 2, a differential 3D-Unet network is constructed, specifically: An improved 3D-Unet network is used as the backbone network. A cross-scale neighborhood semantic fusion module and a semantic query module are proposed to enhance the edge resolution ability of semantics and instance masks. The network is divided into two parts: an encoder and a decoder. The encoder is responsible for encoding and pooling features, and the decoder is responsible for decoding and preliminarily segmenting instances. The encoder and decoder each have four layers to fuse features at different scales.
4. A substation three-dimensional point cloud instance segmentation method according to claim 1, characterized in that: In step 3, the instance refinement module is constructed, specifically: Step 3.1, each voxel is moved according to the offset vector to obtain a series of clustered voxel piles; Step 3.2, under each semantics, the obtained voxel piles are clustered using the DBSCAN algorithm to obtain several types of piles; Step 3.3, assign the same instance score matrix to all points in each obtained pile; Step 3.4, segment the instance according to the instance score matrix to obtain the final segmentation result.