A power equipment detection method based on 3D point cloud saliency learning

Abstract

The present invention discloses a method for detecting power equipment based on 3D point cloud saliency learning, which belongs to the field of deep learning and three-dimensional target detection technology. In view of the problems that background point interference and insufficient key feature extraction of the detection model may occur when detecting substation equipment due to the complexity of the detection target in the existing method, a saliency learning module is constructed in the feature extraction stage on the basis of the Second algorithm, and an enhanced channel attention module and a residual space attention module are designed to focus on the spatial position information of the point cloud, obtain the saliency features of the point cloud in the power scene, reduce the noise interference in the detection scene, and thus achieve the purpose of improving the detection accuracy of the model; the saliency features are sent to the sparse convolution module for feature extraction, and then enter the regional recommendation network module to construct a 3D bounding box as the loss function for overall regression, which further enhances the accuracy of the model prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and three-dimensional target detection, and specifically relates to a power equipment detection method based on 3D point cloud saliency learning. Background Art

[0002] Inspecting power equipment within substations is essential to ensure safe and stable operation. Currently, the main methods for inspecting substation equipment are manual inspection, robotic inspection, and image recognition technology based on video intelligent monitoring. However, manual inspection is inefficient and cannot guarantee accuracy. The latter two methods are limited by the difficulty of measuring distance using image data, making it impossible to accurately inspect substation equipment.

[0003] With the rapid development of computer vision, 3D object detection technology is becoming increasingly mature. To address the increasingly complex maintenance tasks in power grid scenarios, identifying key components using 3D object detection algorithms has become a mainstream trend. However, substation cloud scenarios are complex and the data volume is enormous. Existing methods for detecting substation equipment often encounter issues such as background interference from the numerous detection targets, which can hinder detection and insufficient extraction of key features by the detection model. Therefore, there is an urgent need to develop a substation equipment detection method that can reduce the impact of background interference in complex substation scenarios and capture key features within the scene to meet the needs of safe substation operation and maintenance. Summary of the Invention

[0004] In view of the problems that existing methods for detecting substation equipment may cause background point interference due to the complexity of detection targets and insufficient extraction of key features by the detection model, the present invention provides a power equipment detection method based on 3D point cloud saliency learning.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for detecting power equipment based on 3D point cloud saliency learning, the method comprising the following steps:

[0007] Step 1: Create a KITTI format dataset for power scenarios;

[0008] The specific operations of step 1 to establish the KITTI format dataset for the power scene are as follows: using a lidar to collect point cloud data, using a camera to collect RGB images, using a point cloud annotation tool to annotate the objects in the collected point cloud, and organizing them according to the KITTI dataset label format to obtain a self-built dataset for training and testing; the dataset is divided into a training set, a validation set, and a test set according to the proportion;

[0009] Step 2: Construct a SalSecond object detection model that includes a saliency learning module, a sparse convolution module, and a region recommendation network module; where:

[0010] The saliency learning module enhances channel attention while ensuring the flow of information between groups, and learns different channel features in groups to obtain more refined global features of the point cloud. Subsequently, through point-by-point spatial attention learning, it further obtains salient features in the scene and reduces the impact of background points on detection accuracy.

[0011] The sparse convolution module extracts the obtained salient features;

[0012] The region recommendation network module performs prediction regression by constructing the Widget Over Union (WIoU) loss function and focusing on the positional relationship between the prediction box and the true value box to improve the model regression accuracy. The specific operations are as follows:

[0013] Using the Second target detection model as the baseline model, a saliency learning module is constructed in the feature extraction stage. An enhanced channel attention module and a residual spatial attention module are designed to focus on the spatial location information of the point cloud, reduce noise interference in the detection scene, and obtain salient features of the point cloud in the power scene. These salient features are fed into the sparse convolution module for feature extraction. Subsequently, the WIoU loss is used in the region recommendation network module to construct a 3D bounding box as the overall regression loss function. This solves the problem that the original model's loss function ignores the coupling between different bounding box parameters, resulting in poor regression accuracy.

[0014] Furthermore, in step 2, the traditional target detection model Second is used as the baseline model. In the feature extraction stage, a saliency learning module is constructed, and an enhanced channel attention module (ECA) and a residual space attention module (RSA) are designed to focus on the spatial position information of the point cloud, reduce noise interference in the detection scene, and obtain the saliency features of the point cloud in the power scene. The specific operations are as follows:

[0015] The voxelized point cloud data is input into the network and processed by the enhanced channel attention module to obtain the global feature G. Then, it is processed by the residual space attention module to obtain the salient feature S of the point cloud. The process is as follows:

[0016] Step 2.1: The enhanced channel attention module divides the input channel into three non-interfering sub-networks, and inputs the voxelized point cloud data aggregated by the maximum pooling and average pooling operations separately, focusing on the fine-grained features of the point cloud from different dimensions; then, the Concat operation is performed on the fine-grained features of different dimensions for splicing; to solve the problem of insufficient global feature learning caused by the lack of feature interaction between different groups and reduced information expression ability due to separate calculations of each sub-network, the Shuffle operation is introduced to evenly mix the fine-grained features of different dimensions and then learn them again through the three sub-networks and splice the final results; finally, nonlinear mapping is performed through the activation function to obtain higher-dimensional global features; this process is described by the following formula:

[0017]

[0018] Among them, i represents the sub-networks of different groups, ω i is the parameter learned by the fully connected layer in each sub-network, W max +W avg represents the features of the voxelized point cloud after maximum pooling and average pooling, σ is the Sigmoid activation function; f i (x) represents the global feature obtained by concatenating the results of the first grouping;

[0019] The results obtained from each sub-network are then concatenated and a Shuffle operation is performed. The Shuffle operation is an operation that mixes different features. Specifically, the features calculated from different channels in the first half of the network are first reshaped, transposed, and then reshaped again. They are then evenly mixed, and finally learned again through three different channels. The learned features of different dimensions are concatenated to obtain the global feature. This can avoid information isolation caused by different groups of calculations and enhance the learning ability of data representation. The formula is as follows:

[0020]

[0021] W l Represents the parameters of different fully connected layers during the second grouping operation; σ represents the Sigmoid activation function;

[0022] Step 2.2: The residual spatial attention module adopts a self-attention structure to focus on the spatial distance of the point cloud and the correlation between shape features. It also introduces a residual connection mechanism to strengthen the relationship between input and output and improve the model's ability to capture salient features. Specifically:

[0023] First, the global features obtained by the enhanced channel attention module are mapped into three parameters Q, K, and V through three convolutional layers. Then, the relationship between different points is calculated using the transpose of Q and K. Finally, softmax is used for normalization to obtain the attention weight A and multiply it with V to obtain the salient feature. The process is described by the following formula:

[0024]

[0025] S=V×A+G (4)

[0026] Among them, S is the significant feature obtained, K T Represents matrix transpose, dk is the dimension of key value K;

[0027] Furthermore, in step 2, the region recommendation network module is used to construct the WIoU loss function for regression of the 3D bounding box as a whole. The specific operation to solve the problem that the original model loss function ignores the coupling between different parameters of the bounding box, resulting in poor regression accuracy, is as follows:

[0028] Step 2.3: Construct the WIoU loss function for regression of the 3D bounding box as a whole to solve the problem that the original model loss function ignores the coupling between different parameters of the bounding box, resulting in poor regression accuracy; this process is expressed as the following formula:

[0029] WIoU=L DIoU3D +αv (5)

[0030] Among them, L DIoU3D is the DIoU in a three-dimensional scene, and its calculation formula is as follows:

[0031] L DIoU3D =1+A DIoU3D (B p ,B g ) (6)

[0032]

[0033] Among them, B g is the true target box, B p To predict the bounding box, V p is the area of ​​the prediction box, V g is the area of ​​the true value box, h i For B p With B g The intersection height, R i For B p With B g The volume of the intersection area, b is B p With B gThe distance between the center points, d is the length of the diagonal of the minimum enclosed rectangle formed by the intersection of the prediction box and the true value box; A DIoU3D (B p ,B g ) represents the predicted box B p With the truth box B g DIoU; IoU 3D (B p ,B g ) represents the IoU (intersection over union) between the predicted box and the candidate box in three-dimensional space;

[0034] To make the prediction more accurate and solve the problem that the 3D scene DIoU ignores the consistency of the aspect ratio between the predicted box and the true value box, a penalty term αv is added on the basis of the 3D scene DIoU, where α represents the balance parameter and v represents the correction factor. The calculation process is as follows:

[0035]

[0036] Among them, L1 and L represent the length of the true value box and the predicted box, W1 and W represent the width of the true value box and the predicted box; IoU represents the result of dividing the overlapping part of the predicted box and the true value box by the combined part of the two areas

[0037] Step 3: Training strategy formulation and parameter setting;

[0038] The boundary ranges of the point cloud data on the x, y, and z axes in three-dimensional space are respectively: width W from 0m to 70.4m, height H from -40m to 40m, and depth D from -1m to 5m;

[0039] When performing voxel processing, the voxel size selected is width v w =0.05m, height v h =0.05m, depth v d =0.1m, and limit each voxel to contain a maximum of 5 points, and the upper limit of the number of voxels is 16000;

[0040] During model training, the parameters were set using the Adam optimizer: the initial learning rate was set to 0.003, which was gradually decayed with the increase of training rounds. Each batch processed 4 samples, and a total of 1400 rounds of iterations were performed;

[0041] The momentum parameters β1 and β2 of the Adam optimizer are set to 0.9 and 0.999 respectively, the decay weight of the learning rate decay strategy is 0.01, and the decay factor is 0.8.

[0042] Step 4: Analyze the test results of power equipment components.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] (1) Good accuracy. The performance of this model is evaluated by measuring the average AP value of the points that the machine needs to pay attention to in the suspension insulator, transformer, bushing, and conductor. The experimental results show that the saliency learning module and WIoU loss help improve the model's detection accuracy of substation equipment in the substation scenario.

[0045] (2) Robustness. While achieving real-time detection, the proposed model can overcome the challenges posed by small and complex equipment and background interference in substation scenarios, effectively suppress noise in the detection scene, and effectively improve the problem of insufficient feature extraction in substation scenarios.

[0046] (3) Rich detection types. The present invention can detect types including but not limited to insulators, transformers, bushings and conductors. Therefore, the present invention is a defect detection method with good generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A step-by-step diagram of a power equipment detection method based on 3D point cloud saliency learning;

[0048] Figure 2 This is the overall framework structure diagram;

[0049] Figure 3 To enhance the channel attention module graph;

[0050] Figure 4 is the residual space attention module diagram;

[0051] Figure 5 Schematic diagram of three-dimensional DIoU. DETAILED DESCRIPTION

[0052] To gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description thereof. However, the present invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a comprehensive understanding of the present disclosure.

[0053] A method for detecting power equipment based on 3D point cloud saliency learning, the method comprises the following steps: Figure 1 As shown:

[0054] Step 1: Create a KITTI-formatted dataset for power scenarios. The original data was collected in a substation scenario using an Ouster OS-1-64 laser radar with a horizontal resolution of 2048 pixels and a vertical field of view of 45° (±22.5°). A wide-angle 120-degree camera with a resolution of 1920 x 1080 was used to capture RGB images. The Labelcloud point cloud annotation tool was then used to annotate the objects in the collected point cloud. These objects were then organized according to the KITTI dataset label format to create a self-built dataset suitable for training and testing. This dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The self-built dataset includes points of interest from four categories: insulators, transformers, bushings, and conductors. The distribution of these categories in the dataset is shown in Table 1.

[0055] Table 1 Dataset distribution table

[0056]

[0057] Step 2: Construct the SalSecond target detection model including the saliency learning module, sparse convolution module, and region recommendation network module; its overall model structure is as follows Figure 2 shown; wherein:

[0058] The saliency learning module enhances channel attention while ensuring the flow of information between groups, and learns different channel features in groups to obtain more refined global features of the point cloud. Subsequently, through point-by-point spatial attention learning, it further obtains salient features in the scene and reduces the impact of background points on detection accuracy.

[0059] The sparse convolution module extracts the obtained salient features;

[0060] The region recommendation network performs prediction regression and constructs the WIoU loss function. It also pays attention to the positional relationship between the prediction box and the true value box to improve the model regression accuracy.

[0061] Using the traditional target detection model Second as the baseline model, a saliency learning module is constructed in the feature extraction stage. An enhanced channel attention module and a residual spatial attention module are designed to focus on the spatial location information of the point cloud, reduce noise interference in the detection scene, and obtain saliency features of the point cloud in the power scene. The saliency features are fed into the sparse convolution module for feature extraction, and then into the region recommendation network module to construct the 3D bounding box as the overall regression loss function (WiUoU loss). This solves the problem that the original model's loss function ignores the coupling between different bounding box parameters, resulting in poor regression accuracy.

[0062] Step 2.1: The enhanced channel attention module divides the input channel into three non-interfering sub-networks, and inputs the voxelized point cloud data aggregated by the maximum pooling and average pooling operations respectively, focusing on the fine-grained features of the point cloud from different dimensions; then, the Concat operation is performed on the fine-grained features of different dimensions for splicing; the Shuffle operation is introduced to evenly mix the fine-grained features of different dimensions and then learn again through the three sub-networks and splice the final results; finally, nonlinear mapping is performed through the activation function to obtain higher-dimensional global features; the module structure is as follows Figure 3 As shown; the process is described by the following formula:

[0063]

[0064] Among them, i represents the sub-networks of different groups, ω i is the parameter learned by the fully connected layer in each sub-network, W max +W avg represents the features of the voxelized point cloud after maximum pooling and average pooling, σ is the Sigmoid activation function; f i (x) represents the global feature obtained by concatenating the results of the first grouping;

[0065] The Shuffle operation is an operation that mixes different features. Specifically, the features calculated by different channels in the first half of the network are first reshaped, then reshaped again after transposition, and then evenly mixed. Finally, the features are learned again through three different channels, and the learned features of different dimensions are spliced ​​to obtain the global features. The formula is as follows:

[0066]

[0067] W l Represents the parameters of different fully connected layers during the second grouping operation; σ represents the Sigmoid activation function;

[0068] Step 2.2: The residual spatial attention module adopts a self-attention structure to focus on the spatial distance of the point cloud and the correlation between shape features; introduces a residual connection mechanism to strengthen the relationship between input and output and improve the model's ability to capture significant features; the module structure is as follows Figure 4 Shown: Specifically:

[0069] First, the global features obtained by the enhanced channel attention module are mapped into three parameters Q, K, and V through three convolutional layers. Then, the relationship between different points is calculated using the transpose of Q and K. Finally, softmax is used for normalization to obtain the attention weight A and multiply it with V to obtain the salient feature. The process is described by the following formula:

[0070]

[0071] S=V×A+G (4)

[0072] Among them, S is the significant feature obtained, K T Represents matrix transpose, dk is the dimension of key value K.

[0073] Step 2.3: Construct the WIoU loss function for regression of the 3D bounding box as a whole to solve the problem that the original model loss function ignores the coupling between different parameters of the bounding box, resulting in poor regression accuracy; this process is expressed as the following formula:

[0074] WIoU=L DIoU3D +αv (5)

[0075] Among them, L DIoU3D is the DIoU in a three-dimensional scene, and its calculation formula is as follows:

[0076] L DIoU3D =1+A DIoU3D (B p ,B g ) (6)

[0077]

[0078] Among them, B g is the true target box, B p To predict the bounding box, V p is the area of ​​the prediction box, V g is the area of ​​the true value box, h i For B p With B g The intersection height, R i For B p With B g The volume of the intersection area, b is B p With B g The distance between the center points, d is the length of the diagonal of the minimum enclosed rectangle formed by the intersection of the prediction box and the true value box; A DIoU3D (B p ,B g ) represents the predicted box B p With the truth box B g DIoU; IoU 3D (B p ,B g ) represents the IoU (intersection over union) between the predicted box and the candidate box in three-dimensional space;

[0079] The DIoU diagram of the three-dimensional scene is as follows Figure 5 As shown;

[0080] A penalty term αv is added to the 3D scene DIoU, where α represents the balance parameter and v represents the correction factor. This solves the problem that the 3D scene DIoU ignores the consistency of the aspect ratio between the predicted box and the true value box. The calculation process is as follows:

[0081]

[0082]

[0083] Among them, L1 and L represent the length of the true value box and the predicted box, W1 and W represent the width of the true value box and the predicted box; IoU represents the result of dividing the overlapping part of the predicted box and the true value box by the combined part of the two areas.

[0084] Step 3: Formulate training strategies and set parameters. Train and test the deep learning model architecture constructed based on this method, and detect different devices in the substation scenario. The specific operations are as follows:

[0085] The boundary ranges of the point cloud data on the x, y, and z axes in three-dimensional space are respectively: width W from 0m to 70.4m, height H from -40m to 40m, and depth D from -1m to 5m;

[0086] When performing voxel processing, the voxel size selected is width v w =0.05m, height v h =0.05m, depth v d =0.1m, and limit each voxel to contain a maximum of 5 points, and the upper limit of the number of voxels is 16000;

[0087] During model training, the parameters were set using the Adam optimizer: the initial learning rate was set to 0.003, which was gradually decayed with the increase of training rounds. Each batch processed 4 samples, and a total of 1400 rounds of iterations were performed;

[0088] The momentum parameters β1 and β2 of the Adam optimizer are set to 0.9 and 0.999 respectively, the decay weight of the learning rate decay strategy is 0.01, and the decay factor is 0.8.

[0089] Step 4: Analysis of test results of power equipment components

[0090] The performance of this model was evaluated on a self-built dataset by measuring the average AP value of the machine's focus points in four categories: insulators, transformers, bushings, and conductors. The model was compared with other mainstream target detection models and a visual analysis was performed. The experimental results are shown in Table 2. The experimental results show that the average mean accuracy of the model on the self-built dataset is improved by 2.93% compared with the baseline model. The detection accuracy of the machine's focus points in the four categories of insulators, transformers, bushings, and conductors is improved by 3.43%, 2.03%, 2.89%, and 3.36%, respectively.

[0091] Table 2 Performance comparison of different detection methods on self-built datasets

[0092]

[0093]

[0094] To verify the effectiveness of the saliency learning module and WioU loss, ablation experiments were conducted on a self-built dataset. The results are shown in Table 3. The results show that after constructing the saliency learning module on the baseline network, the detection accuracy of each category increased by 2.19%, 1.28%, 2.00%, and 2.21%, respectively. This shows that the introduction of the saliency learning module improves the model's feature extraction capabilities. To further improve the accuracy of position regression, the original position regression loss function was replaced with the WioU loss. The results show that the detection accuracy of each device increased by 0.69%, 0.62%, 1.22%, and 1.15%, respectively, demonstrating the effectiveness of this method.

[0095] Table 3 Comparison of ablation test results on self-built dataset

[0096]

[0097] Any matters not described in detail in this specification are prior art known to those skilled in the art. Although the above description of the present invention is based on specific embodiments to facilitate understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the specific embodiments. As long as various modifications are within the spirit and scope of the present invention as defined and determined by the appended claims, such modifications will be obvious to those skilled in the art, and all inventions and creations utilizing the concepts of the present invention are protected.

Claims

1. A power equipment detection method based on 3D point cloud saliency learning, characterized in that: The method comprises the following steps: Step 1: Create a KITTI format dataset for power scenarios; Step 2: SalSecond object detection model including saliency learning module, sparse convolution module, and region recommendation network module; Step 3: Training strategy formulation and parameter setting; Step 4: Analysis of test results of power equipment components; Step 2 constructs a SalSecond object detection model including a saliency learning module, a sparse convolution module, and a region recommendation network module; wherein: The saliency learning module enhances channel attention while ensuring the flow of information between groups, and learns different channel features in groups to obtain more refined global features of the point cloud. Subsequently, through point-by-point spatial attention learning, it further obtains salient features in the scene and reduces the impact of background points on detection accuracy. The sparse convolution module extracts the obtained salient features; The region recommendation network module performs prediction regression by constructing the Widget Over Union (WIoU) loss function and focusing on the positional relationship between the prediction box and the true value box to improve the model regression accuracy. Step 2 constructs a SalSecond target detection model including a saliency learning module, a sparse convolution module, and a region recommendation network module; the specific operations are as follows: Using the Second target detection model as the baseline model, a saliency learning module is constructed in the feature extraction stage. An enhanced channel attention module and a residual spatial attention module are designed to focus on the spatial location information of the point cloud, reduce noise interference in the detection scene, and obtain saliency features of the point cloud in the power scene. The saliency features are fed into the sparse convolution module for feature extraction, and then into the region recommendation network module to construct the WIoU loss, a loss function for the 3D bounding box as the whole for regression. This solves the problem that the original model's loss function ignores the coupling between different parameters of the bounding box, resulting in poor regression accuracy. The target detection model Second is used as the benchmark model. In the feature extraction stage, a saliency learning module is constructed. The enhanced channel attention module and the residual space attention module are designed to focus on the spatial position information of the point cloud, reduce the noise interference in the detection scene, and obtain the saliency features of the point cloud in the power scene. The specific operations are as follows: The voxelized point cloud data is input into the network and processed by the enhanced channel attention module to obtain the global feature G. Then, it is processed by the residual space attention module to obtain the salient feature S of the point cloud. The process is as follows: Step 2.1: The enhanced channel attention module divides the input channel into three non-interfering sub-networks and inputs the voxelized point cloud data after the maximum pooling and average pooling operations respectively, focusing on the fine-grained features of the point cloud from different dimensions. Subsequently, the concat operation is performed on the fine-grained features of different dimensions for splicing. The shuffle operation is introduced to evenly mix the fine-grained features of different dimensions and then learn them again through the three sub-networks and splice the final results. Finally, nonlinear mapping is performed through the activation function to obtain higher-dimensional global features. The process is described by the following formula: Among them, i represents the sub-networks of different groups, ω i is the parameter learned by the fully connected layer in each sub-network, W max +W avg represents the features of the voxelized point cloud after maximum pooling and average pooling, σ is the Sigmoid activation function; f i (x) represents the global feature obtained by concatenating the results of the first grouping; The results obtained from each sub-network are then spliced ​​and the Shuffle operation is performed. Specifically, the features calculated from different channels in the first half of the network are reshaped, transposed and reshaped again, and then evenly mixed. Finally, the network is learned again through three different channels, and the learned features of different dimensions are spliced ​​to obtain the global feature G. The formula is as follows: W l represents the parameters of different fully connected layers during the second grouping operation; σ represents the Sigmoid activation function; Step 2.2: The residual spatial attention module adopts a self-attention structure to focus on the spatial distance of the point cloud and the correlation between shape features. It also introduces a residual connection mechanism to strengthen the relationship between input and output and improve the model's ability to capture salient features. Specifically: First, the global features obtained by the enhanced channel attention module are mapped into three parameters Q, K, and V through three convolutional layers. Then, the relationship between different points is calculated using the transpose of Q and K. Finally, softmax is used for normalization to obtain the attention weight A and multiply it by V to obtain the significant feature S. The process is described by the following formula: S=V×A+G (4) Where S is the significant feature obtained, K T Represents matrix transpose, dk is the dimension of key value K; The significant features are sent to the sparse convolution module for feature extraction, and then enter the region recommendation network module to construct the 3D bounding box as the overall regression loss function WIoU loss. The specific operation to solve the problem that the original model loss function ignores the coupling between different parameters of the bounding box and causes the regression accuracy to deteriorate is as follows: Step 2.3: Construct the WIoU loss function for regression of the 3D bounding box as a whole to solve the problem that the original model loss function ignores the coupling between different parameters of the bounding box, resulting in poor regression accuracy; this process is expressed as the following formula: WIoU=L DIoU3D +αv (5) Among them, L DIoU3d is the DIoU in a three-dimensional scene, and its calculation formula is as follows: L DIoU3D =1+A DIoU3D (B p ,B g ) (6) Among them, B g is the true target box, B p To predict the bounding box, V p is the area of ​​the prediction box, V g is the area of ​​the true value box, h i B p With B g The intersection height, R i B p With B g The volume of the intersection area, b is B p With B g The distance between the center points, d is the length of the diagonal of the minimum enclosed rectangle formed by the intersection of the prediction box and the true value box; A DIoU3D (B p ,B g ) represents the predicted box B p With the truth box B g DIoU; IoU 3D (B p ,B g ) represents the IoU between the predicted box and the candidate box in three-dimensional space; A penalty term αv is added to the 3D scene DIoU, where α represents the balance parameter and v represents the correction factor. This solves the problem that the 3D scene DIoU ignores the consistency of the aspect ratio between the predicted box and the true value box. The calculation process is as follows: Among them, L1 and L represent the length of the true value box and the predicted box, W1 and W represent the width of the true value box and the predicted box; IoU represents the result of dividing the overlapping part of the predicted box and the true value box by the combined part of the two areas.

2. The power equipment detection method based on 3D point cloud saliency learning according to claim 1 is characterized in that: The specific operations of step 1 to establish the KITTI format dataset for the power scene are: using a lidar to collect point cloud data, using a camera to collect RGB images, using a point cloud annotation tool to annotate the targets in the collected point cloud, and organizing them according to the KITTI dataset label format to obtain a self-built dataset for training and testing; the self-built dataset is divided into a training set, a validation set, and a test set according to the proportion.

3. The power equipment detection method based on 3D point cloud saliency learning according to claim 2 is characterized in that: The specific steps in step 3 are as follows: The boundary ranges of the point cloud data on the x, y, and z axes in three-dimensional space are respectively: width W from 0m to 70.4m, height H from -40m to 40m, and depth D from -1m to 5m; When performing voxel processing, the voxel size selected is width v w =0.05m, height v h =0.05m, depth v d =0.1m, and limit each voxel to contain a maximum of 5 points, and the upper limit of the number of voxels is 16000; During model training, the parameters were set using the Adam optimizer: the initial learning rate was set to 0.003, which was gradually decayed with the increase of training rounds. Each batch processed 4 samples, and a total of 1400 rounds of iterations were performed; The momentum parameters β1 and β2 of the Adam optimizer are set to 0.9 and 0.999 respectively, the decay weight of the learning rate decay strategy is 0.01, and the decay factor is 0.8.

Images