Indoor three-dimensional point cloud object detection method based on sparse feature enhancement

Abstract

This invention discloses an indoor 3D point cloud object detection method based on sparse feature enhancement: 1. Constructing a feature extraction network based on lightweight spatial-channel joint enhancement to perform multi-stage sparse feature extraction on voxelized point clouds; 2. Constructing a multi-scale feature fusion network based on channel recalibration, fusing the multi-scale sparse features output by the backbone network from top to bottom, and introducing a channel weight attention gating module into the fused features for channel-level weight allocation and semantic recalibration; 3. Inputting the multi-scale fused features into a dual-branch detection head, performing 3D bounding box regression and category classification respectively, to obtain the indoor 3D point cloud object detection results. The indoor object detection method based on sparse feature enhancement provided by this invention solves the problems of insufficient adaptive characterization of key spatial structures and channel importance, limited deep local geometric representation, and redundancy of multi-scale fused features in existing voxel and sparse convolution detection methods.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and artificial intelligence technology, and specifically relates to an indoor 3D point cloud object detection method based on sparse feature enhancement. Background Technology

[0002] Indoor 3D point cloud target detection requires identifying target categories and regressing 3D bounding boxes from disordered, sparse, and unevenly distributed point cloud data. Compared to 2D images, indoor point clouds typically suffer from problems such as dense object distribution, severe occlusion, significant differences in target scale, missing local geometric structures, and complex background structures. Detection models are susceptible to redundant responses and noise interference during feature modeling, leading to insufficient representation of target-related features and thus affecting detection accuracy and localization stability.

[0003] However, existing voxel and sparse convolution detection frameworks still have the following shortcomings in feature representation in complex indoor scenes: First, in the process of extracting low-to-mid-level sparse features, ordinary sparse convolution usually relies on fixed convolution kernels for local neighborhood aggregation, using the same processing method for all channels and spatial locations. It lacks the ability to adaptively model key spatial structures and important channel responses, and in scenes with uneven point cloud distribution and severe occlusion, it is easy to introduce redundant background responses, making it difficult to effectively highlight key target features. Second, after the deep sparse features are progressively downsampled, although the semantic information is continuously enhanced, local geometric details are easily weakened, affecting the fine representation of target boundaries and local structures by deep features, and thus affecting high-precision localization and bounding box regression. Third, in the process of multi-scale feature fusion, there are semantic differences and channel redundancy between features at different levels. Traditional fusion methods usually cannot finely screen the contributions of different channels, which can easily reduce the consistency and discriminativeness of fused features, affecting the stability and accuracy of the final detection results.

[0004] Therefore, how to effectively highlight target-related features and suppress redundant responses while maintaining the efficiency of sparse convolution, and how to construct adaptive screening and enhancement mechanisms in the feature extraction and multi-scale fusion stages, has become a key issue in improving the performance of indoor 3D point cloud target detection. There is an urgent need for an indoor object detection method aimed at improving the quality of sparse voxel features, optimizing the sparse feature representation process from aspects such as spatial-channel joint modeling, deep local structure reconstruction, and fusion feature channel recalibration, thereby improving detection accuracy, localization stability, and inference efficiency in complex indoor scenes. Summary of the Invention

[0005] The purpose of this invention is to provide an indoor object detection method based on sparse feature enhancement, which solves the problems of insufficient adaptive characterization of key spatial structures and channel importance, limited expression of deep local geometry, and redundancy of multi-scale fusion features in existing voxel and sparse convolution detection methods.

[0006] The technical solution adopted in this invention is: an indoor 3D point cloud object detection method based on sparse feature enhancement, the specific method of which is as follows: Step 1: Construct a feature extraction network based on lightweight spatial-channel joint enhancement, perform multi-stage sparse feature extraction on voxelized point clouds, introduce sparse tensor channel-spatial attention (STCA) in the shallow layer, and introduce sparse local residual convolution (SLR-Conv) in the deep layer. Step 2: Construct a multi-scale feature fusion network based on channel recalibration. The multi-scale sparse features output by the backbone network are fused from top to bottom. The fused features are then introduced into a channel weight attention gating module for channel-level weight allocation and semantic recalibration. Step 3: Input the multi-scale fusion features into the dual-branch detection head, and perform 3D bounding box regression and category classification respectively to obtain the indoor 3D point cloud object detection results.

[0007] The invention is further characterized by: The specific method for step 1 is as follows: Step 1.1: Construct an initial sparse feature encoding layer. Perform local spatial aggregation, channel expansion, and structured downsampling on the low-dimensional input features of the voxelized point cloud to obtain the initial sparse features. Step 1.2: Construct a multi-stage sparse residual feature extraction layer, and form sparse feature representations of different scales by downsampling layer by layer and connecting with the residual; Step 1.3: In the low-to-mid-level feature extraction stage, the Sparse Tensor Channel-Spatial Attention Module (STCA) is introduced to co-model the local geometry and channel response, generate selective attention weights, and adaptively recalibrate the original sparse features. Step 1.4: In the deep feature extraction stage, a sparse local residual convolution module (SLR-Conv) is introduced to perform local spatial reconstruction in the low-dimensional channel space, and to achieve the refinement and stable enhancement of deep semantic features through residual connections.

[0008] The specific method for step 1.1 is as follows: Step 1.1.1: Perform sparse voxelization on the input point cloud to obtain a sparse tensor containing effective voxel coordinates and initial voxel features; Step 1.1.2: Use a sparse convolution with a kernel size of 3×3×3 and a stride of 2 to perform local neighborhood feature aggregation and channel mapping on the sparse tensor to complete the initial spatial downsampling;

[0009] in, Indicates the input sparse tensor; This indicates a sparse convolution operation, specifically a 2x spatial downsampling convolution with a kernel size of 3×3×3 and a stride of 2. Step 1.1.3: Perform batch normalization and non-linear activation on the convolutional output sequentially, and then perform structured compression through sparse max pooling to form the initial sparse feature representation:

[0010] in, Indicates the initial extracted features; This indicates a batch normalization operation; This indicates a max pooling operation, where the maximum pooling step size and receptive field are 2, achieving 2x spatial downsampling. This represents the activation function.

[0011] The specific method for step 1.2 is as follows: Step 1.2.1: Construct multiple consecutive sparse residual feature extraction stages, each stage consisting of several sparse residual feature extraction modules, gradually expanding the receptive field and improving semantic expressive power:

[0012] in, This represents the feature extraction layer in the i-th stage; This represents the output feature of the i-th layer; This represents the N sparse residual feature extraction downsampling modules of the i-th stage feature extraction layer; Step 1.2.2: In each sparse residual feature extraction module, local neighborhood feature aggregation and semantic enhancement are completed through sparse convolution, batch normalization, and nonlinear activation. Input features and transformed features are then fused through residual connections.

[0013] in, Represents input features; Indicates a semantic deepening branch of features; Indicates the spatial resolution downsampling branch; This represents a 3×3×3 sparse convolution operation; Step 1.2.3: In the first sparse residual feature extraction module of each stage, set up a downsampling branch. When the stride is greater than 1, use convolutional downsampling mapping instead of identity mapping.

[0014] Complete the reduction of spatial resolution and adjustment of channel dimensions.

[0015] The specific method for step 1.3 is as follows: Step 1.3.1: Perform a 3×3×3 sparse convolution on the input sparse voxel features to aggregate the local neighborhood geometric context without changing the sparse coordinate distribution, thus obtaining local spatial features.

[0016] Step 1.3.2: Feed the local spatial features into a low-dimensional channel mapping structure consisting of two 1×1×1 sparse convolutions. First, compress the channel dimension to extract a compact channel importance representation. Then, enhance the inter-channel dependencies through non-linear activation. Finally, restore the original channel dimension and generate channel-spatial attention weights through Sigmoid activation.

[0017] Step 1.3.3: Multiply the channel-spatial attention weights with the original input sparse voxel features on a voxel-by-voxel and channel-by-channel basis to output the enhanced features after adaptive recalibration.

[0018] in, Represents local spatial features; Indicates channel-spatial attention weights; This indicates that the product of voxels and channels is multiplied. This indicates the enhanced features of the STCA output.

[0019] The specific method for step 1.4 is as follows: Step 1.4.1: Perform a 1×1×1 sparse convolution on the sparse features output from the deep layer to compress the channel dimension into a low-dimensional feature space:

[0020] Step 1.4.2: Perform 3×3×3 sparse convolution in the low-dimensional feature space to perform secondary local spatial modeling of deep semantic features, thereby strengthening the fine-grained geometric structure weakened by the downsampling process.

[0021] Step 1.4.3: Map the enhanced features back to the original channel dimensions using 1×1×1 sparse convolution, and then obtain the recovered features through batch normalization and non-linear activation.

[0022] Step 1.4.4: Add the recovered features to the module input features using residual summation to obtain the deep enhancement features output by the sparse local residual convolution module:

[0023] in, , representing the sparse voxel feature tensor output by the sparse residual feature extraction module. N This indicates the number of voxel features in the sparse residual feature extraction module of this layer. C Dimensions representing voxel features; These represent the low-dimensional channel features, geometric features, and restored features after convolution in the sparse residual feature extraction module, respectively. This represents an i×i×i sparse convolution operation; This represents the residual output feature tensor.

[0024] The specific method for step 2 is as follows: Step 2.1: Upsample the input sparse features using generative transpose convolution, and then obtain aligned upsampled features through batch normalization and non-linear activation:

[0025] in, Indicates transpose; This represents a generative convolution with a kernel size of 3×3×3 and a stride of 2. Step 2.2: Fuse the upsampled high-level semantic features with the corresponding downsampled features from the previous layer in the backbone network element by element to obtain fused features at different scales.

[0026] Step 2.3: Perform sparse convolution on the cross-layer fusion features to re-aggregate features within the local receptive field, alleviating the scale differences and feature aliasing problems caused by cross-layer fusion, and obtaining stable reconstituted features:

[0027] Step 2.4: Embed features of different scales into the same feature space and map them to the fixed input channel dimension required by the detection head.

[0028] Step 2.5: Introduce the Channel Weight Attention Gating Module (CWAG) to perform semantic relabeling on the fused multi-scale features.

[0029] The specific method for step 2.5 is as follows: Step 2.5.1: Perform global average pooling on the input fused features to compress the spatial dimension and obtain the channel description vector that reflects the overall semantic distribution.

[0030] Step 2.5.2: Model the nonlinear dependency between channels using two fully connected layers, enhance the representation with ReLU activation in between, and reduce the hidden layer dimension with a channel compression ratio r to control the number of parameters and computational cost while ensuring modeling capability. Finally, generate channel weights with Sigmoid activation.

[0031] in, This indicates that the number of output channel dimensions is A fully connected layer; Indicates the channel compression ratio; The number of channel dimensions is expressed as A fully connected layer; Step 2.5.3: Broadcast the channel weights to the corresponding positions of the fused features, and perform channel-by-channel weighting on the fused features to highlight key semantic channels and suppress redundant channel responses.

[0032] The specific method for step 3 is as follows: Step 3.1, Design of a lightweight, anchor-free detection head using sparse convolution: Step 3.1.1: Using the multi-scale fused voxel features output by the neck network as input, the features are fed into the bounding box regression branch and the category classification branch respectively. The two branches share the input features and are responsible for the 3D bounding box parameter prediction and target category prediction tasks respectively. Step 3.1.2: Using a point-based regression approach, with each valid voxel as a candidate center, the 6-dimensional size parameters of the center offset and axis-aligned 3D bounding box are directly predicted through sparse convolution. The axis alignment IoU loss is only optimized for positive samples. Step 3.1.3: Output the class confidence logits for each voxel through sparse convolution, convert them into class probabilities through Sigmoid activation, and use Focal Loss during training to alleviate the class imbalance problem; Step 3.1.4: After completing the predictions on features at different scales, the prediction results at each scale are fused together in the inference stage to output the final target category and 3D bounding box, thereby achieving efficient end-to-end prediction of sparse voxel features. Step 3.2, Multi-condition joint constraint label allocation strategy.

[0033] The specific method for step 3.2 is as follows: Step 3.2.1: Based on the scale differences of different target categories, preset the mapping relationship between categories and feature levels, retain only the points in the voxel level that match the target category as candidate positive samples, and filter out unmatched samples from the level dimension. Step 3.2.2: For each real target, select the K points closest to the target center from the candidate points that meet the hierarchical constraints as positive samples to stabilize the number of positive samples and alleviate the sample imbalance problem caused by scale differences and point cloud sparsity. Step 3.2.3: For voxel points that match multiple targets simultaneously, assign them to the nearest target to ensure that each voxel point corresponds to only one true bounding box, avoid supervision conflicts, and improve the stability of label assignment.

[0034] The beneficial effects of this invention are: This invention presents an indoor object detection method based on sparse feature enhancement. Using SFE-Det as the overall detection framework, it enhances sparse point cloud feature representation from three levels: spatial-channel joint modeling, deep local feature reconstruction, and fused feature channel recalibration. Compared to a detection process consisting only of conventional sparse convolution, feature pyramids, and the detection head sequence, this invention focuses on constructing a multi-level information filtering mechanism that runs through the feature extraction and feature fusion stages. This mechanism adaptively strengthens key semantic information and suppresses redundant responses, improving detection accuracy, localization stability, and robustness in complex indoor scenes without significantly increasing computational overhead. Attached Figure Description

[0035] Figure 1 This is a structural diagram of the Sparse Tensor Channel-Spatial Attention Module (STCA) in the indoor object detection method of this invention; Figure 2 This is a structural diagram of the sparse local residual convolution module (SLR-Conv) in the indoor object detection method of this invention; Figure 3 This is a structural diagram of the lightweight channel attention gating module (CWAG) in the indoor object detection method of the present invention; Figure 4 This is a visualized indoor object detection result diagram of scene one in the indoor object detection method of the present invention; Figure 5 This is a visualized image of the indoor object detection results in scenario two of the indoor object detection method of this invention. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0037] This invention presents an indoor 3D point cloud object detection method based on sparse feature enhancement, the specific method of which is as follows: Step 1: Sparse voxelize the input indoor 3D point cloud, construct a feature extraction network based on lightweight spatial-channel joint enhancement, and perform multi-stage sparse feature extraction on the voxelized point cloud. In the low-to-mid-level feature extraction stage, the sparse tensor channel-space attention module STCA is introduced to co-model the local geometric structure and channel response and adaptively recalibrate the sparse features. In the deep feature extraction stage, the sparse local residual convolution module SLR-Conv is introduced to perform local spatial reconstruction and residual enhancement of deep semantic features. The specific method is as follows: Step 1.1: Construct an initial sparse feature encoding layer. Perform local spatial aggregation, channel expansion, and structured downsampling on the low-dimensional input features of the voxelized point cloud to obtain the initial sparse features. The specific method is as follows: Step 1.1.1: Perform sparse voxelization on the input point cloud to obtain a sparse tensor containing effective voxel coordinates and initial voxel features; Step 1.1.2: Use a sparse convolution with a kernel size of 3×3×3 and a stride of 2 to perform local neighborhood feature aggregation and channel mapping on the sparse tensor to complete the initial spatial downsampling; (1) in, Indicates the input sparse tensor; This indicates a sparse convolution operation, specifically a 2x spatial downsampling convolution with a kernel size of 3×3×3 and a stride of 2. Step 1.1.3: Perform batch normalization and nonlinear activation on the convolutional output sequentially to stabilize the feature distribution and enhance the nonlinear expression. Then, perform structured compression through sparse max pooling to form the initial sparse feature representation. (2) in, Indicates the initial extracted features; This indicates a batch normalization operation; This indicates a max pooling operation, where the maximum pooling step size and receptive field are 2, achieving 2x spatial downsampling. This represents the activation function.

[0038] Step 1.2: Construct a multi-stage sparse residual feature extraction layer, forming sparse feature representations at different scales through layer-by-layer downsampling and concatenation with the residuals; the specific method is as follows: Step 1.2.1: Construct multiple consecutive sparse residual feature extraction stages, each stage consisting of several components, to gradually expand the receptive field and improve semantic expressive power;

[0039] in, This represents the feature extraction layer in the i-th stage; This represents the output feature of the i-th layer; This represents the N sparse residual feature extraction downsampling modules of the i-th stage feature extraction layer; Step 1.2.2: In each sparse residual feature extraction module, BasicBlock, local neighborhood feature aggregation and semantic enhancement are completed through sparse convolution, batch normalization, and nonlinear activation. Input features and transformed features are then fused through residual connections.

[0040] in, Represents input features; Indicates a semantic deepening branch of features; Indicates the spatial resolution downsampling branch; This represents a 3×3×3 sparse convolution operation; Step 1.2.3: In the first sparse residual feature extraction module BasicBlock of each stage, set up a downsampling branch. When the stride is greater than 1, use convolutional downsampling mapping instead of identity mapping:

[0041] Complete the reduction of spatial resolution and adjustment of channel dimensions.

[0042] Step 1.3: In the low-to-mid-level feature extraction stage, the Sparse Tensor Channel-Spatial Attention Module (STCA) is introduced to co-model the local geometry and channel response, generate selective attention weights, and adaptively recalibrate the original sparse features. Unlike ordinary sparse convolution, which uses fixed weights for all channels and spatial locations, this invention introduces a sparse tensor channel-space attention module (STCA) in the mid-to-low-level feature extraction stage. Figure 1 As shown, by using a cascaded structure of spatial convolution and channel mapping, we can achieve joint modeling of local geometric context and channel importance while keeping the sparse voxel coordinates unchanged. The specific method is as follows: Step 1.3.1: Perform a 3×3×3 sparse convolution on the input sparse voxel features to aggregate the local neighborhood geometric context without changing the sparse coordinate distribution, thus obtaining local spatial features.

[0043] Step 1.3.2: Feed the local spatial features into a low-dimensional channel mapping structure consisting of two 1×1×1 sparse convolutions. First, compress the channel dimension to extract a compact channel importance representation. Then, enhance the inter-channel dependencies through non-linear activation. Finally, restore the original channel dimension and generate channel-spatial attention weights through Sigmoid activation.

[0044] Step 1.3.3: Multiply the channel-spatial attention weights with the original input sparse voxel features on a voxel-by-voxel and channel-by-channel basis to output the enhanced features after adaptive recalibration. (8); in, Represents local spatial features; Indicates channel-spatial attention weights; This indicates that the product of voxels and channels is multiplied. This indicates the enhanced features of the STCA output.

[0045] Step 1.4: In the deep feature extraction stage, a sparse local residual convolution module (SLR-Conv) is introduced to perform local spatial reconstruction in the low-dimensional channel space, and to achieve the refinement and stable enhancement of deep semantic features through residual connections.

[0046] Unlike conventional residual modules that rely solely on basic sparse convolutions for feature transfer at deeper layers, this invention introduces a sparse local residual convolution module (SLR-Conv) in the deep feature extraction stage, such as... Figure 2 As shown, a lightweight approach involving channel compression, spatial reconstruction, and residual enhancement is used to supplement deep semantic features with fine-grained geometric structures. The specific method is as follows: Step 1.4.1: Perform a 1×1×1 sparse convolution on the sparse features output from the deep layer stage to compress the channel dimension into a low-dimensional feature space, reducing the computational cost of subsequent spatial convolutions.

[0047] Step 1.4.2: Perform 3×3×3 sparse convolution in the low-dimensional feature space to perform secondary local spatial modeling of deep semantic features, thereby strengthening the fine-grained geometric structure weakened by the downsampling process.

[0048] Step 1.4.3: Map the enhanced features back to the original channel dimensions using 1×1×1 sparse convolution, and then obtain the recovered features through batch normalization and non-linear activation.

[0049] Step 1.4.4: Add the recovered features to the module input features using residual summation to obtain the deep enhancement features output by the sparse local residual convolution module:

[0050] in, , representing the sparse voxel feature tensor output by the sparse residual feature extraction module.N This indicates the number of voxel features in the sparse residual feature extraction module of this layer. C Dimensions representing voxel features; These represent the low-dimensional channel features, geometric features, and restored features after convolution in the sparse residual feature extraction module, respectively. This represents an i×i×i sparse convolution operation; This represents the residual output feature tensor.

[0051] Step 1 of this invention does not simply stack and reuse traditional sparse convolutional backbones. Instead, it focuses on improving the discriminative power and representational quality of sparse features. During multi-stage feature extraction, it simultaneously embeds two lightweight enhancement mechanisms: Sparse Tensor Channel-Spatial Attention (STCA) and Sparse Local Residual Convolution (SLR-Conv). STCA, while maintaining the sparse voxel coordinates, achieves joint modeling of local geometric context and channel importance, enabling the network to adaptively focus on key target regions and enhance effective channel responses. SLR-Conv, through a lightweight path of channel compression, sparse convolution, and residual fusion, supplements deep, high-semantic features with fine-grained geometric details with low computational overhead. The synergistic effect of these two mechanisms allows the backbone network to perform multi-scale feature extraction while simultaneously filtering key responses and refining deep structural details, thereby improving the effectiveness and robustness of sparse point cloud features from the source.

[0052] Step 2: Construct a multi-scale feature fusion network based on channel recalibration. Perform top-down fusion on the multi-scale sparse features output by the backbone network. The fused features are then subjected to a channel weight attention gating module (CWAG) for channel-level weight allocation and semantic recalibration, suppressing redundancy and highlighting key semantic information. The specific method is as follows: Step 2.1: Upsample deep sparse features using generative transpose convolution to explicitly generate new sparse spatial coordinates and restore spatial resolution; specifically: upsample the input sparse features using generative transpose convolution to restore spatial resolution and reconstruct the sparse coordinate structure, and then obtain aligned upsampled features through batch normalization and nonlinear activation:

[0053] in, Indicates transpose; This represents a generative convolution with a kernel size of 3×3×3 and a stride of 2. Step 2.2: The upsampled high-level semantic features are fused element-by-element with the corresponding layer's backbone network features to obtain cross-layer fused features. Specifically, the upsampled high-level semantic features are fused element-by-element with the downsampled features of the previous layer in the backbone network to obtain fused features at different scales.

[0054] Step 2.3: Locally reshape and smooth the cross-layer fusion features to eliminate semantic inconsistencies and feature aliasing caused by the direct superposition of features at different scales. Specifically, sparse convolution is performed on the cross-layer fusion features to re-aggregate features within the local receptive field, alleviating the scale differences and feature aliasing problems caused by cross-layer fusion, and obtaining stable reshaped features.

[0055] Step 2.4: Perform unified channel mapping on output features at different scales to embed them into the same feature space to adapt to the detection head; specifically, embed features at different scales into the same feature space and map them to the fixed input channel dimensions required by the detection head.

[0056] Step 2.5: Introduce the Channel Weight Attention Gating Module (CWAG) to perform semantic relabeling on the fused multi-scale features.

[0057] Unlike existing fusion networks that only perform feature space alignment and lack channel filtering mechanisms, this invention introduces a channel weight attention gating module (CWAG), such as... Figure 3 As shown, channel-level recalibration is performed on the fused features to achieve efficient semantic channel filtering without the need for spatial attention modeling. The specific method is as follows: Step 2.5.1: Perform global average pooling on the input fused features to compress the spatial dimension and obtain the channel description vector that reflects the overall semantic distribution.

[0058] Step 2.5.2: Model the nonlinear dependencies between channels using two fully connected layers (Linear1, Linear2), with ReLU activation to enhance the representation. Simultaneously, reduce the hidden layer dimension using a channel compression ratio r to control the number of parameters and computational cost while maintaining modeling capability. Finally, generate channel weights using Sigmoid activation.

[0059] in, This indicates that the number of output channel dimensions is A fully connected layer; Indicates the channel compression ratio; The number of channel dimensions is expressed as A fully connected layer; Step 2.5.3: Broadcast the channel weights to the corresponding positions of the fused features, and perform channel-by-channel weighting on the fused features to highlight key semantic channels and suppress redundant channel responses.

[0060] In the multi-scale feature fusion design of step 2 of this invention, instead of using the traditional top-down feature stacking and upsampling operation, a lightweight channel weight adaptive gating module (CWAG) is introduced in the fusion stage to suppress channel redundancy and enhance the discriminative power of the fused features. This module no longer performs spatial attention modeling repeatedly, but directly performs channel-level adaptive recalibration on the fused features after the multi-scale sparse features have been spatially aligned. This dynamically highlights the semantic channels that contribute more to 3D object detection, effectively alleviating the information redundancy problem in the fusion process and significantly improving the expression quality and detection stability of multi-scale features with almost no increase in inference time.

[0061] Step 3: Input the multi-scale fused features into the dual-branch detector head, and perform 3D bounding box regression and category classification respectively to obtain the indoor 3D point cloud object detection results. The specific method is as follows: Step 3.1, Design of a lightweight, anchor-free detection head using sparse convolution: Step 3.1.1: Using the multi-scale fused voxel features output by the neck network as input, the features are fed into the bounding box regression branch and the category classification branch respectively. The two branches share the input features and are responsible for the 3D bounding box parameter prediction and target category prediction tasks respectively. Step 3.1.2: Using a point-based regression approach, with each valid voxel as a candidate center, the 6-dimensional size parameters of the center offset and axis-aligned 3D bounding box are directly predicted through sparse convolution. The axis alignment IoU loss is only optimized for positive samples. Step 3.1.3: Output the class confidence logits for each voxel through sparse convolution, convert them into class probabilities through Sigmoid activation, and use Focal Loss during training to alleviate the class imbalance problem; Step 3.1.4: After completing the predictions on features at different scales, the prediction results at each scale are fused together in the inference stage to output the final target category and 3D bounding box, thereby achieving efficient end-to-end prediction of sparse voxel features. Step 3.2: Multi-condition joint constraint label assignment strategy. The specific method is as follows: Step 3.2.1: Based on the scale differences of different target categories, preset the mapping relationship between categories and feature levels, retain only the points in the voxel level that match the target category as candidate positive samples, and filter out unmatched samples from the level dimension. Step 3.2.2: For each real target, select the K points closest to the target center from the candidate points that meet the hierarchical constraints as positive samples to stabilize the number of positive samples and alleviate the sample imbalance problem caused by scale differences and point cloud sparsity. Step 3.2.3: For voxel points that match multiple targets simultaneously, assign them to the nearest target to ensure that each voxel point corresponds to only one true bounding box, avoid supervision conflicts, and improve the stability of label assignment.

[0062] To further verify the technical effects of the present invention, qualitative and quantitative experiments were conducted in typical indoor scenarios. For example... Figure 4 , 5 As shown in the figure, the method of the present invention can achieve accurate and stable detection even in complex situations such as dense occlusion or small targets.

[0063] Example 1 This embodiment is based on an indoor 3D point cloud object detection method with sparse feature enhancement, as detailed below: Step 1: Construct a feature extraction network based on lightweight spatial-channel joint enhancement, perform multi-stage sparse feature extraction on voxelized point clouds, introduce sparse tensor channel-spatial attention (STCA) in the shallow layer, and introduce sparse local residual convolution (SLR-Conv) in the deep layer. Step 2: Construct a multi-scale feature fusion network based on channel recalibration. The multi-scale sparse features output by the backbone network are fused from top to bottom. The fused features are then introduced into a channel weight attention gating module for channel-level weight allocation and semantic recalibration. Step 3: Input the multi-scale fusion features into the dual-branch detection head, and perform 3D bounding box regression and category classification respectively to obtain the indoor 3D point cloud object detection results.

[0064] Example 2 This embodiment is based on an indoor 3D point cloud object detection method with sparse feature enhancement. Building upon Embodiment 1, the specific method for step 1 is as follows: Step 1.1: Construct an initial sparse feature encoding layer. Perform local spatial aggregation, channel expansion, and structured downsampling on the low-dimensional input features of the voxelized point cloud to obtain the initial sparse features. Step 1.2: Construct a multi-stage sparse residual feature extraction layer, and form sparse feature representations of different scales by downsampling layer by layer and connecting with the residual; Step 1.3: In the low-to-mid-level feature extraction stage, the Sparse Tensor Channel-Spatial Attention Module (STCA) is introduced to co-model the local geometry and channel response, generate selective attention weights, and adaptively recalibrate the original sparse features. Step 1.4: In the deep feature extraction stage, a sparse local residual convolution module (SLR-Conv) is introduced to perform local spatial reconstruction in the low-dimensional channel space, and to achieve the refinement and stable enhancement of deep semantic features through residual connections.

[0065] Example 3 This embodiment is based on an indoor 3D point cloud object detection method with sparse feature enhancement. Based on Embodiment 2, the specific method for step 1.1 is as follows: Step 1.1.1: Perform sparse voxelization on the input point cloud to obtain a sparse tensor containing effective voxel coordinates and initial voxel features; Step 1.1.2: Use a sparse convolution with a kernel size of 3×3×3 and a stride of 2 to perform local neighborhood feature aggregation and channel mapping on the sparse tensor to complete the initial spatial downsampling; (1) in, Indicates the input sparse tensor; This indicates a sparse convolution operation, specifically a 2x spatial downsampling convolution with a kernel size of 3×3×3 and a stride of 2. Step 1.1.3: Perform batch normalization and non-linear activation on the convolutional output sequentially, and then perform structured compression through sparse max pooling to form the initial sparse feature representation: (2) in, Indicates the initial extracted features; This indicates a batch normalization operation; This indicates a max pooling operation, where the maximum pooling step size and receptive field are 2, achieving 2x spatial downsampling. This represents the activation function.

[0066] Example 4 This embodiment is based on an indoor 3D point cloud object detection method with sparse feature enhancement. Based on Embodiment 2, the specific method for step 1.2 is as follows: Step 1.2.1: Construct multiple consecutive sparse residual feature extraction stages, each stage consisting of several sparse residual feature extraction modules, gradually expanding the receptive field and improving semantic expressive power:

[0067] in, This represents the feature extraction layer in the i-th stage; This represents the output feature of the i-th layer; This represents the N sparse residual feature extraction downsampling modules of the i-th stage feature extraction layer; Step 1.2.2: In each sparse residual feature extraction module, local neighborhood feature aggregation and semantic enhancement are completed through sparse convolution, batch normalization, and nonlinear activation. Input features and transformed features are then fused through residual connections.

[0068] in, Represents input features; Indicates a semantic deepening branch of features; Indicates the spatial resolution downsampling branch; This represents a 3×3×3 sparse convolution operation; Step 1.2.3: In the first sparse residual feature extraction module of each stage, set up a downsampling branch. When the stride is greater than 1, use convolutional downsampling mapping instead of identity mapping.

[0069] Complete the reduction of spatial resolution and adjustment of channel dimensions.

[0070] Example 5 This embodiment is based on an indoor 3D point cloud object detection method with sparse feature enhancement. Based on Embodiment 2, the specific method for step 1.3 is as follows: Step 1.3.1: Perform a 3×3×3 sparse convolution on the input sparse voxel features to aggregate the local neighborhood geometric context without changing the sparse coordinate distribution, thus obtaining local spatial features.

[0071] Step 1.3.2: Feed the local spatial features into a low-dimensional channel mapping structure consisting of two 1×1×1 sparse convolutions. First, compress the channel dimension to extract a compact channel importance representation. Then, enhance the inter-channel dependencies through non-linear activation. Finally, restore the original channel dimension and generate channel-spatial attention weights through Sigmoid activation.

[0072] Step 1.3.3: Multiply the channel-spatial attention weights with the original input sparse voxel features on a voxel-by-voxel and channel-by-channel basis to output the enhanced features after adaptive recalibration. (8); in, Represents local spatial features; Indicates channel-spatial attention weights; This indicates that the product of voxels and channels is multiplied. This indicates the enhanced features of the STCA output.

[0073] Example 6 This embodiment is based on an indoor 3D point cloud object detection method with sparse feature enhancement. Based on Embodiment 2, the specific method for step 1.4 is as follows: Step 1.4.1: Perform a 1×1×1 sparse convolution on the sparse features output from the deep layer to compress the channel dimension into a low-dimensional feature space:

[0074] Step 1.4.2: Perform 3×3×3 sparse convolution in the low-dimensional feature space to perform secondary local spatial modeling of deep semantic features, thereby strengthening the fine-grained geometric structure weakened by the downsampling process.

[0075] Step 1.4.3: Map the enhanced features back to the original channel dimensions using 1×1×1 sparse convolution, and then obtain the recovered features through batch normalization and non-linear activation.

[0076] Step 1.4.4: Add the recovered features to the module input features using residual summation to obtain the deep enhancement features output by the sparse local residual convolution module:

[0077] in, , representing the sparse voxel feature tensor output by the sparse residual feature extraction module. N This indicates the number of voxel features in the sparse residual feature extraction module of this layer. C Dimensions representing voxel features; These represent the low-dimensional channel features, geometric features, and restored features after convolution in the sparse residual feature extraction module, respectively. This represents an i×i×i sparse convolution operation; This represents the residual output feature tensor.

[0078] Example 7 This embodiment is based on an indoor 3D point cloud object detection method with sparse feature enhancement. Building upon Embodiment 1, step 2 is specifically described as follows: Step 2.1: Upsample deep sparse features using generative transpose convolution to explicitly generate new sparse spatial coordinates and restore spatial resolution; specifically: upsample the input sparse features using generative transpose convolution to restore spatial resolution and reconstruct the sparse coordinate structure, and then obtain aligned upsampled features through batch normalization and nonlinear activation:

[0079] in, Indicates transpose; This represents a generative convolution with a kernel size of 3×3×3 and a stride of 2. Step 2.2: The upsampled high-level semantic features are fused element-by-element with the corresponding layer's backbone network features to obtain cross-layer fused features. Specifically, the upsampled high-level semantic features are fused element-by-element with the downsampled features of the previous layer in the backbone network to obtain fused features at different scales.

[0080] Step 2.3: Locally reshape and smooth the cross-layer fusion features to eliminate semantic inconsistencies and feature aliasing caused by the direct superposition of features at different scales. Specifically, sparse convolution is performed on the cross-layer fusion features to re-aggregate features within the local receptive field, alleviating the scale differences and feature aliasing problems caused by cross-layer fusion, and obtaining stable reshaped features.

[0081] Step 2.4: Perform unified channel mapping on output features at different scales to embed them into the same feature space to adapt to the detection head; specifically, embed features at different scales into the same feature space and map them to the fixed input channel dimensions required by the detection head.

[0082] Step 2.5: Introduce the Channel Weight Attention Gating Module (CWAG) to perform semantic relabeling on the fused multi-scale features.

[0083] Unlike existing fusion networks that only perform feature space alignment but lack a channel filtering mechanism, this invention introduces a channel weight attention gating module (CWAG) to perform channel-level recalibration on the fused features. This achieves efficient semantic channel filtering without requiring spatial attention modeling. The specific method is as follows: Step 2.5.1: Perform global average pooling on the input fused features to compress the spatial dimension and obtain the channel description vector that reflects the overall semantic distribution.

[0084] Step 2.5.2: Model the nonlinear dependencies between channels using two fully connected layers (Linear1, Linear2), with ReLU activation to enhance the representation. Simultaneously, reduce the hidden layer dimension using a channel compression ratio r to control the number of parameters and computational cost while maintaining modeling capability. Finally, generate channel weights using Sigmoid activation.

[0085] in, This indicates that the number of output channel dimensions is A fully connected layer; Indicates the channel compression ratio; The number of channel dimensions is expressed as A fully connected layer; Step 2.5.3: Broadcast the channel weights to the corresponding positions of the fused features, and perform channel-by-channel weighting on the fused features to highlight key semantic channels and suppress redundant channel responses.

[0086] Example 8 This embodiment is based on an indoor 3D point cloud object detection method with sparse feature enhancement. Building upon Embodiment 1, the specific method for step 3 is as follows: Step 3.1, Design of a lightweight, anchor-free detection head using sparse convolution: Step 3.1.1: Using the multi-scale fused voxel features output by the neck network as input, the features are fed into the bounding box regression branch and the category classification branch respectively. The two branches share the input features and are responsible for the 3D bounding box parameter prediction and target category prediction tasks respectively. Step 3.1.2: Using a point-based regression approach, with each valid voxel as a candidate center, the 6-dimensional size parameters of the center offset and axis-aligned 3D bounding box are directly predicted through sparse convolution. The axis alignment IoU loss is only optimized for positive samples. Step 3.1.3: Output the class confidence logits for each voxel through sparse convolution, convert them into class probabilities through Sigmoid activation, and use Focal Loss during training to alleviate the class imbalance problem; Step 3.1.4: After completing the predictions on features at different scales, the prediction results at each scale are fused together in the inference stage to output the final target category and 3D bounding box, thereby achieving efficient end-to-end prediction of sparse voxel features. Step 3.2: Multi-condition joint constraint label assignment strategy. The specific method is as follows: Step 3.2.1: Based on the scale differences of different target categories, preset the mapping relationship between categories and feature levels, retain only the points in the voxel level that match the target category as candidate positive samples, and filter out unmatched samples from the level dimension. Step 3.2.2: For each real target, select the K points closest to the target center from the candidate points that meet the hierarchical constraints as positive samples to stabilize the number of positive samples and alleviate the sample imbalance problem caused by scale differences and point cloud sparsity. Step 3.2.3: For voxel points that match multiple targets simultaneously, assign them to the nearest target to ensure that each voxel point corresponds to only one true bounding box, avoid supervision conflicts, and improve the stability of label assignment.

Claims

1. A method for indoor 3D point cloud object detection based on sparse feature enhancement, characterized in that, Specifically as follows: Step 1: Construct a feature extraction network based on lightweight spatial-channel joint enhancement, perform multi-stage sparse feature extraction on voxelized point clouds, introduce sparse tensor channel-spatial attention (STCA) in the shallow layer, and introduce sparse local residual convolution (SLR-Conv) in the deep layer. Step 2: Construct a multi-scale feature fusion network based on channel recalibration. The multi-scale sparse features output by the backbone network are fused from top to bottom. The fused features are then introduced into a channel weight attention gating module for channel-level weight allocation and semantic recalibration. Step 3: Input the multi-scale fusion features into the dual-branch detection head, and perform 3D bounding box regression and category classification respectively to obtain the indoor 3D point cloud object detection results.

2. The method of claim 1, wherein, The specific method for step 1 is as follows: Step 1.1: Construct an initial sparse feature encoding layer. Perform local spatial aggregation, channel expansion, and structured downsampling on the low-dimensional input features of the voxelized point cloud to obtain the initial sparse features. Step 1.2: Construct a multi-stage sparse residual feature extraction layer, and form sparse feature representations of different scales by downsampling layer by layer and connecting with the residual; Step 1.3: In the low-to-mid-level feature extraction stage, the Sparse Tensor Channel-Spatial Attention Module (STCA) is introduced to co-model the local geometry and channel response, generate selective attention weights, and adaptively recalibrate the original sparse features. Step 1.4: In the deep feature extraction stage, a sparse local residual convolution module (SLR-Conv) is introduced to perform local spatial reconstruction in the low-dimensional channel space, and to achieve the refinement and stable enhancement of deep semantic features through residual connections.

3. The method of claim 2, wherein the method further comprises: The specific method for step 1.1 is as follows: Step 1.1.1: Perform sparse voxelization on the input point cloud to obtain a sparse tensor containing effective voxel coordinates and initial voxel features; Step 1.1.2: Use a sparse convolution with a kernel size of 3×3×3 and a stride of 2 to perform local neighborhood feature aggregation and channel mapping on the sparse tensor to complete the initial spatial downsampling; （1） wherein, represents inputting a sparse tensor; represents a sparse convolution operation, which is specifically a 2x spatial down-sampling convolution with a kernel size of 3x3x3 and a step size of 2 Step 1.1.3: Perform batch normalization and non-linear activation on the convolutional output sequentially, and then perform structured compression through sparse max pooling to form the initial sparse feature representation: （2） in, Indicates the initial extracted features; This indicates a batch normalization operation; This indicates a max pooling operation, where the maximum pooling step size and receptive field are 2, achieving 2x spatial downsampling. This represents the activation function.

4. The method of claim 3, wherein, The specific method for step 1.2 is as follows: Step 1.2.1: Construct multiple consecutive sparse residual feature extraction stages, each stage consisting of several sparse residual feature extraction modules, gradually expanding the receptive field and improving semantic expressive power: in, This represents the feature extraction layer in the i-th stage; This represents the output feature of the i-th layer; This represents the N sparse residual feature extraction downsampling modules of the i-th stage feature extraction layer; Step 1.2.2: In each sparse residual feature extraction module, local neighborhood feature aggregation and semantic enhancement are completed through sparse convolution, batch normalization, and nonlinear activation. Input features and transformed features are then fused through residual connections. wherein, denotes input features; denotes feature semantic deepening branch; denotes spatial resolution down-sampling branch; denotes 3x3x3 sparse convolution operation; Step 1.2.3: In the first sparse residual feature extraction module of each stage, set up a downsampling branch. When the stride is greater than 1, use convolutional downsampling mapping instead of identity mapping. Complete the reduction of spatial resolution and adjustment of channel dimensions.

5. The method of claim 4, wherein, The specific method for step 1.3 is as follows: Step 1.3.1: Perform a 3×3×3 sparse convolution on the input sparse voxel features to aggregate the local neighborhood geometric context without changing the sparse coordinate distribution, thus obtaining local spatial features. Step 1.3.2: Feed the local spatial features into a low-dimensional channel mapping structure consisting of two 1×1×1 sparse convolutions. First, compress the channel dimension to extract a compact channel importance representation. Then, enhance the inter-channel dependencies through non-linear activation. Finally, restore the original channel dimension and generate channel-spatial attention weights through Sigmoid activation. Step 1.3.3: Multiply the channel-spatial attention weights with the original input sparse voxel features on a voxel-by-voxel and channel-by-channel basis to output the enhanced features after adaptive recalibration. in, Represents local spatial features; Indicates channel-space attention weights; This indicates that the product of voxels and channels is multiplied. This indicates the enhanced features of the STCA output.

6. The method of claim 5, wherein, The specific method for step 1.4 is as follows: Step 1.4.1: Perform a 1×1×1 sparse convolution on the sparse features output from the deep layer to compress the channel dimension into a low-dimensional feature space: Step 1.4.2: Perform 3×3×3 sparse convolution in the low-dimensional feature space to perform secondary local spatial modeling of deep semantic features, thereby strengthening the fine-grained geometric structure weakened by the downsampling process. Step 1.4.3: Map the enhanced features back to the original channel dimensions using 1×1×1 sparse convolution, and then obtain the recovered features through batch normalization and non-linear activation. Step 1.4.4: Add the recovered features to the module input features using residual summation to obtain the deep enhancement features output by the sparse local residual convolution module: in, , representing the sparse voxel feature tensor output by the sparse residual feature extraction module. N This indicates the number of voxel features in the sparse residual feature extraction module of this layer. C Dimensions representing voxel features; These represent the low-dimensional channel features, geometric features, and restored features after convolution in the sparse residual feature extraction module, respectively. This represents an i×i×i sparse convolution operation; This represents the residual output feature tensor.

7. The method of claim 1, wherein, The specific method for step 2 is as follows: Step 2.1: Upsample the input sparse features using generative transpose convolution, and then obtain aligned upsampled features through batch normalization and non-linear activation: wherein denotes the transpose; denotes a generative convolution with kernel size 3x3x3 and stride 2; Step 2.2: Fuse the upsampled high-level semantic features with the corresponding downsampled features from the previous layer in the backbone network element by element to obtain fused features at different scales. Step 2.3: Perform sparse convolution on the cross-layer fusion features to re-aggregate features within the local receptive field, alleviating the scale differences and feature aliasing problems caused by cross-layer fusion, and obtaining stable reconstituted features: Step 2.4: Embed features of different scales into the same feature space and map them to the fixed input channel dimension required by the detection head. Step 2.5: Introduce the Channel Weight Attention Gating Module (CWAG) to perform semantic relabeling on the fused multi-scale features.

8. The method of claim 7, wherein, The specific method for step 2.5 is as follows: Step 2.5.1: Perform global average pooling on the input fused features to compress the spatial dimension and obtain the channel description vector that reflects the overall semantic distribution. Step 2.5.2: Model the nonlinear dependency between channels using two fully connected layers, enhance the representation with ReLU activation in between, and reduce the hidden layer dimension with a channel compression ratio r to control the number of parameters and computational cost while ensuring modeling capability. Finally, generate channel weights with Sigmoid activation. in, This indicates that the number of output channel dimensions is A fully connected layer; Indicates the channel compression ratio; The number of channel dimensions is expressed as A fully connected layer; Step 2.5.3: Broadcast the channel weights to the corresponding positions of the fused features, and perform channel-by-channel weighting on the fused features to highlight key semantic channels and suppress redundant channel responses.

9. The method of claim 1, wherein, The specific method for step 3 is as follows: Step 3.1, Design of a lightweight, anchor-free detection head using sparse convolution: Step 3.1.1: Using the multi-scale fused voxel features output by the neck network as input, the features are fed into the bounding box regression branch and the category classification branch respectively. The two branches share the input features and are responsible for the 3D bounding box parameter prediction and target category prediction tasks respectively. Step 3.1.2: Using a point-based regression approach, with each valid voxel as a candidate center, the 6-dimensional size parameters of the center offset and axis-aligned 3D bounding box are directly predicted through sparse convolution. The axis alignment IoU loss is only calculated for positive samples and optimized. Step 3.1.3: Output the class confidence logits for each voxel through sparse convolution, convert them into class probabilities through Sigmoid activation, and use Focal Loss during training to alleviate the class imbalance problem; Step 3.1.4: After completing the predictions on features at different scales, the prediction results at each scale are fused together in the inference stage to output the final target category and 3D bounding box, thereby achieving efficient end-to-end prediction of sparse voxel features. Step 3.2, Multi-condition joint constraint label allocation strategy.

10. The method of claim 9, wherein, The specific method for step 3.2 is as follows: Step 3.2.1: Based on the scale differences of different target categories, preset the mapping relationship between categories and feature levels, retain only the points in the voxel level that match the target category as candidate positive samples, and filter out unmatched samples from the level dimension. Step 3.2.2: For each real target, select the K points closest to the target center from the candidate points that meet the hierarchical constraints as positive samples to stabilize the number of positive samples and alleviate the sample imbalance problem caused by scale differences and point cloud sparsity. Step 3.2.3: For voxel points that match multiple targets simultaneously, assign them to the nearest target to ensure that each voxel point corresponds to only one true bounding box, avoid supervision conflicts, and improve the stability of label assignment.

Images