A three-dimensional target detection method, device, equipment and storage medium

Abstract

The application discloses a three-dimensional target detection method, device and equipment and a storage medium. The method comprises the following steps: acquiring three-dimensional point cloud data collected by a vehicle-mounted laser radar, pre-processing the three-dimensional point cloud data, and performing multi-resolution space grid division to obtain a multi-scale division result; performing lightweight feature coding on the multi-scale division result to generate a multi-scale pseudo-image tensor, and obtaining a multi-level feature map after backbone feature extraction; performing bidirectional cross-scale feature pyramid on the multi-level feature map to complete preliminary feature fusion, and then sequentially performing parallel channel and spatial attention calibration, cross-scale attention fusion and level propagation to obtain an enhanced multi-scale feature map; and finally outputting the category, spatial position, size and orientation information of the three-dimensional target by a lightweight detection head on the enhanced multi-scale feature map. The application improves the inference speed on the vehicle-mounted platform, has excellent comprehensive average detection accuracy, and effectively balances the detection real-time performance and accuracy.

Description

Technical Field

[0001] This application relates to the fields of autonomous driving and intelligent transportation environment perception technology, and in particular to a three-dimensional target detection method, device, equipment and storage medium. Background Technology

[0002] As a core functional module of vehicle-mounted LiDAR environmental perception systems, 3D target detection's real-time performance and accuracy directly determine the vehicle's path planning and obstacle avoidance capabilities. Vehicle-mounted LiDAR can acquire high-precision 3D point cloud data independent of lighting conditions, providing reliable stereo geometric information for environmental perception. However, the inherent sparsity and non-uniformity of point cloud data make the features of distant targets and small objects easily obscured by background noise. Simultaneously, vehicle-mounted edge computing platforms generally suffer from limited computing power and tight memory bandwidth, placing stringent demands on the deployment efficiency of the perception system.

[0003] Currently, mainstream 3D object detection algorithms still face significant challenges in balancing accuracy and efficiency. Voxel-based methods consume enormous amounts of computation and memory, making real-time operation on embedded platforms difficult; point-based processing paradigms experience exponentially increasing computational demands in large-scale scenes. While bird's-eye view projection methods, such as PointPillars, significantly improve inference speed, their original architecture suffers from obvious parameter and computational redundancy. Furthermore, the collapse of point clouds into a 2D plane results in the loss of fine geometric details in the vertical direction, and insufficient single-scale feature representation leads to poor detection performance for distant small targets and heavily occluded targets. Conventional improvement schemes that merely increase network depth or width to enhance performance revert to a state of high redundancy, contradicting the initial goal of lightweight deployment.

[0004] Therefore, how to maintain an extremely high inference frame rate while removing the structural redundancy of the original network and compensating for the information loss caused by dimensionality compression, and build a 3D object detection system that combines lightweight deployment characteristics with the ability to perceive complex scenes at multiple scales, is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the three-dimensional target detection method, apparatus, device, and storage medium provided in this application embodiment can maintain an extremely high inference frame rate while removing structural redundancy from the original network and compensating for information loss caused by dimensionality compression, thus constructing a three-dimensional target detection system that combines lightweight deployment characteristics with multi-scale complex scene perception capabilities. The three-dimensional target detection method, apparatus, device, and storage medium provided in this application embodiment are implemented as follows: This application provides a three-dimensional target detection method, including: Acquire 3D point cloud data collected by vehicle-mounted LiDAR, preprocess the 3D point cloud data and perform multi-resolution spatial mesh division to obtain multi-scale division results; The multi-scale segmentation results are subjected to lightweight feature encoding processing to obtain a multi-scale pseudo-image tensor; The multi-scale pseudo-image tensor is subjected to backbone feature extraction processing to obtain a multi-level feature map; The multi-level feature maps are subjected to bidirectional cross-scale feature pyramid construction and fusion processing to obtain initial fused features; The initial fused features are subjected to parallel channel and spatial attention calibration to obtain calibrated multi-scale features; The calibrated multi-scale features are subjected to cross-scale attention fusion and hierarchical propagation processing to obtain an enhanced multi-scale feature map; The enhanced multi-scale feature map is subjected to lightweight detection head inference processing to obtain the category, spatial location, size, and orientation information of the three-dimensional target.

[0006] In some embodiments, the lightweight feature encoding process performed on the multi-scale segmentation result to obtain a multi-scale pseudo-image tensor includes: The geometric center offset and feature dimensionality reduction are performed on the multi-scale partitioning results to obtain a seven-dimensional point feature vector; The seven-dimensional point feature vector is subjected to high-dimensional feature mapping to obtain the mapped point features; The mapped point features are subjected to max pooling aggregation to obtain a feature vector of fixed length; The fixed-length feature vectors are subjected to two-dimensional grid spatial scattering processing to obtain multi-scale pseudo-image tensors corresponding to different resolutions.

[0007] In some embodiments, the process of extracting backbone features from the multi-scale pseudo-image tensor to obtain multi-level feature maps includes: The multi-scale pseudo-image tensor is subjected to depthwise separable convolution feature extraction to obtain initial backbone features; The initial backbone features are downsampled to obtain the downsampled backbone features; The downsampled backbone features are subjected to global average pooling to obtain channel descriptors; The channel descriptors are processed by channel relationship learning to obtain channel attention weight vectors; The downsampled backbone features and the channel attention weight vector are subjected to feature recalibration processing to obtain multi-level feature maps of three different levels.

[0008] In some embodiments, the bidirectional cross-scale feature pyramid construction and fusion process of the multi-level feature maps to obtain initial fused features includes: Semantic information transfer processing is performed on the multi-level feature maps to obtain semantically enhanced shallow features; Geometric detail transfer processing is performed on the multi-level feature maps to obtain deep features with enhanced details; The semantically enhanced shallow features and the detail-enhanced deep features are subjected to fast normalization and weighted fusion processing to obtain the initial fused features.

[0009] In some embodiments, the parallel channel and spatial attention calibration process performed on the initial fused features to obtain calibrated multi-scale features includes: The initial fusion features are subjected to efficient channel attention processing to obtain a channel weight map; The initial fusion features are subjected to lightweight spatial attention processing to obtain a spatial weight map; Adaptive weighted fusion processing is performed on the initial fusion features, the channel weight map, and the spatial weight map to obtain calibrated multi-scale features.

[0010] In some embodiments, the step of performing cross-scale attention fusion and hierarchical propagation processing on the calibrated multi-scale features to obtain an enhanced multi-scale feature map includes: The calibrated multi-scale features are subjected to bilinear interpolation resolution alignment processing to obtain multi-scale features with uniform resolution. Global average pooling and concatenation are performed on the multi-scale features of the unified resolution to obtain the scale description matrix; The scale description matrix is ​​subjected to self-attention calculation to obtain the initial fusion weights of features at different scales; The initial fusion weights are subjected to hierarchical attention propagation processing to obtain the corrected fusion weights; The multi-scale features with the unified resolution and the modified fusion weights are weighted and fused, and residual connection processing is performed to obtain an enhanced multi-scale feature map.

[0011] This application provides a three-dimensional target detection device, comprising: The acquisition module is used to acquire three-dimensional point cloud data collected by the vehicle-mounted LiDAR, and to preprocess the three-dimensional point cloud data and perform multi-resolution spatial mesh division processing to obtain multi-scale division results. The processing module is used to perform lightweight feature encoding processing on the multi-scale segmentation results to obtain a multi-scale pseudo-image tensor. The processing module is also used to perform backbone feature extraction processing on the multi-scale pseudo-image tensor to obtain a multi-level feature map. The processing module is also used to perform bidirectional cross-scale feature pyramid construction and fusion processing on the multi-level feature maps to obtain initial fused features; The processing module is also used to perform parallel channel and spatial attention calibration on the initial fused features to obtain calibrated multi-scale features; The processing module is also used to perform cross-scale attention fusion and hierarchical propagation processing on the calibrated multi-scale features to obtain an enhanced multi-scale feature map. The inference module is used to perform lightweight detection head inference processing on the enhanced multi-scale feature map to obtain the category, spatial location, size, and orientation information of the three-dimensional target.

[0012] The computer device provided in this application includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the method described in this application.

[0013] The computer-readable storage medium provided in this application embodiment stores a computer program thereon, which, when executed by a processor, implements the method provided in this application embodiment.

[0014] This application provides a 3D target detection method, apparatus, device, and storage medium. It acquires 3D point cloud data from a vehicle-mounted LiDAR, preprocesses the point cloud data, and performs multi-resolution spatial mesh partitioning to obtain multi-scale partitioning results. The multi-scale partitioning results are then subjected to lightweight feature encoding to generate multi-scale pseudo-image tensors. After backbone feature extraction, multi-level feature maps are obtained. A bidirectional cross-scale feature pyramid is used to perform preliminary feature fusion on the multi-level feature maps. Parallel channel and spatial attention calibration, cross-scale attention fusion, and hierarchical propagation are then performed sequentially to obtain enhanced multi-scale feature maps. Finally, a lightweight detection head outputs the category, spatial location, size, and orientation information of the 3D target from the enhanced multi-scale feature maps. This approach maintains an extremely high inference frame rate while eliminating structural redundancy in the original network and compensating for information loss due to dimensionality compression. It constructs a 3D target detection system that combines lightweight deployment with multi-scale complex scene perception capabilities, thus solving the technical problems mentioned in the background art. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram illustrating the implementation process of a three-dimensional target detection method provided in an embodiment of this application; Figure 2 This application provides a schematic diagram illustrating the implementation process for obtaining multi-scale pseudo-image tensors in an embodiment of the present application. Figure 3 This is a schematic diagram of the structure of a three-dimensional target detection device provided in an embodiment of this application; Figure 4 This application provides an overall architecture diagram of a three-dimensional target detection method according to an embodiment of the present application. Figure 5 A lightweight three-level strategy network diagram of "encoding-backbone-detection" is provided for embodiments of this application; Figure 6 A spatial partitioning diagram of multi-resolution point columns is provided for embodiments of this application; Figure 7 An architectural diagram of a bidirectional cross-scale feature pyramid provided in this application embodiment; Figure 8 This is a network architecture diagram for cross-scale attention feature fusion provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] The following description of some technologies involved in the embodiments of this application is provided to aid understanding and should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, some descriptions of well-known functions and structures are omitted in the following description.

[0019] The overall architecture of a three-dimensional target detection method provided in this application embodiment is as follows: Figure 4 As shown, the process comprises two core stages: Stage 1 is the lightweight network stage, which includes lightweight Pillar feature encoding, a 2D deep separable backbone network, and a simplified detection head; Stage 2 is the feature enhancement network stage, which includes multi-scale feature extraction and fusion, parallel channels and spatial attention mechanisms, and cross-scale attention feature fusion. After the raw point cloud data undergoes these two stages of pipeline processing, the final output is 3D bounding box information such as the target's category, location, size, and orientation.

[0020] The three-level lightweight strategy network of "encoding-backbone-detection" in this application embodiment is as follows: Figure 5 As shown, the input point cloud first generates a pseudo image through a Pillar coding network with point feature dimensionality reduction and shared weights, then enters a backbone network based on depthwise separable convolution, and finally outputs the detection result through a lightweight detection head with a shared feature extraction layer.

[0021] Figure 1 This is a schematic diagram illustrating the implementation flow of a three-dimensional target detection method provided in an embodiment of this application, including steps 101 to 107. Wherein, Figure 1 This is merely one execution order shown in the embodiments of this application and does not represent the only execution order for a 3D target detection method. The execution order can be adjusted to achieve the desired final result. Figure 1 The steps shown can be performed in parallel or in reverse order.

[0022] Step 101: Acquire the 3D point cloud data collected by the vehicle-mounted LiDAR, perform preprocessing and multi-resolution spatial mesh generation on the 3D point cloud data to obtain multi-scale partitioning results.

[0023] In this embodiment, the 3D point cloud data of the current environment is acquired in real time by an onboard LiDAR. The input point cloud set is shown in the formula:

[0024] in, For the i-th lidar point, The input is a set of 3D point clouds (all LiDAR points). The total number of point clouds, Let these be the physical coordinates of a point in three-dimensional space. The laser reflection intensity is set. An effective detection range is defined for the input point cloud, filtering out invalid point cloud data exceeding this range. Simultaneously, fine-grained, medium-grained, and coarse-grained Pillar meshes of three different sizes are constructed on the bird's-eye view plane. Multi-resolution Pillar spatial partitioning is as follows: Figure 6 As shown, fine-grained meshes are dedicated to capturing the fine local geometry of small targets such as pedestrians and cyclists in the distance, medium-grained meshes achieve a basic balance between detail preservation and computational overhead, and coarse-grained meshes provide a larger initial receptive field for large vehicles in the foreground.

[0025] For the species scale ( The spatial resolution of the pseudo-image feature map it generates The calculation is shown in the formula:

[0026] in, The function is an up-rounding function to ensure that the point cloud at the boundary can be completely mapped. For the first The height of the pseudo-image at various scales, No. The width of the pseudo-image at each scale. Define the effective detection range for the input point cloud: X-axis Y-axis . For the first The Pillar mesh size introduces an adaptive threshold based on height and laser reflection intensity to discard noise points that clearly belong to flat ground or low-altitude background in advance, and limits the number of maximum non-empty Pillars at the three scales to avoid memory overflow.

[0027] Step 102: Perform lightweight feature encoding on the multi-scale segmentation results to obtain the multi-scale pseudo-image tensor.

[0028] In this embodiment, for all points within each non-empty Pillar, the offset relative to the geometric center of all points within that Pillar is calculated, generating a seven-dimensional point feature vector. The absolute offsets from the grid center that have low signal-to-noise ratios, as used in the original method, are then removed. The formula is as follows:

[0029] in, It is the arithmetic mean (geometric center) of all points within the current Pillar. This is the local offset of the point relative to the geometric center within Pillar.

[0030] The seven-dimensional point feature vectors within all Pillars across three scales are uniformly fed into a two-layer fully connected network with shared parameters for high-dimensional feature mapping, avoiding a surge in front-end parameters due to multi-scale partitioning. Max pooling is then applied to the point features within each mapped Pillar, aggregating them into a fixed-length Pillar feature vector. This operation ensures the model's permutation invariance to the point cloud input order. Assuming each Pillar retains T valid points, after independent mapping, they are aggregated into a fixed-length Pillar feature vector through symmetric max pooling, as shown in the formula:

[0031] Where C represents the number of feature channels, which is set to 64 in this paper. The Pillar feature encoding network consists of two fully connected layers (containing BN and ReLU), mapping each 7-dimensional feature point to... High-dimensional features. MaxPool takes the maximum value for each channel along the number of points.

[0032] Based on the original spatial coordinates of each Pillar, the corresponding Pillar feature vector is distributed into a two-dimensional grid, forming three dense pseudo-image tensors with different resolutions.

[0033] Step 103: Perform backbone feature extraction on the multi-scale pseudo-image tensor to obtain multi-level feature maps.

[0034] In this embodiment, multi-scale pseudo-image tensors are input into a lightweight two-dimensional backbone network constructed based on depthwise separable convolution. Depthwise separable convolution completely replaces the standard convolution operation in the backbone network, decomposing the convolution process into two steps: spatial filtering and cross-channel feature mixing, which significantly reduces the computational load.

[0035] Let the size of the input feature map be... The number of input channels and the number of output channels are respectively and The kernel size is The floating-point computation cost of standard convolution is shown in the formula:

[0036] Depthwise separable convolution decomposes it into two steps: depthwise convolution (spatial filtering) and pointwise convolution (cross-channel blending). The computational complexity formula is shown below:

[0037] The formula for the compression ratio is shown below:

[0038] The backbone network contains multiple downsampling blocks, with a lightweight channel attention module embedded after each downsampling block. Global average pooling is performed on the feature maps input to the channel attention modules to obtain channel descriptors. The formula for obtaining the channel descriptors by applying global average pooling to the input feature maps is shown below:

[0039] in, For the input feature map at the th On each channel, the spatial coordinates are The characteristic response value at that location.

[0040] After two layers of fully connected excitation networks (compression ratio) Learning channel relationships to generate attention weight vectors The formula is shown below:

[0041] in, It is Sigmoid. For a reduced-dimensional matrix, This is an upgraded matrix. ReLU is a non-linear activation function. Finally, feature recalibration is performed. . This represents a channel-wise multiplication operation in the spatial dimension. The downsampled backbone features are adaptively recalibrated using channel attention weight vectors, ultimately outputting three multi-level feature maps at different levels, corresponding to different resolutions and semantic depths.

[0042] Step 104: Perform bidirectional cross-scale feature pyramid construction and fusion processing on the multi-level feature maps to obtain the initial fused features.

[0043] In this embodiment, a bidirectional cross-scale feature pyramid is constructed based on multi-level feature maps, and its architecture is as follows: Figure 7 As shown, it contains two parallel paths: a top-down semantic enhancement path and a bottom-up detail enhancement path.

[0044] A top-down semantic enhancement path is constructed, which transfers strong semantic information contained in deep features to shallow features through upsampling operations, thereby enhancing the semantic expressive power of shallow features. A bottom-up detail enhancement path is constructed, which injects fine geometric boundary information contained in shallow features into deep features through downsampling operations, thereby improving the localization accuracy of deep features. At each fusion node of the bidirectional path, a fast normalization weighting strategy is adopted, which autonomously allocates the relative importance of features at different scales through learnable scalar weights, replacing the traditional element addition or concatenation method.

[0045] During the fusion process, traditional element addition or concatenation is abandoned, and a lightweight, fast normalization weighting strategy is introduced. For output nodes... Assuming the input feature set is The fusion formula is as follows:

[0046] in, For learnable scalar weights activated by ReLU, constant This mechanism ensures numerical stability, enabling the network to autonomously learn the relative importance of features at different scales. For the first to participate in the integration Input feature maps.

[0047] Step 105: Perform parallel channel and spatial attention calibration on the initial fused features to obtain calibrated multi-scale features.

[0048] In this embodiment, parallel channel and spatial attention calibration is performed on the initial fused features at each feature node of the bidirectional cross-scale feature pyramid. Global average pooling is performed on the initial fused features through an efficient channel attention branch, and one-dimensional convolution is used to extract cross-channel dependencies and generate a channel weight map. The results of max pooling and average pooling along the channel dimension are calculated separately by a lightweight spatial attention branch and concatenated, then a spatial weight map is generated through depthwise separable convolution. The output features of the channel attention branch and the spatial attention branch are balanced using learnable scalar weights, and the initial fused features are adaptively weighted and fused.

[0049] For the input feature map Minimize parameter overhead. Perform global average pooling to obtain channel descriptors. Then, use an adaptive kernel size of One-dimensional convolution extracts channel dependencies, as shown in the following formula:

[0050] in , This means taking the nearest odd number. This represents the number of feature channels in the current layer. Weights are generated and calibrated as shown in the following formula:

[0051]

[0052] in The Sigmoid activation function is given by the formula: Map the output to The interval serves as a gating mechanism. This is a one-dimensional convolution operation that performs sliding window computation along the channel dimension to capture local cross-channel dependencies. This is the final channel attention weight vector. This is an element-wise multiplication along the channel dimension. The output feature map after channel attention weighting.

[0053] Max pooling and average pooling are calculated along the channel dimension, concatenated, and then fed into the input. Depth-separable convolutions generate spatial weight maps. The formula is shown below:

[0054]

[0055] in The input is the original feature map. To perform global average pooling on the feature map along the channel dimension, To perform global max pooling on the feature map along the channel dimension, This is for channel splicing operations. For use Depth convolution of convolution kernels. For position The corresponding spatial attention weights. For the original feature map at location eigenvalues ​​at that location This represents the output feature value at position (i,j) after spatial attention weighting.

[0056] Utilizing learnable scalar weights (Initially 0.5) Balance the two attention features, as shown in the formula below:

[0057] in, This is the final attention-enhanced output feature map.

[0058] Step 106: Perform cross-scale attention fusion and hierarchical propagation processing on the calibrated multi-scale features to obtain an enhanced multi-scale feature map.

[0059] In this embodiment, bilinear interpolation is performed on the calibrated multi-scale features to uniformly align feature maps of different scales to a medium resolution. Global average pooling is then performed on the feature maps of each scale after unifying the resolution, compressing them into corresponding feature description vectors. All feature description vectors are then concatenated to form a scale description matrix. The dependencies between features of different scales are calculated based on a self-attention mechanism, and initial fusion weights for features of different scales are dynamically allocated. A hierarchical attention propagation mechanism is introduced so that the attention calculation of the current layer is affected by the attention information of the previous layer. The contribution of the attention information of the previous layer is adjusted by an attenuation coefficient to obtain the corrected fusion weights. The multi-scale features of the unified resolution are weighted and fused according to the corrected fusion weights and connected with the residual branch features.

[0060] Network architectures for cross-scale attention feature fusion, such as Figure 8 As shown.

[0061] Feature maps of different scales Uniform alignment to medium resolution was achieved through bilinear interpolation. Global average pooling is used to compress the vector into a description vector. spliced ​​into a matrix The self-attention mechanism across computational scales is illustrated in the following formula:

[0062]

[0063] in, These are three learnable linear transformation matrices used to map input features to different spaces. The query vector determines the information we are interested in. This is a key vector that describes information at each position. This is a value vector representing the features that actually carry information, which will be weighted and summed according to the attention weights later. This represents the attention matrix, where each element indicates the degree of association between two locations; a larger value indicates a stronger association. Normalize the scaled similarity matrix, converting the weight of each position into... The values ​​between these ranges, with all weights summing to 1, yield the final attention weight matrix. .

[0064] Feature weight generation The fusion weights are obtained by row normalization. .

[0065] To ensure consistency in cross-level feature fusion, a hierarchical attention propagation mechanism is introduced, namely the first... The attention matrix of the first layer is affected by the second layer. The effect of the layer (attenuation coefficient) The formula is shown below:

[0066] in, This is the updated attention weight matrix for the k-th layer (or the k-th iteration).

[0067] The final fused features adopt a residual structure, as shown in the following formula: .

[0068] in, This is the final output feature map after cross-scale attention fusion. For the first Aligned feature maps at various scales These are learnable weights used to dynamically adjust the contribution ratio of features at different scales to the final result. For the residual connection part, for the input features The purpose of performing a 1×1 convolution is to adjust the number of channels so that it matches the feature dimension after the previous weighted summation, thereby achieving residual connections.

[0069] Step 107: Perform lightweight detection head inference processing on the enhanced multi-scale feature map to obtain the category, spatial location, size, and orientation information of the three-dimensional target.

[0070] In this embodiment, the total loss function consists of three parts, with weight ratios set as follows: The formula is shown below:

[0071] in, This represents the total loss function value during model training. These are the weighting coefficients for classification loss, regression loss, and directional loss, respectively. Classification loss for the target For the target position regression loss, Classify the loss according to the target direction.

[0072] FocalLoss is used to alleviate the extreme imbalance problem in scenes where there are far more background points than foreground objects, forcing the network to focus on hard samples. The formula is shown below:

[0073] set up It is a commonly used hyperparameter combination in object detection tasks. It is the Focal Loss function, which is specifically designed to solve the problem of extreme imbalance between positive and negative samples in object detection. This represents the model's predicted probability for the target category. This is the positive and negative sample balancing coefficient, used to reduce the weight of a large number of background samples.

[0074] To eliminate the ambiguity of the regression angle period (i.e., the regression jump caused by a 180-degree difference between the target's forward and reverse directions), a sinusoidal difference angle decomposition encoding is used, and the loss calculation is transformed using the trigonometric identity, as shown in the following formula:

[0075] in, The target orientation angle predicted by the model. This represents the actual target orientation angle.

[0076] Predictions and targets are regressed separately. and This ensures that the loss tends to be 0 when the angle difference is 0 or 180 degrees.

[0077] Binary cross-entropy (BCELoss) is used specifically to help distinguish the absolute forward and reverse orientation of a target.

[0078] The AdamW optimizer is used, with weight decay set to 0.01 to suppress overfitting. The learning rate scheduling employs the OneCycleLR strategy: in the first 40% of iterations, the learning rate is... Linear preheating to peak During this stage, the network quickly crosses local optima; subsequently, in the remaining 60% of the steps, it undergoes smooth annealing using a cosine function to decay the data until it approaches zero, ensuring final fine convergence. Data augmentation techniques include random flipping and global scaling. ), global rotation ( ) and translational disturbances.

[0079] This application's embodiments avoid the pitfalls of traditional methods that sacrifice efficiency for improved accuracy or excessively reduce performance for lightweight design by first systematically removing network redundancy to ensure real-time performance and then precisely enhancing accuracy through multi-scale and attention mechanisms. The solution is highly modular with smooth integration between components, enabling stable operation on resource-constrained vehicle platforms. It also possesses accurate perception capabilities for multiple types of targets in complex traffic scenarios, providing a practical solution for the large-scale mass production and deployment of autonomous driving LiDAR perception systems.

[0080] In the above Figure 1 Based on the above, this application embodiment also provides a schematic diagram of the implementation process for obtaining multi-scale pseudo-image tensors. For example... Figure 2 As shown, steps 201 to 204 are included: Step 201: Calculate the geometric center offset and perform feature dimensionality reduction on the multi-scale partitioning results to obtain a seven-dimensional point feature vector.

[0081] In this embodiment, each non-empty Pillar in the multi-scale partitioning result is traversed, and the arithmetic mean of all points within the Pillar is calculated to obtain the geometric center of the Pillar. For each point within the Pillar, its offset relative to the geometric center of the Pillar in the three coordinate axes of three-dimensional space is calculated. The initial features of the point are formed by combining the three-dimensional physical coordinates of each point and the laser reflection intensity. The absolute offsets of points with low signal-to-noise ratios relative to the grid center of the Pillar in the original method are removed, and the feature dimension of each point is reduced from nine dimensions to seven dimensions to obtain a seven-dimensional point feature vector, thereby reducing the parameter redundancy of subsequent fully connected layers.

[0082] Step 202: Perform high-dimensional feature mapping on the seven-dimensional point feature vector to obtain the mapped point features.

[0083] In this embodiment, a feature encoding network consisting of two fully connected layers is constructed, and the network is implemented using grouped convolution. The seven-dimensional point feature vectors within all Pillars at three scales—fine-grained, medium-grained, and coarse-grained—are uniformly fed into the same set of feature encoding networks for high-dimensional feature mapping. Through a parameter sharing mechanism, the problem of a surge in the number of front-end encoder parameters caused by multi-scale partitioning is avoided. Without significantly increasing the computational overhead, the low-dimensional point features are mapped to high-dimensional feature representations, resulting in the mapped point features.

[0084] Step 203: Perform max pooling on the mapped point features to obtain a feature vector of fixed length.

[0085] In this embodiment, a symmetric max pooling operation is performed on the high-dimensional features of all points within each non-empty Pillar after mapping. For each feature channel dimension, the maximum response value of all points within that Pillar in that channel is selected as the output value of that channel. Through the max pooling operation, the variable number of point features within each Pillar are aggregated into a Pillar feature vector of fixed length. This operation simultaneously ensures the permutation invariance of the model to the input order of the point cloud, eliminating the impact of the disorder of the point cloud data.

[0086] Step 204: Perform two-dimensional grid spatial scattering processing on the fixed-length feature vectors to obtain multi-scale pseudo-image tensors corresponding to different resolutions.

[0087] In this embodiment, based on the original spatial coordinates of each Pillar on the bird's-eye view plane, the corresponding Pillar feature vector is placed at the corresponding position on the two-dimensional grid. For grid positions without non-empty Pillars, zero-value features are filled. The above scattering operation is performed on the Pillar features at three scales: fine-grained, medium-grained, and coarse-grained, respectively, to finally form three dense two-dimensional pseudo-image tensors with different spatial resolutions, thus completing the conversion of three-dimensional point cloud data into two-dimensional pseudo-images.

[0088] This application's embodiments reduce the point feature dimension from nine to seven, eliminating the absolute offset term of the grid center with low signal-to-noise ratio, thus reducing unnecessary computation and noise interference. It employs an intra-group weight-sharing encoding network, avoiding the problem of a surge in front-end parameters caused by multi-scale partitioning, achieving multi-granularity feature extraction without significantly increasing computational overhead. Through max-pooling aggregation, it ensures the model's permutation invariance to the point cloud input order and converts variable-length point features into fixed-length vectors, laying the foundation for subsequent two-dimensional convolution processing. The resulting multi-scale pseudo-image tensor retains spatial geometric information of different granularities while possessing extremely low computational and storage overhead.

[0089] In some embodiments, the process of extracting backbone features from the multi-scale pseudo-image tensor to obtain a multi-level feature map includes: extracting depthwise separable convolutional features from the multi-scale pseudo-image tensor to obtain initial backbone features.

[0090] Specifically, multi-scale pseudo-image tensors are input into a lightweight 2D backbone network constructed based on depthwise separable convolutions. Depthwise separable convolutions completely replace all standard convolution operations in the backbone network. Each convolution operation is decomposed into two independent steps: depthwise convolution and pointwise convolution. Depthwise convolution is responsible for performing spatial filtering operations on each input channel individually, while pointwise convolution is responsible for performing cross-channel mixing operations on features from different channels. This decomposition method significantly reduces the computational load of the backbone network. Without significantly losing feature representation capabilities, it adapts to the computing power constraints of the vehicle-mounted edge computing platform to obtain the initial backbone features.

[0091] Furthermore, the initial backbone features are downsampled to obtain the downsampled backbone features.

[0092] Specifically, the backbone network is configured with three consecutive downsampling blocks. Each downsampling block halves the feature map spatial size and doubles the number of feature channels through a depthwise separable convolution with a stride of 2. Through stepwise downsampling operations, the receptive field of the features is gradually expanded, and feature information at different levels of abstraction is extracted. After the operation of the three downsampling blocks is completed in sequence, three backbone features with different spatial resolutions are obtained.

[0093] Furthermore, global average pooling is performed on the downsampled backbone features to obtain channel descriptors.

[0094] Specifically, a lightweight channel attention module is embedded at the output of each downsampling block to perform global average pooling on the downsampled backbone features output by the downsampling block; global mean calculation is performed on the spatial dimension of the feature map to compress the two-dimensional feature map of each channel into a single value, forming a channel descriptor with the same length as the number of feature channels, which is used to characterize the global feature information contained in each channel.

[0095] Furthermore, channel relationship learning is performed on the channel descriptors to obtain the channel attention weight vector.

[0096] Specifically, the channel descriptors are input into an activation network consisting of two fully connected layers. The first fully connected layer compresses the dimension of the channel descriptors to one-sixteenth of the original number of channels. After processing by a non-linear activation function, the second fully connected layer restores the dimension to the original number of channels. The two fully connected layers learn the interdependencies between the features of different channels and generate the attention weight value corresponding to each channel. The weight values ​​of all channels together constitute the channel attention weight vector.

[0097] Furthermore, the downsampled backbone features and channel attention weight vectors are subjected to feature recalibration processing to obtain multi-level feature maps of three different levels.

[0098] Specifically, the generated channel attention weight vector is multiplied channel by channel with the corresponding downsampled backbone feature map. The features of different channels are adaptively weighted and adjusted to enhance the channel feature response containing key information and suppress irrelevant channel features containing background noise. After completing the feature recalibration operation corresponding to the three downsampled blocks in sequence, three multi-level feature maps of different levels are finally output, corresponding to the shallow fine spatial detail features, the middle comprehensive semantic features, and the deep strong semantic abstract features, respectively.

[0099] This application's embodiments completely replace standard convolutions with depthwise separable convolutions, significantly reducing the computational load of the backbone network and enabling it to adapt to the computing power constraints of automotive edge platforms, thus ensuring the system's real-time inference capabilities. Simultaneously, a lightweight channel attention module is embedded after each downsampling block, effectively compensating for the weak channel interaction capabilities of depthwise separable convolutions. This module can adaptively learn the importance of different channel features, enhance the response of channels containing key target information, and suppress interference from background noise channels. The final output consists of three multi-level feature maps at different levels, respectively covering shallow fine spatial details, mid-level comprehensive semantics, and deep strong semantic abstraction information, providing a rich and high-quality feature foundation for subsequent cross-scale feature fusion.

[0100] In some embodiments, bidirectional cross-scale feature pyramid construction and fusion processing is performed on multi-level feature maps to obtain initial fused features, including: semantic information transfer processing is performed on multi-level feature maps to obtain semantically enhanced shallow features.

[0101] Specifically, based on the three different levels of multi-level feature maps output by the backbone network, a top-down semantic enhancement path is constructed. Starting from the deep, strong semantic features at the top layer, the spatial resolution is increased to match the resolution of the next layer feature map through upsampling. The upsampled deep semantic features are combined with the original features of the corresponding layer, enabling the shallow features to obtain deep global semantic information and enhancing the ability of the shallow features to distinguish the target semantic category. The semantic information is then passed layer by layer from the deepest to the shallowest layer to obtain semantically enhanced shallow features.

[0102] Furthermore, geometric detail transfer processing is performed on the multi-level feature maps to obtain deep features with enhanced details.

[0103] Specifically, a bottom-up detail enhancement path is constructed synchronously, starting from the shallow, fine geometric features at the bottom layer. The spatial resolution of these features is reduced to match the resolution of the feature map at the previous layer through downsampling. The downsampled shallow geometric details are then combined with the original features at the corresponding level, allowing the deep features to obtain the fine boundaries and precise location information of the shallow layer, thus improving the target localization accuracy of the deep features. This process is repeated layer by layer to transfer geometric details from the shallowest to the deepest layer, resulting in deep features with enhanced details.

[0104] Furthermore, the semantically enhanced shallow features and the detail-enhanced deep features are subjected to fast normalization and weighted fusion processing to obtain the initial fused features.

[0105] Specifically, at each fusion node in the bidirectional path, a fast normalization weighting strategy is adopted to replace the traditional element-wise addition or feature concatenation method; a learnable scalar weight is assigned to each feature input to the fusion node, and this weight participates in the fusion calculation after being activated by a linear rectified function, while a constant is introduced to ensure numerical stability; through the learnable scalar weight, the network can autonomously learn the relative importance of features at different scales in different scenarios; the above weighted fusion operation is performed on the semantic enhancement features and detail enhancement features at all levels, and finally the initial fusion features of the layer enhancement are obtained.

[0106] The bidirectional cross-scale feature pyramid constructed in this application addresses the problem of insufficient information transmission in traditional unidirectional feature pyramids. A top-down semantic enhancement path transmits strong semantic information from deep layers to shallow layers, improving the ability of shallow features to distinguish target categories. A bottom-up detail enhancement path injects fine geometric boundary information from shallow layers into deeper layers, improving the target localization accuracy of deep features. A fast normalized weighted fusion strategy replaces traditional element-wise addition or feature concatenation, enabling the network to autonomously learn the relative importance of features at different scales in different scenarios, making the fusion method more flexible and efficient. The resulting initial fused features possess both rich semantic information and precise spatial detail information, providing high-quality input for subsequent attention calibration.

[0107] In some embodiments, parallel channel and spatial attention calibration processing is performed on the initial fused features to obtain calibrated multi-scale features, including: performing efficient channel attention processing on the initial fused features to obtain a channel weight map.

[0108] Specifically, at each feature node of the bidirectional cross-scale feature pyramid, a global average pooling operation is performed on the initial fused features to compress the two-dimensional feature map of each channel into a global feature value, forming a channel descriptor. An efficient channel attention mechanism with minimal parameter overhead is adopted, using a one-dimensional convolution with an adaptive kernel size to process the channel descriptor and extract the interaction dependencies between different channels. Attention weight values ​​corresponding to each channel are generated through a non-linear activation function, and the weight values ​​of all channels together constitute a channel weight map, which is used to characterize the importance of different channel features.

[0109] Furthermore, the initial fused features are subjected to lightweight spatial attention processing to obtain a spatial weight map.

[0110] Specifically, for the same initial fused feature, global max pooling and global average pooling operations are performed along the channel dimension to obtain two single-channel feature maps. The two single-channel feature maps are then concatenated along the channel dimension to form a dual-channel feature map. The concatenated feature map is then input into a depthwise separable convolutional layer for processing to generate a single-channel spatial weight map, which is used to characterize the importance of different spatial locations in the feature map. This process uses depthwise separable convolution instead of standard convolution, which significantly reduces the computational cost while ensuring spatial attention effects.

[0111] Furthermore, the initial fusion features, channel weight map, and spatial weight map are subjected to adaptive weighted fusion processing to obtain calibrated multi-scale features.

[0112] Specifically, the channel weight map is multiplied channel-wise with the initial fused features to obtain a channel-weighted feature map; the spatial weight map is multiplied spatially with the initial fused features to obtain a spatially weighted feature map; a learnable scalar weight with an initial value of 0.5 is set, and this scalar weight is used to adaptively balance and fuse the channel-weighted feature map and the spatially weighted feature map; through the above operations, the model is guided to focus on the key region containing the foreground target from two orthogonal dimensions of channels and space, suppressing the interference of background noise, and finally obtaining the calibrated multi-scale features.

[0113] This application's parallel channel and spatial attention calibration mechanism adaptively adjusts features from two orthogonal dimensions: channel and space. This enables accurate capture of foreground target regions amidst massive background noise. The efficient channel attention branch captures cross-channel dependencies with minimal parameter overhead, enhancing the expressive power of key feature channels. The lightweight spatial attention branch generates a spatial weight map based on depthwise separable convolutions, guiding the model to focus on key spatial locations containing the target. By adaptively balancing the outputs of the two attention paths using learnable scalar weights, the limitations of a single attention mechanism are avoided, further improving feature discriminative power, especially enhancing the feature capture capability for challenging samples such as distant small targets and heavily occluded targets.

[0114] In some embodiments, cross-scale attention fusion and hierarchical propagation processing are performed on the calibrated multi-scale features to obtain an enhanced multi-scale feature map, including: performing bilinear interpolation resolution alignment processing on the calibrated multi-scale features to obtain multi-scale features with uniform resolution.

[0115] Specifically, calibrated multi-scale features at three different resolutions—fine-grained, medium-grained, and coarse-grained—are acquired. Using the resolution of the medium-grained features as a benchmark, downsampling is performed on the fine-grained features, and upsampling is performed on the coarse-grained features. All sampling processes employ bilinear interpolation. Through these operations, the feature maps at the three different scales are uniformly adjusted to the same spatial resolution, resulting in multi-scale features with a unified resolution, laying the foundation for subsequent feature interactions between scales.

[0116] Furthermore, global average pooling and concatenation are performed on the multi-scale features with uniform resolution to obtain the scale description matrix.

[0117] Specifically, a global average pooling operation is performed on each feature map after uniform resolution, and the feature values ​​of each channel are averaged in the spatial dimension to compress each two-dimensional feature map into a one-dimensional feature description vector. The three feature description vectors corresponding to fine-grained, medium-grained and coarse-grained features are concatenated along the dimensional direction to form a scale description matrix, which can comprehensively represent the global information of features at different scales.

[0118] Furthermore, the scale description matrix is ​​processed by self-attention calculation to obtain the initial fusion weights of features at different scales.

[0119] Specifically, query vectors, key vectors, and value vectors are generated based on the scale description matrix. The degree of association between features at different scales is obtained by calculating the similarity between the query vector and the key vector. The similarity results are normalized to obtain the initial fusion weight corresponding to each scale feature. This weight can reflect the relative importance of features at different scales in the current detection task.

[0120] Furthermore, the initial fusion weights are subjected to hierarchical attention propagation processing to obtain the corrected fusion weights.

[0121] Specifically, a hierarchical attention propagation mechanism is introduced to obtain the attention information calculated by the previous layer; the attention information of the previous layer is multiplied by a preset decay coefficient and then added to the initial fusion weight of the current layer; the influence of the attention information of the previous layer on the current layer is adjusted by the decay coefficient to ensure the consistency and coherence of cross-level feature fusion, and finally the corrected fusion weight is obtained.

[0122] Furthermore, the multi-scale features with uniform resolution are weighted and fused with the corrected fusion weights, and residual connection processing is performed to obtain an enhanced multi-scale feature map.

[0123] Specifically, based on the corrected fusion weights, the three scale features of the same resolution are summed element-wise to obtain the weighted fused features. At the same time, residual branches are constructed, and the original medium-granularity calibration features are adjusted for the number of channels through a convolutional layer to become residual features. The weighted fused features and the residual features are added element-wise, and the gradient vanishing problem is avoided through residual connections, finally obtaining a highly robust enhanced multi-scale feature map.

[0124] This application's embodiments model the dependencies between features at different scales through a cross-scale self-attention mechanism, enabling dynamic allocation of fusion weights and achieving "location-specific" feature integration. This approach is more adaptable to the drastic changes in target scale in autonomous driving scenarios compared to fixed-weight fusion methods. The introduction of a hierarchical attention propagation mechanism ensures the consistency and coherence of cross-level feature fusion, avoiding conflicts between features at different levels. Combined with a residual connection structure, it effectively alleviates the gradient vanishing problem during deep network training, improving the model's training stability and convergence accuracy. The resulting enhanced multi-scale feature map possesses strong robustness and feature representation capabilities, significantly improving the model's detection performance in complex scenes, sparse point cloud regions, and challenging samples.

[0125] While this application provides the method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in this embodiment is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the methods shown in this embodiment or the accompanying drawings can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0126] like Figure 3 As shown in the illustration, this application also provides a three-dimensional target detection device 300. The device includes: The acquisition module 301 is used to acquire three-dimensional point cloud data collected by the vehicle-mounted LiDAR, preprocess the three-dimensional point cloud data and perform multi-resolution spatial mesh division to obtain multi-scale division results.

[0127] The processing module 302 is used to perform lightweight feature encoding processing on the multi-scale segmentation results to obtain a multi-scale pseudo-image tensor.

[0128] The processing module 302 is also used to perform backbone feature extraction processing on the multi-scale pseudo-image tensor to obtain multi-level feature maps.

[0129] The processing module 302 is also used to construct and fuse bidirectional cross-scale feature pyramids on multi-level feature maps to obtain initial fused features.

[0130] The processing module 302 is also used to perform parallel channel and spatial attention calibration on the initial fused features to obtain calibrated multi-scale features.

[0131] The processing module 302 is also used to perform cross-scale attention fusion and hierarchical propagation processing on the calibrated multi-scale features to obtain an enhanced multi-scale feature map.

[0132] The inference module 303 is used to perform lightweight detection head inference processing on the enhanced multi-scale feature map to obtain the category, spatial location, size and orientation information of the three-dimensional target.

[0133] Some modules in the apparatus described in this application can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0134] The apparatus or module described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. For ease of description, the above apparatus is described by dividing it into various modules according to their functions. When implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware. Of course, a module that implements a certain function can also be implemented by combining multiple sub-modules or sub-units.

[0135] The methods, apparatus, or modules described in this application can be implemented in a computer-readable program code manner. The controller can be implemented in any suitable manner, such as a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of a memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code manner, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included within it for implementing various functions can also be considered as structures within the hardware component. Alternatively, the device used to implement various functions can be viewed as either a software module that implements the method or a structure within a hardware component.

[0136] This application also provides an apparatus, the apparatus comprising: a processor; a memory for storing processor-executable instructions; wherein, when the processor executes the executable instructions, it implements the method described in this application.

[0137] This application also provides a non-volatile computer-readable storage medium storing a computer program or instructions thereon, which, when executed, enables the method described in this application embodiment to be implemented.

[0138] Furthermore, in the various embodiments of the present invention, each functional module can be integrated into a processing module, or each module can exist independently, or two or more modules can be integrated into a single module.

[0139] The aforementioned storage media include, but are not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions.

[0140] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, or it can be embodied in the process of data migration. The computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0141] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this application can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0142] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of this application.

Claims

1. A three-dimensional target detection method, characterized in that, include: Acquire 3D point cloud data collected by vehicle-mounted LiDAR, preprocess the 3D point cloud data and perform multi-resolution spatial mesh division to obtain multi-scale division results; The multi-scale segmentation results are subjected to lightweight feature encoding processing to obtain a multi-scale pseudo-image tensor; The multi-scale pseudo-image tensor is subjected to backbone feature extraction processing to obtain a multi-level feature map; The multi-level feature maps are subjected to bidirectional cross-scale feature pyramid construction and fusion processing to obtain initial fused features; The initial fused features are subjected to parallel channel and spatial attention calibration to obtain calibrated multi-scale features; The calibrated multi-scale features are subjected to cross-scale attention fusion and hierarchical propagation processing to obtain an enhanced multi-scale feature map; The enhanced multi-scale feature map is subjected to lightweight detection head inference processing to obtain the category, spatial location, size, and orientation information of the three-dimensional target.

2. The method according to claim 1, characterized in that, The step of performing lightweight feature encoding processing on the multi-scale segmentation results to obtain a multi-scale pseudo-image tensor includes: The geometric center offset and feature dimensionality reduction are performed on the multi-scale partitioning results to obtain a seven-dimensional point feature vector; The seven-dimensional point feature vector is subjected to high-dimensional feature mapping to obtain the mapped point features; The mapped point features are subjected to max pooling aggregation to obtain a feature vector of fixed length; The fixed-length feature vectors are subjected to two-dimensional grid spatial scattering processing to obtain multi-scale pseudo-image tensors corresponding to different resolutions.

3. The method according to claim 1, characterized in that, The process of extracting backbone features from the multi-scale pseudo-image tensor to obtain multi-level feature maps includes: The multi-scale pseudo-image tensor is subjected to depthwise separable convolution feature extraction to obtain initial backbone features; The initial backbone features are downsampled to obtain the downsampled backbone features; The downsampled backbone features are subjected to global average pooling to obtain channel descriptors; The channel descriptors are processed by channel relationship learning to obtain channel attention weight vectors; The downsampled backbone features and the channel attention weight vector are subjected to feature recalibration processing to obtain multi-level feature maps of three different levels.

4. The method according to claim 1, characterized in that, The process of constructing and fusing a bidirectional cross-scale feature pyramid on the multi-level feature maps to obtain initial fused features includes: Semantic information transfer processing is performed on the multi-level feature maps to obtain semantically enhanced shallow features; Geometric detail transfer processing is performed on the multi-level feature maps to obtain deep features with enhanced details; The semantically enhanced shallow features and the detail-enhanced deep features are subjected to fast normalization and weighted fusion processing to obtain the initial fused features.

5. The method according to claim 1, characterized in that, The initial fused features are subjected to parallel channel and spatial attention calibration to obtain calibrated multi-scale features, including: The initial fusion features are subjected to efficient channel attention processing to obtain a channel weight map; The initial fusion features are subjected to lightweight spatial attention processing to obtain a spatial weight map; Adaptive weighted fusion processing is performed on the initial fusion features, the channel weight map, and the spatial weight map to obtain calibrated multi-scale features.

6. The method according to claim 1, characterized in that, The process of performing cross-scale attention fusion and hierarchical propagation on the calibrated multi-scale features to obtain an enhanced multi-scale feature map includes: The calibrated multi-scale features are subjected to bilinear interpolation resolution alignment processing to obtain multi-scale features with uniform resolution. Global average pooling and concatenation are performed on the multi-scale features of the unified resolution to obtain the scale description matrix; The scale description matrix is ​​subjected to self-attention calculation to obtain the initial fusion weights of features at different scales; The initial fusion weights are subjected to hierarchical attention propagation processing to obtain the corrected fusion weights; The multi-scale features with the unified resolution and the modified fusion weights are weighted and fused, and residual connection processing is performed to obtain an enhanced multi-scale feature map.

7. A three-dimensional target detection device, characterized in that, include: The acquisition module is used to acquire three-dimensional point cloud data collected by the vehicle-mounted LiDAR, and to preprocess the three-dimensional point cloud data and perform multi-resolution spatial mesh division processing to obtain multi-scale division results. The processing module is used to perform lightweight feature encoding processing on the multi-scale segmentation results to obtain a multi-scale pseudo-image tensor. The processing module is also used to perform backbone feature extraction processing on the multi-scale pseudo-image tensor to obtain a multi-level feature map. The processing module is also used to perform bidirectional cross-scale feature pyramid construction and fusion processing on the multi-level feature maps to obtain initial fused features; The processing module is also used to perform parallel channel and spatial attention calibration on the initial fused features to obtain calibrated multi-scale features; The processing module is also used to perform cross-scale attention fusion and hierarchical propagation processing on the calibrated multi-scale features to obtain an enhanced multi-scale feature map. The inference module is used to perform lightweight detection head inference processing on the enhanced multi-scale feature map to obtain the category, spatial location, size, and orientation information of the three-dimensional target.

8. A computer device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images