Point cloud target detection model training method and device based on asymmetric feature consistency

Abstract

This invention discloses a training method and apparatus for a point cloud object detection model based on asymmetric feature consistency, relating to the fields of autonomous driving perception and computer vision technology. The method includes: acquiring source domain point cloud data and inputting it into a model to obtain object detection loss, and extracting geometric and intensity features respectively; perturbing the geometric features to generate enhanced geometric features, and mapping the geometric, enhanced geometric, and intensity features to high-dimensional representations of the original geometric, enhanced geometric, and intensity features respectively; determining the geometric self-consistency loss based on the original and enhanced geometric feature representations; performing a gradient stopping operation on the original geometric feature representation, and combining it with the intensity feature representation to determine the asymmetric feature alignment loss; finally, weighted and integrated the above losses into a total loss to update the model parameters. This invention, by establishing unidirectional gradient blocking, utilizes intensity semantics to assist geometric feature learning while completely blocking noise pollution caused by cross-domain intensity bias, significantly improving the model's single-source domain generalization ability.

Description

Technical Field

[0001] This invention relates to the fields of autonomous driving perception and computer vision technology, and in particular to a method and apparatus for training a point cloud target detection model based on asymmetric feature consistency. Background Technology

[0002] Currently, 3D target detection from LiDAR point clouds is a key technology in the large-scale deployment of autonomous driving systems in the real world. Existing deep learning-based 3D target detection methods typically assume that training and test data are independently and identically distributed. However, in practical applications, due to differences in LiDAR hardware and environmental variations, there are significant domain shifts in the point cloud distribution between different datasets, leading to severe performance degradation of the models. LiDAR point cloud data contains geometric coordinates and reflection intensity information, with reflection intensity information reflecting the material properties of the target and possessing important semantic value. However, in cross-domain scenarios, the intensity distribution between different datasets often exhibits drastic and irregular shifts. To avoid the risk of intensity domain shift, existing technologies typically choose to directly remove reflection intensity information and rely solely on geometric coordinates for detection. This results in the loss of rich semantic information about the object's surface material contained in the intensity information, limiting the model's detection capabilities, especially significantly impacting the detection performance of small targets such as pedestrians and riders. Therefore, how to effectively utilize the semantic value of intensity information during the training phase while preventing intensity domain bias from negatively contaminating the geometric backbone network has become an urgent technical problem to be solved. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a point cloud target detection model training method and device based on asymmetric feature consistency.

[0004] To solve the above technical problems, the technical solution of the present invention is as follows: A training method for a point cloud object detection model based on asymmetric feature consistency includes: Obtain point cloud data from the source domain, input it into the point cloud target detection model to obtain target detection loss, and extract geometric features and intensity features in the same spatial dimension from the point cloud data respectively; The geometric features are perturbed at the feature level to generate enhanced geometric features, and the geometric features, the enhanced geometric features, and the intensity features are mapped to the original geometric feature representation, the enhanced geometric feature representation, and the intensity feature representation, respectively. Based on the original geometric feature representation and the enhanced geometric feature representation, the geometric self-consistency loss is determined; A gradient-stopping operation is performed on the original geometric feature representation, and an asymmetric feature alignment loss is determined based on the original geometric feature representation after the gradient-stopping operation and the intensity feature representation. The target detection loss, the geometric self-consistency loss, and the asymmetric feature alignment loss are weighted and integrated into a total loss, and the model parameters of the point cloud target detection model are updated based on the total loss.

[0005] The present invention also provides a point cloud target detection model training device based on asymmetric feature consistency, comprising: The basic feature extraction module is used to acquire point cloud data from the source domain, input it into the point cloud target detection model to obtain target detection loss, and extract geometric features and intensity features in the same spatial dimension from the point cloud data respectively. The feature mapping module is used to perform feature-level perturbation on the geometric features to generate enhanced geometric features, and to map the geometric features, the enhanced geometric features and the intensity features to the original geometric feature representation, the enhanced geometric feature representation and the intensity feature representation, respectively. The self-consistency loss determination module determines the geometric self-consistency loss based on the original geometric feature representation and the enhanced geometric feature representation; An asymmetric alignment loss determination module is used to perform a gradient stopping operation on the original geometric feature representation and determine the asymmetric feature alignment loss based on the original geometric feature representation after performing the gradient stopping operation and the intensity feature representation; The parameter update module is used to weight and integrate the target detection loss, the geometric self-consistency loss, and the asymmetric feature alignment loss into a total loss, and update the model parameters of the point cloud target detection model based on the total loss.

[0006] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described training method.

[0007] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described training method.

[0008] The beneficial effects of this invention are: (1) This invention, by constructing a geometry-intensity dual-flow BEV architecture, for the first time safely utilizes reflection intensity information discarded by mainstream methods in cross-domain detection. Through an asymmetric gradient stopping operation, the intensity flow is forced to unidirectionally align with the geometry flow, which not only allows the material semantics in the intensity information to assist the geometry backbone network in learning more robust representations, but also strictly prevents the intensity domain shift under different sensors and environments from contaminating the geometric features. Ablation experiments show that the symmetric variant without gradient stopping causes the 3D mAP to plummet from 44.14% to 40.61%, verifying that the core mechanism of this invention effectively avoids negative transfer.

[0009] (2) This invention designs a feature-level stochastic augmentation and a dual-projection head architecture within the geometric flow, and forces the backbone network to learn features invariant to various perturbations through self-consistency loss. Ablation experiments show that the performance is optimal when both augmentation operations are used simultaneously (3DmAP 44.14%), while the performance drops to 41.67% after removing all augmentations. This mechanism significantly improves the model's resistance to domain offset factors such as changes in lidar beam density and long-range sparse scanning.

[0010] (3) This invention utilizes an occupancy mask to calculate the consistency loss only in the effective lidar echo region, avoiding the calculation of meaningless loss on blank grids and reducing noise interference. This not only speeds up the training convergence speed but also prevents the adverse effects of feature drift in open areas on the backbone network.

[0011] (4) In this invention, the geometric projection head, the enhanced geometric projection head, and all data augmentation operations are only used during the training phase and are completely removed during inference. Only the lightweight intensity projection branch is retained, and a simple convolutional layer is used to integrate the features into the backbone. Therefore, the final deployed model has a very similar number of parameters and inference speed to the baseline detector, meeting the stringent real-time requirements of autonomous driving. Experiments show that embedding Asym-GRCL into detectors with three different architectures—Part-A², PointPillars, and CenterPoint—all bring stable performance improvements, verifying its versatility. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating the point cloud target detection model training method based on asymmetric feature consistency provided by this invention; Figure 2 The overall architecture flowchart of the point cloud target detection model training method based on asymmetric feature consistency provided for the embodiment; Figure 3 This is a schematic diagram illustrating the process of using an occupancy mask to filter the grid in the feature tensor and calculate the loss in the embodiment. Figure 4 This is a visualization of the asymmetric gradient flow within the Asym-GRCL framework in the embodiment; Figure 5A schematic diagram of the point cloud target detection model training device based on asymmetric feature consistency provided by the present invention; Figure 6 A schematic diagram of the computer device provided by the present invention. Detailed Implementation

[0014] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0015] Example 1: As Figure 1 As shown, this embodiment provides a point cloud object detection model training method based on asymmetric feature consistency. This method aims to address the difficulty in utilizing intensity information in cross-domain scenarios and the potential for negative transfer. By constructing a dual-stream architecture of geometric and intensity feature flows and introducing an asymmetric gradient stopping mechanism, it effectively prevents intensity domain bias from back-contaminating geometric features while fully utilizing intensity information to supplement semantics, thereby improving the cross-domain generalization performance of the point cloud object detection model. The overall framework diagram of the above method is shown below. Figure 2 As shown. Although this invention is compatible with various 3D point cloud detection networks, for ease of comparison with other similar methods, Voxel R-CNN will be used as the baseline model for this invention. The specific solution includes the following steps: Step S101: Obtain point cloud data from the source domain, input it into the point cloud target detection model to obtain the target detection loss, and extract geometric features and intensity features in the same spatial dimension from the point cloud data.

[0016] Specifically, given in the source domain D S A frame of raw LiDAR point cloud data collected in [the context of the data]. This contains N discrete three-dimensional laser reflection points. Each point... Each contains at least two dimensions of measurement information: one is the geometric coordinates describing its three-dimensional spatial position in the coordinate system of the data acquisition vehicle. Secondly, scalar reflection intensity information describing the physical reflective material characteristics. .

[0017] First, the collected raw point cloud data is input into the point cloud object detection model for forward propagation. The downstream detection head then performs bounding box regression and category classification tasks to obtain the baseline object detection loss. .

[0018] While acquiring the detection loss, geometric features and intensity features in the same spatial dimension are extracted separately, as follows: Geometric Feature Extraction: In a standard 3D detection preprocessing workflow, the detector first regularizes the disordered and unevenly distributed raw point cloud by voxelization or cylinderization, mapping it into a regular 3D or 2D mesh. Subsequently, local features are extracted using 3D sparse convolution or a simple multilayer perceptron (MLP) network, and finally projected along the vertical Z-axis, compressing it into a two-dimensional BEV plane. In the standard workflow, the sparse and irregular input point cloud is first finely voxelized into a 3D mesh distribution, followed by hierarchical feature processing and downsampling through a multilayer 3D sparse backbone network to efficiently extract voxel-by-voxel spatial occupancy features. After semantic encoding in 3D space, the network compresses the feature depth along the vertical dimension of the model along the Z-axis and seamlessly integrates with several standard 2D convolutional modules to obtain high-density 2D BEV features. The geometric BEV features obtained after extraction and projection by the backbone network are denoted as... Where B is the batch size, C g H×W represents the channel depth of the backbone network output features, while H×W represents the length and width spatial resolution dimensions of the BEV mesh system in physical space. Here, the geometric BEV feature F... g It has already highly condensed the high-level semantic and complex geometric topological information of the object from this perspective, and it will be directly sent to the downstream detection head to complete the bounding box regression and category classification tasks.

[0019] One detail to note is that, since it originates from the original 3D backbone detection network containing all channel information, although the geometric BEV feature F... g Named and defined as a "geometric flow" used to extract topological structures, its actual feature space inevitably includes some intensity features. Such a deep parameter mapping is prone to overfitting when trained in a single source domain, which becomes a major factor leading to the collapse of generalization performance during domain transfer. Asym-GRCL addresses this by introducing a clear, independent intensity-supervised semantic flow, thereby decoupling and eliminating redundant intensity domain offset information within the dataset.

[0020] Intensity Feature Extraction: The geometric features of a point cloud are represented by the x, y, z coordinates of points, while the intensity features of the point cloud are represented solely by r. To construct an intensity flow that aligns with deep geometric features, it is necessary to directly extract and construct single-channel intensity BEV features containing spatial context from the lossless raw laser point cloud coordinate point set. Unlike the geometric flow constructed using voxel feature extraction networks, the intensity branch design is very simple and lightweight, using only the raw intensity semantic information. Specifically, according to the same physical resolution spatial size as the voxelization of the backbone network, the continuous BEV coordinate system is divided into discrete grid cells. Let a specific grid cell on the discretized two-dimensional BEV plane be the global coordinate (u, v), and let... This represents the original set of lidar points falling within this voxel. The intensity value at position (u, v) is obtained by calculating the average pooling of the points: ,in, This represents the reflection intensity value of point k. Average pooling is an efficient method for processing point clouds. It can smooth the data and filter out extreme intensity values ​​caused by scattering or lidar thermal noise to a great extent, while significantly reducing computational complexity. In this way, a single-channel intensity BEV feature is explicitly constructed as the intensity BEV feature, denoted as […]. This prepared the second crucial input tensor for the subsequent Asym-GRCL framework.

[0021] Step S102: Perform feature-level perturbation on the geometric features to generate enhanced geometric features, and map the geometric features, enhanced geometric features, and intensity features to the original geometric feature representation, enhanced geometric feature representation, and intensity feature representation, respectively.

[0022] Specifically, the goal of the geometric branch is to force the backbone network to learn invariant features of point cloud targets under extreme external perturbations and cross-domain offsets, thereby mitigating the negative impact of point cloud deformation or sensor noise from different target domains on deep geometric semantics from the source. To avoid the model collapsing into a simple identity mapping during consistency learning, the geometric branch does not directly constrain the output of the original backbone network, but instead projects it onto a more discriminative high-dimensional embedding space. This design includes two parallel branches: the original path and the enhancement path, designed to force the two projectors to output highly consistent representations even when faced with input perturbations, thereby improving the finally learned geometric features F. g Robustness.

[0023] To simulate the complex distribution shifts that may occur in the target domain as realistically as possible during the source domain training phase, the enhancement path of the geometric flow introduces a feature-level stochastic enhancement operator. Geometric BEV features of the backbone network output Perform explicit perturbation. The resulting enhanced geometry formula is: .

[0024] Specifically, enhancement operators It includes two operations: spatial feature dropout and feature Gaussian noise. For spatial feature Dropout, and for geometric BEV feature F g For each spatial grid location (u,v), a binary spatial mask following a Bernoulli distribution is generated, with probability... Set the channel feature vector at a specific location to zero, and then use... The system rescales to maintain the expected feature activations during training and testing. This operation effectively simulates point cloud sparsity caused by different LiDAR beams or long-range scanning at the feature level, as well as physical occlusion common in complex traffic scenes.

[0025] For the characteristic Gaussian noise, inject a random noise tensor that follows a zero-mean Gaussian distribution into the geometric BEV feature. ,in, This operation aims to simulate minute drifts in point cloud coordinates and high-frequency local noise interference caused by sensor hardware aging, reflectivity decay, or extreme weather.

[0026] Unlike conventional preprocessing enhancements applied to the raw point cloud data input, the enhancement operations described above are directly applied to the high-dimensional feature maps encoded by the backbone network. This design is based on two considerations: firstly, feature-level enhancement avoids the enormous computational overhead of directly transforming the original large-scale point cloud; secondly, injecting perturbations directly into the high-dimensional features can, to some extent, limit the model's overfitting to local textures in a specific domain, while not disrupting the target-level three-dimensional relative geometric topology, thus enabling the backbone network to extract robust geometric features with cross-domain universality.

[0027] To handle pure features and perturbated features separately, a dual-projection head architecture was customized for the original path and the enhanced path. (Geometric projection head) F g Projecting into the embedded space, enhancing the geometric projection head Structure and The same but with independent parameters, applied to the enhanced geometric features. In the current implementation, and Both consist of two layers The multilayer perceptron (MLP) structure, composed of convolutional cascades, can efficiently achieve cross-channel information fusion and complex nonlinear mapping.

[0028] The original features and enhanced features are mapped and extracted by the two projection heads into an E-dimensional feature embedding space with the same dimension and shared parameters, which is formalized as follows: The enhanced geometric feature representations are generated respectively. and original geometric feature representation ,in, The two projectors do not share parameters, which is intended to prevent the network from learning simple identity mappings. The network is forced to go beyond shallow numerical differences to dig deeper and align the stable and invariant deep geometric semantic features between the original path and the perturbed path.

[0029] For intensity flow mapping, the aim is to extract semantic context from reflection intensity data as a complementary modality to geometric information. This branch is designed to extract stable surface texture and physical material context from the raw but physically meaningful reflection intensity data, and transform it into a parallel complementary modality for geometric structure information. The network structure of intensity flow is more complex than that of geometric flow. This is due to the input intensity BEV features. The last dimension, derived directly from the original point cloud, is single-channel and sparse, and highly sensitive to sensor characteristics and calibration differences. This makes the intensity BEV feature extremely susceptible to interference from domain offset characteristics, limiting the effectiveness of channel blending when using a 1×1 convolution similar to the geometric flow branch. Therefore, the intensity flow branch has been designed differently from the geometric flow branch. It uses an intensity projection mapping head with a larger receptive field. Dominantly, the projection head internally stacks multiple layers of standard 3×3 convolutions, along with corresponding nonlinear activations and normalization to effectively expand the receptive field and aggregate semantics of surrounding relevant regions. Intensity features are input into the intensity projection head to generate intensity feature representations. This more complex network design not only fully expands single-channel information to an E-dimensional feature depth that matches the geometric flow, but also utilizes the extended neighborhood receptive field to compensate for the gaps caused by missing local point clouds. As a result, the network can more effectively mine specific texture patterns that are independent of sensor system noise and target distribution differences in sparse reflections.

[0030] The three features mentioned above have the following dimensions: , , .

[0031] Step S103: Determine the geometric self-consistency loss based on the original geometric feature representation and the enhanced geometric feature representation.

[0032] Specifically, before feeding the data into the consistency loss calculation, this embodiment performs a pre-filtering step to completely eliminate interference from the inherent empty voxel grids in the lidar point cloud. Addressing the sparsity of lidar point clouds, an occupancy mask design is introduced to avoid calculating the consistency loss in empty voxel regions. An occupancy mask is generated based on intensity features to indicate the effective grid positions where the reflection intensity values ​​meet set conditions; the mask is defined. The expression for each grid is: ,in, It is an indicator function. This is the set threshold. This mask is used to explicitly indicate the grid locations with valid reflection intensity values ​​in the intensity BEV feature, ensuring that subsequent consistency adversarial loss is calculated only on specific regions of interest (ROIs) with valid LiDAR return signals, eliminating meaningless alignment due to invalid background holes.

[0033] For the specific physical process of executing spatial filtering, please refer to [link / reference]. Figure 3 The system will use high-dimensional feature tensors , , Tensor flattening is performed on the spatial grid resolution dimension, transforming it into a (BHW,E) dimensional matrix. Simultaneously, the occupancy mask M is flattened to the form (BHW,1). Subsequently, element-wise filtering is performed using the flattened mask as grid position indices, removing feature vectors corresponding to empty grids that do not meet the set threshold condition (i.e., M=0). Only the set of feature vectors corresponding to valid grid coordinate positions (M=1) is extracted and retained.

[0034] In the boundary condition control verification of this embodiment, the size of the effective feature set retained after filtering is statistically analyzed in real time. If the current number of effective samples is lower than the set minimum threshold, the device directly interrupts the alignment calculation, returns a loss value of 0, and skips the current batch. If the effective sample size meets the threshold, the extracted features are normalized along the sample dimension, thereby outputting the normalized effective original geometric feature representation, effective enhanced geometric feature representation, and effective intensity feature representation. Through this purely objective tensor normalization operation, it is ensured that the subsequent consistency loss function focuses on the effective region.

[0035] To meet the input requirements of the final alignment formula, after obtaining the effective original geometric feature representation, effective enhanced geometric feature representation, and effective intensity feature representation, the system normalizes them along the sample dimension to project them onto a unified scale space. After the above spatial region filtering and normalization processing, the final output feature tensors are respectively aligned to the effective original geometric feature representations. Effectively enhances geometric feature representation and effective intensity characteristics representation .

[0036] Before determining the specific loss, based on the aforementioned effective feature combinations in the normalized plane, this embodiment employs the Barlow Twins method to systematically quantify the degree of interleaving of features in deep semantics. It introduces any combination of two masked feature representations that are in the normalized plane. and This theory calculates their cross-correlation matrix in the feature dimension. This system quantifies the degree of cross-correlation between these two branch features in deep semantics, where N measures the effective sample size extracted by the current batch mask. The alignment optimization objective of this application is to force the calculated cross-correlation matrix to converge to the ideal identity matrix through continuous gradient descent. Its cross-correlation consistency loss function... The calculation formula is defined as follows: .

[0037] The above formula contains two optimization constraints with explicit physical and algorithmic evolutionary significance: First item As an invariant term, it forces the diagonal elements of the cross-correlation matrix to tend to 1, making the characterization invariant to different perturbations; Second item To remove redundant terms, off-diagonal elements are forced to approach 0, and mutual exclusion and decoupling are enforced between feature channels to remove redundancy between dimensions of representation. It is a hyperparameter that weighs the importance of these two objectives.

[0038] Finally, the normalized effective original geometric feature representation (as a general term) and effectively enhanced geometric feature representation (as a general term) Substituting this into the previously constructed Barlow Twins consistency loss calculation architecture, the system first calculates the cross-correlation matrix between the two independent paths within the geometric flow. Furthermore, the geometric self-consistency loss is determined within the fully differentiable computational graph. The specific formula is expressed as follows: .

[0039] By minimizing this geometric self-consistency regularization loss during the training optimization process This approach forces the model's original 3D backbone network to overcome shallow numerical perturbations, deeply mining and extracting high-order topological features that are highly geometrically invariant to spatial deformations and physical occlusions caused by sensor parameters, weather changes, and other factors. In this step, gradient flow on the computation graph is allowed to propagate normally in both directions, and error gradients are smoothly backpropagated to the 3D sparse convolutional backbone network and all preceding geometric parameter layers, achieving explicit cross-domain generalization optimization of the geometric feature extraction backbone.

[0040] Step S104: Perform a stopping gradient operation on the original geometric feature representation, and determine the asymmetric feature alignment loss based on the original geometric feature representation and intensity feature representation after performing the stopping gradient operation.

[0041] Specifically, since the spatial distribution of reflection intensity in different physical datasets and under different weather conditions has a highly and irregular domain-sensitive shift, once the gradient of intensity loss is backpropagated and the geometric backbone parameters are forcibly changed, the originally stable geometric spatial topology learning will be severely polluted by noise, resulting in obvious "negative transfer" overfitting.

[0042] To fully utilize the surface material and texture semantics inherent in the intensity features, while simultaneously isolating them from the geometric skeleton at the physical level, this method introduces an asymmetric cross-modal alignment mechanism employing a stopping gradient operation. See [link to relevant documentation] Figure 4 When calculating the cross-modal alignment loss, the system explicitly embeds the stop-gradient operator on the geometric flow side of the computation graph. This forcibly truncates the gradient flow of the node during backpropagation. Subsequently, the system bases its analysis on the original geometric feature representation and the normalized effective intensity feature representation after performing the gradient-stopping operation. The asymmetric feature alignment loss was determined using the Barlow Twins mechanism. The specific equation is defined as follows: ,in, This is the cross-correlation consistency loss function.

[0043] According to this asymmetric loss formula, during error backpropagation, the cross-modal alignment error gradient flow generated on the intensity side flows backward along the computation graph to the original geometric feature representation. When the position is determined, the gradient operator will be stopped. The constructed unidirectional gradient blocking mechanism effectively truncates the gradient, preventing it from propagating further back into the 3D sparse convolutional backbone network. This design ensures that the intermodal interaction evolution is unidirectional and asymmetric: the reflection intensity flow can unidirectionally guide and inject its inherent material and morphological semantics into the geometric feature representation space, while the domain-specific intensity distribution deviation noise it carries is completely isolated, preventing any negative pollution to the backbone perception network and fundamentally eliminating cross-domain negative transfer at the underlying mathematical mechanism level.

[0044] Step S105: The target detection loss, geometric self-consistency loss, and asymmetric feature alignment loss are weighted and integrated into a total loss, and the model parameters of the point cloud target detection model are updated based on the total loss.

[0045] Specifically, the proposed Asym-GRCL architecture is a highly decoupled, plug-and-play multi-task auxiliary training module that can be integrated into common 3D detection pipelines. During the multi-task joint training and optimization phase of the model, the system uses the first task loss (object detection loss) obtained from forward propagation. ), and the geometric self-consistency loss generated in parallel by the auxiliary network. Alignment loss with asymmetric features The data is weighted and integrated according to specific weighting coefficients to calculate and construct the final total loss that guides the evolution of the entire network. The weighted summation formula is detailed below: .

[0046] In the above formula, This is a balance factor used to balance the relative strength between the internal self-constraints of the geometric flow and the cross-modal feature alignment constraints (in the general scenario of this embodiment, The value can be flexibly set to 1.5. This is the weighting factor that controls the gradient weights of the entire auxiliary consistency learning system in the overall joint training multi-task (in this embodiment, (Flexible setting: 0.2).

[0047] The entire multi-task network is based on this total loss function. Perform standard error backpropagation. Dynamically and synchronously update the parameters of all network layers in the point cloud object detection model that are not gradient-blocked (including the 3D sparse convolution backbone, geometric projection head) using stochastic gradient descent (SGD) or the AdamW optimizer. Enhanced geometric projection head and intensity projection head Differentiable weights and biases are used until the total loss function converges to a preset threshold range.

[0048] It is important to note that in the Asym-GRCL module constructed in this invention, the geometric projection head, enhanced geometric projection head, and all data augmentation operations are only enabled during the fully supervised training (model network optimization) phase and are completely removed from the network computation graph during inference. In actual online deployment for inference or testing, this invention retains only the lightweight intensity projection branch and uses a simple convolutional layer to integrate the extracted intensity features into the backbone network. By eliminating complex auxiliary alignment calculations and redundant high-dimensional mapping structures, the final deployed model has a very similar number of parameters and inference speed to the baseline detector. Therefore, the method of this invention can perfectly balance the high real-time requirements of industrial deployments while retaining the generalized 3D topology learned by the backbone network during training and effectively utilizing intensity information to supplement cross-domain semantics during the inference phase, thus achieving extremely high cross-domain generalization accuracy.

[0049] Example 2: This example provides a training device for a point cloud object detection model based on asymmetric feature consistency. See also... Figure 5The device includes: a basic feature extraction module 201, a feature mapping module 202, a self-consistency loss determination module 203, an asymmetric alignment loss determination module 204, and a parameter update module 205.

[0050] Specifically, the input of the basic feature extraction module 201 is communicatively connected to the point cloud data stream of sensors such as LiDAR, and its output establishes internal data stream connections with the main task loss calculation unit and the feature mapping module of the point cloud target detection model, respectively. The basic feature extraction module 201 executes the following specific control and calculation behaviors: Forward loss acquisition unit: retrieves raw LiDAR point cloud data from the training source domain. each of the points Each contains at least two dimensions of measurement information: one is the geometric coordinates describing its three-dimensional spatial position in the coordinate system of the data acquisition vehicle. Secondly, scalar reflection intensity information describing the physical reflective material characteristics. This unit controls the input of point cloud data into the point cloud target detection model to be trained, performs a complete forward propagation calculation, and outputs a 3D bounding box and class probability, based on which a benchmark target detection loss is generated. It is then either temporarily stored or directly sent to the parameter update module.

[0051] The geometric feature flow construction unit controls the meshing and voxelization or cylindrical segmentation of the disordered raw point cloud data. It then extracts voxel space occupancy features layer by layer through the model's built-in 3D sparse convolutional backbone network. Next, in the latent space, it controls the feature compression and planar convergence of the mesh features along the vertical Z-axis, ultimately producing a high-density 2D backbone geometric bird's-eye view (BEV) feature at the output, which serves as the geometric BEV feature. .

[0052] Intensity feature flow construction unit: Controlling the division of the continuous bird's-eye view coordinate system into discrete grid cells (u,v) according to the grid spatial resolution and physical boundaries that perfectly correspond to the geometric features. Statistically calculating the subset of point clouds falling within each specific grid area. The reflectivity is explicitly calculated using average pooling: After smoothing and filtering out high-frequency random noise from the sensor, a single-channel intensity BEV feature is independently constructed at the output end, which serves as the intensity BEV feature. .

[0053] The feature mapping module 202 refines the execution of complex control behaviors related to feature perturbation enhancement, high-dimensional latent space mapping representation, occupancy mask generation, and spatial region slicing filtering. The input of this module is connected to the basic feature extraction module 201, and the output is connected in parallel to the subsequent self-consistency loss determination module and asymmetric alignment loss determination module. The feature mapping module is further subdivided logically into: Feature-level random augmentation operator unit: directly acts on the geometric features F of the backbone network's encoded output. g This unit controls and executes at least one of the spatial feature dropout operation and the feature injection Gaussian noise operation to efficiently simulate target domain sparseness, long-distance scan attenuation, and large-area physical occlusion in the latent space, thereby producing enhanced geometric features at the output. .

[0054] Parameter-independent multi-projection head mapping unit: Contains three channel-level convolutional mapping matrices that are completely decoupled and independent in the feature embedding space. Geometric projection head. It contains two 1×1 convolutional cascaded multilayer perceptrons, responsible for processing pure geometric features F. g Mapped to the original geometric feature representation in a high-dimensional latent space Enhanced geometry projection head They have the same structure but completely independent parameters, and are responsible for enhancing geometric features. Mapping to enhanced geometric feature representation Intensity projection head It employs an asymmetric, complex structural design, internally stacking multiple layers of standard 3×3 convolutional blocks combined with nonlinear activation layers and normalization layers to expand the local receptive field and fill mesh gaps. This is responsible for mapping the intensity feature I to an intensity feature representation of the same dimension. .

[0055] Occupancy mask generation and spatial slicing filtering unit: This unit is based on single-channel intensity BEV characteristics. The real-time grid activation value is used to trigger the generation of a binary occupancy mask matrix via an indicator function. (threshold) =0), explicitly indicating the effective region of interest (ROI) where the entity target is located. Subsequently, this unit controls the high-dimensional feature tensor... , , Perform Flatten transformation in the planar spatial dimension, and use the flattened mask to perform element-level spatial cross-filtering to remove empty background meshes with a mask indicator of 0.

[0056] Boundary threshold verification subunit: Real-time monitoring and determination of the total capacity of valid grid samples retained in the current batch. If the number of valid samples is lower than the preset safety threshold, short-circuit control logic is triggered, the device directly returns an auxiliary loss value of 0 and forcibly skips the loss calculation for the current batch; if the sample capacity meets the threshold, the valid grid sample set is controlled to undergo standard normalization projection along the sample dimension, and finally three sets of scale-plane aligned hard core data streams are output at the module output end: valid original geometric feature representation. Effectively enhances geometric feature representation and effective intensity characteristics representation .

[0057] The input of the self-consistency loss determination module 203 is connected to the geometry flow output channel of the feature mapping module 202, and its output is connected to the parameter update module. The self-consistency loss determination module internally incorporates a Barlow Twins cross-correlation coefficient optimization matrix algorithm. This module receives the normalized valid original geometric feature representation... With effective enhancement of geometric feature representation Then, the module performs feature channel-level dot product cross-correlation operations on both components within the differentiable computation graph to calculate the corresponding cross-correlation matrix C. Subsequently, this module performs consistency objective loss optimization using the following formula: By minimizing the invariant terms and redundancy removal The joint constraints determine the geometric self-consistency loss used to explicitly constrain the topological invariance of the geometric space at the module output. The computation path within this module does not intercept gradient backpropagation in any way, allowing the backpropagation error gradient to pass smoothly.

[0058] The asymmetric alignment loss determines the 204 modules, which constitute the core "physical one-way firewall" of cross-modal and cross-domain generalization defense. Its inputs receive valid original geometric feature representations from the geometric flow branches. Effective intensity characteristics from intensity flow branches Its output is connected to the second auxiliary loss input of the parameter update module.

[0059] The asymmetric alignment loss determination module 204 contains a key deployment of stopping gradient operator logic: Gradient flow truncation operator unit: upon receiving a valid original geometric feature representation At that time, control forces a stopping gradient operation to be applied to the backpropagation computation layer surface. ; Asymmetric alignment computation unit: based on the original geometric feature representation after performing the stopped gradient operation and the normalized effective intensity feature representation. Using the Barlow Twins cross-correlation alignment architecture, the cross-modal asymmetric feature alignment loss is determined by the following formula. : .

[0060] Due to the explicit physical blocking effect of the gradient flow truncation operator unit, when the error backpropagation algorithm is started, the material alignment error gradient generated by the intensity flow branch will be forcibly truncated in place by the operator after flowing to the original geometric feature representation node on the geometric side. This completely isolates the backpropagation and penetration of the physical distribution offset unique to the intensity domain and sensor calibration noise into the 3D sparse convolution backbone network, ensuring that the modal alignment exhibits unidirectional asymmetric characteristics.

[0061] The input of the parameter update module 205 is connected to the physical bus of the main detection loss output of the basic feature extraction module 201, the output of the self-consistency loss determination module 203, and the output of the asymmetric alignment loss determination module 204, respectively. Its output is connected to the adjustable parameter matrix of all differentiable layers in the model.

[0062] The parameter update module executes the following specific control behaviors: Multi-task loss weighted integration unit: responsible for integrating the received target detection loss Geometric self-consistency loss and asymmetric feature alignment loss Based on the weight coefficients set by the hyperparameters, a multidimensional linear combination is performed, and the total loss guiding the optimization of the entire network parameters is calculated and determined according to the following formula. : ; Among them, the balance factor The dynamic ratio is 1.5, used to control the stiffness between the model's internal self-stability and cross-modal alignment; global weight factor. The dynamic ratio is 0.2, which is used to control the adjustment range of the auxiliary module on the gradient of the main task.

[0063] Backpropagation gradient optimization unit: based on the total loss The unit performs standard stochastic gradient descent or AdamW gradient evolution algorithms. By dynamically calculating gradient values, it synchronously updates all unstopped gradient operations in the point cloud object detection model along the backhaul path. The parameters of the entire differentiable network layers (including the weight parameter matrices of the 3D sparse convolutional backbone layer, geometric projector, and intensity projector in the initial extracted network) are blocked until the total loss function completely converges to the set optimal convergence interval, thereby completing the entity training of the high generalization model.

[0064] See Figure 6This embodiment also provides a computer device, the components of which may include, but are not limited to: one or more processors or processing units, system memory, and buses connecting different system components (including system memory and processing units).

[0065] A bus refers to one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. Examples of these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0066] Computer systems / servers typically include a variety of computer system-readable media. These media can be any available media that can be accessed by the computer system / server, including volatile and non-volatile media, and removable and non-removable media.

[0067] System memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The computer device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system may be used to read and write non-removable, non-volatile magnetic media. Disk drives for reading and writing to removable non-volatile disks (e.g., "floppy disks") and optical disc drives for reading and writing to removable non-volatile optical discs (e.g., CD-ROMs, DVD-ROMs, or other optical media) may be provided. In these cases, each drive may be connected to a bus via one or more data media interfaces. The memory may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0068] A program / utility having a set (at least one) of program modules can be stored, for example, in memory. Such program modules include—but are not limited to—an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The program modules typically perform the functions and / or methods described in the embodiments of this invention.

[0069] Computer devices can also communicate with one or more external devices (such as keyboards, pointing devices, monitors, etc.). This communication can be done through input / output (I / O) interfaces. Furthermore, computer devices can communicate with one or more networks (such as local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapters.

[0070] The processing unit executes the functions and / or methods described in the embodiments of the present invention by running programs stored in the system memory.

[0071] The aforementioned computer program can be stored in a computer storage medium, that is, the computer storage medium is encoded with a computer program, which, when executed by one or more computers, causes one or more computers to perform the method flow and / or device operation shown in the above embodiments of the present invention.

[0072] With the development of time and technology, the meaning of "medium" has become increasingly broad. The dissemination of computer programs is no longer limited to tangible media; they can also be downloaded directly from the network. Any combination of one or more computer-readable media can be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. Computer-readable storage media can be, for example,—but not limited to—electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0073] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0074] The program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0075] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0076] In addition to the above embodiments, the present invention may have other implementation methods; all technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A training method for a point cloud target detection model based on asymmetric feature consistency, characterized in that: include: Obtain point cloud data from the source domain, input it into the point cloud target detection model to obtain target detection loss, and extract geometric features and intensity features in the same spatial dimension from the point cloud data respectively; The geometric features are perturbed at the feature level to generate enhanced geometric features, and the geometric features, the enhanced geometric features, and the intensity features are mapped to the original geometric feature representation, the enhanced geometric feature representation, and the intensity feature representation, respectively. Based on the original geometric feature representation and the enhanced geometric feature representation, the geometric self-consistency loss is determined; A gradient-stopping operation is performed on the original geometric feature representation, and an asymmetric feature alignment loss is determined based on the original geometric feature representation after the gradient-stopping operation and the intensity feature representation. The target detection loss, the geometric self-consistency loss, and the asymmetric feature alignment loss are weighted and integrated into a total loss, and the model parameters of the point cloud target detection model are updated based on the total loss.

2. The point cloud target detection model training method based on asymmetric feature consistency according to claim 1, characterized in that: Extracting geometric features and intensity features in the same spatial dimension from the point cloud data includes: The point cloud data is voxelized or cylindricalized, and the grid features are extracted through a three-dimensional sparse backbone network. The grid features are then compressed into a two-dimensional bird's-eye view feature map along the vertical direction as the geometric features. According to the grid resolution corresponding to the geometric features, the continuous bird's-eye view coordinate system is divided into discrete grid cells, and the reflection intensity values ​​of the point cloud data falling within the same grid cell are subjected to average pooling processing to extract the single-channel bird's-eye view intensity features as the intensity features.

3. The point cloud target detection model training method based on asymmetric feature consistency according to claim 1, characterized in that: Generating enhanced geometric features by performing feature-level perturbation on the geometric features includes: performing at least one of spatial feature dropout operation and feature Gaussian noise injection operation on the geometric features; The spatial feature Dropout operation is as follows: all channel feature vectors at a specific spatial grid position in the geometric feature are set to zero with a preset probability, and then rescaled to keep the expected feature activation unchanged. The injection feature Gaussian noise operation is to inject a random noise tensor that follows a zero-mean Gaussian distribution into the geometric feature.

4. The point cloud target detection model training method based on asymmetric feature consistency according to claim 1, characterized in that: Mapping the geometric features, the enhanced geometric features, and the intensity features to their original geometric feature representations, enhanced geometric feature representations, and intensity feature representations, respectively, includes: The geometric features are input into a geometric projection head for mapping to generate the original geometric feature representation; The enhanced geometric features are input into the enhanced geometric projection head for mapping to generate the enhanced geometric feature representation; The intensity features are input into an intensity projection head for mapping to generate the intensity feature representation; The geometric projection head, the enhanced geometric projection head, and the intensity projection head have independent parameters, and the intensity projection head used to map the intensity features includes multiple stacked convolutional blocks.

5. The point cloud target detection model training method based on asymmetric feature consistency according to claim 1, characterized in that: The determination of the asymmetric feature alignment loss includes: The asymmetric feature alignment loss is calculated using the following formula: ,in, This is the normalized representation of the original geometric features. This is the normalized representation of the intensity feature. For cross-correlation consistency loss function, To stop the gradient operation.

6. The point cloud target detection model training method based on asymmetric feature consistency according to claim 1, characterized in that: The weighted integration of the target detection loss, the geometric self-consistency loss, and the asymmetric feature alignment loss into a total loss includes: The total loss is determined according to the following formula. : ,in, For target detection loss, For geometric self-consistency loss, For asymmetric feature alignment loss, As a balance factor, This is the weighting factor.

7. The point cloud target detection model training method based on asymmetric feature consistency according to claim 1, characterized in that: Before determining the geometric self-consistency loss and the asymmetric feature alignment loss, the method further includes: An occupancy mask is generated based on the intensity features to indicate the valid grid positions where the reflection intensity value meets the set conditions; The original geometric feature representation, the enhanced geometric feature representation, and the intensity feature representation are spatially filtered using the occupancy mask to extract the features corresponding to the effective grid positions, thereby obtaining effective original geometric feature representation, effective enhanced geometric feature representation, and effective intensity feature representation. Based on the effective original geometric feature representation, the effective enhanced geometric feature representation, and the effective intensity feature representation, the geometric self-consistency loss and the asymmetric feature alignment loss are calculated.

8. A training device for a point cloud target detection model based on asymmetric feature consistency, characterized in that, include: The basic feature extraction module is used to acquire point cloud data from the source domain, input it into the point cloud target detection model to obtain target detection loss, and extract geometric features and intensity features in the same spatial dimension from the point cloud data respectively. The feature mapping module is used to perform feature-level perturbation on the geometric features to generate enhanced geometric features, and to map the geometric features, the enhanced geometric features and the intensity features to the original geometric feature representation, the enhanced geometric feature representation and the intensity feature representation, respectively. The self-consistency loss determination module determines the geometric self-consistency loss based on the original geometric feature representation and the enhanced geometric feature representation; An asymmetric alignment loss determination module is used to perform a gradient stopping operation on the original geometric feature representation and determine the asymmetric feature alignment loss based on the original geometric feature representation after performing the gradient stopping operation and the intensity feature representation; The parameter update module is used to weight and integrate the target detection loss, the geometric self-consistency loss, and the asymmetric feature alignment loss into a total loss, and update the model parameters of the point cloud target detection model based on the total loss.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images