A cooperative perception method for intelligent connected vehicles based on 3D point clouds in intersection environments

By employing a 3D point cloud collaborative perception method in intelligent connected vehicles, and utilizing ConvLSTM networks and multi-scale feature fusion modules, the communication delay and spatiotemporal state asynchrony issues of intelligent connected vehicles in urban intersection environments are resolved, thereby improving perception accuracy and range.

CN118823535BActive Publication Date: 2026-11-17NORTH CHINA UNIVERSITY OF TECHNOLOGY
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202410882262.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-03
Publication Date
2026-11-17
Estimated Expiration
2044-07-03

AI Technical Summary

Technical Problem

In complex urban intersection environments, intelligent connected vehicles struggle to accurately detect moving targets due to limitations in the detection range of onboard sensors and the impact of occlusion. Existing collaborative perception methods suffer from low communication transmission efficiency and asynchrony between the spatiotemporal states of intelligent agents, with communication delay uncertainty having a significant impact, especially in complex intersection scenarios.

Method used

A collaborative perception method based on 3D point clouds is adopted. By establishing coordinate systems for each vehicle, extracting column features, V2X communication, compensating for time feature delay, and fusing spatial feature information, a ConvLSTM network and a multi-scale feature fusion module are designed to achieve temporal and spatial synchronization of point cloud features, reduce communication delay lag effects, and enhance environmental perception capabilities.

Benefits of technology

It improves the perception capabilities of intelligent connected vehicles in dynamic scenarios, reduces the lag effect caused by network latency, enhances the understanding of local fine-grained and global contextual information, and expands the range of environmental perception.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118823535B_ABST
    Figure CN118823535B_ABST
Patent Text Reader

Abstract

This invention provides a collaborative perception method for intelligent connected vehicles based on 3D point clouds in intersection environments. The method includes: S1 establishing coordinate systems for each vehicle and transforming the coordinates of point clouds collected from surrounding vehicles to the vehicle's own coordinate system; S2 processing the raw point cloud data of each vehicle using a volumetric feature extraction network to extract local spatial features; S3 transmitting the features collected by CAV2 to CAV1 using V2X communication; S4 capturing the temporal dependence of point cloud feature sequences and predicting feature state changes through a temporal feature delay compensation module; S5 fusing multi-resolution spatial feature information from multiple vehicle perspectives through a spatial feature information fusion compensation module; and S6 outputting various vehicle parameter information through a fully connected layer. This invention collaboratively perceives the point cloud features of the surrounding environment through vehicle-to-vehicle communication technology and improves the target detection performance of intelligent connected vehicles in long-distance, occluded scenarios at intersections by utilizing spatiotemporal feature compensation fusion, effectively mitigating the lag effect caused by network transmission communication delays.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, specifically to environmental perception technology, and particularly to a collaborative perception method for intelligent connected vehicles based on three-dimensional point clouds in an intersection environment. Background Technology

[0002] In complex urban intersection environments, intelligent connected vehicles often struggle to accurately detect surrounding moving targets solely relying on their own environmental perception sensors due to limitations in onboard sensor detection range and occlusion. Collaborative perception methods based on 3D point clouds can improve vehicle environmental awareness by fusing point cloud information from multiple intelligent connected vehicles, effectively overcoming the limitations of single-vehicle intelligent perception.

[0003] Existing collaborative perception methods suffer from low communication transmission efficiency and spatiotemporal asynchrony of agents in complex intersection scenarios. Chinese patent application publication number CN 117956507 A discloses a collaborative perception and data fusion method for intelligent connected vehicles. This method optimizes the offloading of collaborative perception and fusion tasks for multi-agent groups in intelligent connected vehicles through a two-layer collaborative perception and multi-source heterogeneous data fusion network model. However, it does not consider communication bandwidth limitations and needs improvement in handling communication delay uncertainties. Chinese patent application publication number CN 116893424 A discloses a multi-vehicle collaborative perception method, device, and storage medium based on lidar. This method proposes that the vehicle first performs precise registration of the point cloud information sent by the collaborating vehicle and then combines it with the vehicle's point cloud for information fusion. While this can reduce the blind spot of a single vehicle's perception, the precise registration of the collaborating vehicle's original point cloud results in excessive point cloud transmission volume, making it difficult for network communication bandwidth to meet transmission requirements. This inevitably leads to communication delays, and in intersection environments, the environmental point cloud information received by the intelligent connected vehicle is spatiotemporally out of sync with the actual environmental state information at the current moment. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a collaborative perception method for intelligent connected vehicles based on 3D point clouds in intersection environments, which aims to alleviate the lag effect caused by communication delays and improve target detection performance.

[0005] The technical solution adopted in this invention has the following steps:

[0006] A cooperative perception method for intelligent connected vehicles based on 3D point clouds in an intersection environment, the method comprising the following steps:

[0007] S1. Establish the coordinate system of each vehicle and transform the point cloud coordinates collected by the surrounding vehicles CAV2 to the coordinate system of the vehicle CAV1.

[0008] S2. Use a volumetric feature extraction network to process the raw point cloud data of each vehicle and extract local spatial features of the point cloud.

[0009] S3. Use V2X communication to transmit the features collected by CAV2 to CAV1;

[0010] S4. By using the time feature delay compensation module, the time dependence of point cloud feature sequences is captured and feature state changes are predicted, thereby reducing the feature lag effect caused by network communication transmission delay.

[0011] S5. By fusing multi-resolution spatial feature information from multiple vehicle perspectives through the spatial feature information fusion and compensation module, the range of environmental perception area is expanded.

[0012] S6 outputs vehicle target confidence, bounding box center point coordinates, geometric dimensions, and heading angle through a fully connected layer.

[0013] Preferably, in step S1, the coordinate transformation process is as follows: the local Cartesian coordinate systems of the self-vehicle CAV1 and the surrounding vehicle CAV2 are represented by X1O1Y1 and X2O2Y2 respectively. When the surrounding vehicle CAV2 enters the effective communication range, the self-vehicle CAV1 sends its pose state information to CAV2. The transformation matrix T is calculated by combining the relative pose information of the two vehicles. The point cloud P2 collected by CAV2 is projected onto the coordinate system of CAV1 using T to obtain the transformed point cloud P′1, so that the point cloud state space of the two vehicles is synchronized.

[0014] Preferably, in step S2, the geometric center of each vehicle is taken as the center point, with the CAV1 movement direction [y min ,y max ], horizontal direction [x min ,x max The region of interest is defined as a uniformly sized grid with a grid size of m. The PointPillar method is used to extract volumetric point cloud features without vertical height constraints within the grid, generating pseudo-image features f with dimensions (H, W, C).

[0015] Preferably, in step S3, the feature transmission process is as follows: the CAV2 feature map size is compressed to (H, W, C′), and it is sent to CAV1 via the V2X communication network. CAV1 then transmits the received feature f... v The dimensions are projected back to (H, W, C).

[0016] Preferably, in step S4, the historical frame feature sequence of CAV2 point cloud is used as input, and the features of future moments are predicted by the ConvLSTM algorithm to achieve time synchronization of the point cloud states of the two vehicles.

[0017] Preferably, the loss function of the ConvLSTM network is:

[0018]

[0019] Among them, Y t and These represent the true features and predicted features at time t, respectively.

[0020] Preferably, in step S5, the spatial feature information fusion compensation module includes: a cross-vehicle feature information aggregation submodule, a multi-resolution feature extraction submodule, and a multi-scale feature fusion submodule;

[0021] The cross-vehicle feature information aggregation submodule is used to align and stitch together the pseudo-image features of two vehicles; the multi-resolution feature extraction submodule is used to capture target details and global semantic information, and comprehensively perceive obstacle targets of different sizes; the multi-scale feature fusion submodule combines the advantages of shallow and deep features to perform multi-scale feature fusion, and generates feature expressions with rich semantic information.

[0022] Preferably, the cross-vehicle feature information aggregation process is as follows: define candidate aggregation regions with height H1 and width W1 (H1>H, W1>W), calculate the overlapping region of the two pseudo-image features, and perform channel-level alignment aggregation on the overlapping region using the Maxout operation.

[0023] Preferably, the multi-resolution feature extraction network includes four rounds of extraction operations, which extract shallow features F1 and deep features F2, F3, and F4 respectively.

[0024] Preferably, the multi-scale feature fusion submodule uses a bidirectional multi-scale information feature pyramid network to fuse features at different resolutions, and the convolutional fusion calculation is as follows:

[0025]

[0026] Among them, I i O represents the input features, and O represents the output value after weighted feature fusion. i and w j The learnable weights are ε = 0.001.

[0027] The present invention has the following beneficial effects:

[0028] The time delay compensation module based on ConvLSTM designed in this invention has the ability to predict spatiotemporal sequence data, and can extract target spatial information, enabling CAV1 to predict the characteristics of future moments based on the received vehicle historical characteristics, reducing the lag effect caused by network latency, and improving the vehicle's ability to perceive dynamic scenes.

[0029] The spatial feature information fusion compensation module designed in this invention includes three sub-modules: cross-vehicle feature information aggregation, multi-resolution feature extraction, and multi-scale feature fusion. It is used to fuse environmental point cloud features from surrounding vehicles, effectively enhancing the network's ability to understand local fine-grained information and global contextual information of the point cloud. Attached Figure Description

[0030] Figure 1 The attached figure is an overall flowchart of the present invention;

[0031] Figure 2 The attached figure is a schematic diagram of the coordinate system transformation of this invention;

[0032] Figure 3 The attached figure is a schematic diagram of cross-vehicle feature information aggregation according to the present invention;

[0033] Figure 4 The attached figure is a schematic diagram of the multi-resolution feature extraction network of the present invention;

[0034] Figure 5 The attached figure is a schematic diagram of the bidirectional multi-scale information feature pyramid network of the present invention. Detailed implementation method:

[0035] The invention will now be further described with reference to the accompanying drawings.

[0036] like Figure 1 As shown, this invention provides a cooperative perception method for intelligent connected vehicles based on 3D point clouds in an intersection environment. The method steps are as follows:

[0037] Step 1: Establish the coordinate system for each vehicle and transform the point cloud coordinates collected by the surrounding vehicles (CAV2) to the coordinate system of the vehicle (CAV1).

[0038] Step 2: Use a volumetric feature extraction network to process the raw point cloud data collected by each vehicle and extract local spatial features of the point cloud;

[0039] Step 3: Use V2X communication to transmit the features collected by CAV2 to CAV1;

[0040] Step 4: Capture the time dependence of point cloud feature sequences and predict feature state changes through the time feature delay compensation module to reduce the feature lag effect caused by network communication transmission delay.

[0041] Step 5: Use the spatial feature information fusion and compensation module to fuse multi-resolution spatial feature information from multiple vehicle perspectives to expand the environmental perception area.

[0042] Step 6: Output vehicle target confidence, bounding box center point coordinates, geometric dimensions, and heading angle through the fully connected layer.

[0043] like Figure 2 As shown, in Step 1, a coordinate system for each vehicle is established, and coordinate transformation is performed to fuse the environmental point cloud information collected by CAV1 and CAV2, so that the coordinate systems of the two vehicles are aligned. Specifically:

[0044] The local Cartesian coordinate systems of the vehicle CAV1 and the surrounding vehicle CAV2 are denoted as X1O1Y1 and X2O2Y2, respectively.

[0045] When a nearby vehicle CAV2 enters the effective communication range, the vehicle CAV1 sends its pose status information to CAV2.

[0046] The transformation matrix T is calculated by combining the relative pose information of the two workshops. The point cloud P2 collected by CAV2 is projected onto the coordinate system of CAV1 using T to obtain the transformed point cloud P′1, thereby realizing the synchronization of the point cloud state space of the two vehicles.

[0047] The specific calculation process is as follows:

[0048] θ = θ1 - θ2

[0049]

[0050] P′1=T*P2

[0051] Where θ1 and θ2 are the heading angles of CAV1 and CAV2, respectively, θ is the included heading angle between the two vehicles, and dx and dy are the horizontal and vertical distances between the two vehicles in the CAV2 coordinate system, respectively.

[0052] In Step 2, the geometric center of each vehicle is taken as the center point, and the direction of CAV1 movement [y] is used as the center point. min y max ], horizontal direction [x min x max The region of interest is divided into uniformly sized grids with a grid size of m. The PointPillar method is used to extract volumetric point cloud features without vertical height constraints within the grid, generating pseudo-image features f with size (H, W, C).

[0053] The feature extraction process of the PointPillar method is as follows:

[0054] An augmentation operation is performed on the scanned point p(x, y, z, r) after coordinate transformation, where r represents the reflection intensity. The arithmetic mean of the Euclidean distances from point p to all points within the cylinder is calculated and denoted as (x...). c y c , z c The offset of point p relative to the center of the cylinder along the x-axis and y-axis is denoted as x. p y pThe augmented point p is represented as (x, y, z, r, x). c y c , z c x p y p ), dimension D = 9;

[0055] Create a dense tensor of size (D, P, N), where P is the number of non-empty cylinders and N is a threshold for the number of points within each cylinder. If the number of points within a cylinder is greater than N, randomly sample the points within that cylinder; otherwise, zero-padded.

[0056] Effective feature augmentation of dense tensors is achieved by applying a linear layer, BatchNorm, and the ReLU activation function. A tensor of size (C, P, N) is generated, and then pooling is performed on each volume pillar to obtain a tensor of size (C, P), where C represents the number of channels.

[0057] The generated (C, P) tensor is mapped onto a two-dimensional plane based on the index value of each cylinder, generating a pseudo-image feature f of size (H, W, C), where the height H and width W are calculated as follows:

[0058] W = (x max -x min ) / m, H=(y max -y min ) / m

[0059] In step 3, the feature transmission process is as follows:

[0060] The feature map is progressively compressed along the channel dimension, and a 1×1 convolution kernel is used to compress the compressed feature f. v The number of channels C is reduced to C′, and the compressed feature f v The dimensions are (H, W, C′);

[0061] In a networked environment, CAV2 will feature f v Send to CAV1 via the network;

[0062] After receiving the features, CAV1 performs a reverse operation using a 1×1 convolution kernel, transforming the features f... v The dimensions are projected back to (H, W, C).

[0063] In Step 4, the historical frame feature sequence of CAV2 point cloud is used as input, and a time delay compensation module based on ConvLSTM is designed to predict the features of future moments. By integrating multi-scale spatial information and temporal relationships, the time synchronization of the point cloud states of the two vehicles is achieved.

[0064] The ConvLSTM network replaces the matrix multiplication operations between input gates and state gates, and between state gates, in traditional LSTM networks with convolution operations. This allows it to extract target spatial information, enabling CAV1 to predict future vehicle features based on received historical vehicle characteristics. This reduces the lag effect caused by network latency and improves the vehicle's ability to perceive dynamic scenes. The computation process is as follows:

[0065] i t =σ(W i *h t-1 +W i *x t +b i )

[0066] f t =σ(W f *h t-1 +W f *x t +b f )

[0067] o t =σ(W o *h t-1 +W o *x t +b o )

[0068] c t =f t ·c t-1 +i t ·tanh(W c *h t-1 W c *x t +b c )

[0069] h t =o t ·tanh(c t )

[0070] Among them, i t f t o t These represent the input gate, forget gate, and output gate, respectively; c represents the internal state unit; h represents the hidden state; * represents the convolution operation; and W represents the input gate, forget gate, and output gate, respectively. i W f W o and W c These are the input gate, forget gate, output gate, and long memory weights, respectively. σ and tanh are the activation functions, and b... i b f b o and b c These represent the bias respectively.

[0071] The loss function L of the ConvLSTM network uses Euclidean distance to constrain the error between the predicted features and the true features, and is calculated as follows:

[0072]

[0073] Among them, Y t and These represent the true features and predicted features at time t, respectively.

[0074] In Step 5, the spatial feature information fusion compensation module includes: a cross-vehicle feature information aggregation submodule, a multi-resolution feature extraction submodule, and a multi-scale feature fusion submodule.

[0075] like Figure 3 As shown, the cross-vehicle feature information aggregation submodule is used to align and concatenate the predicted features generated by CAV2 after time delay compensation with the local environment features obtained by CAV1 to generate aggregated features, specifically:

[0076] The center coordinates of CAV1 and CAV2 in the global coordinate system are set to (x1, y1) and (x2, y2) respectively, and the local environmental features acquired by the two vehicles are converted into pseudo-image features of size (H, W, C).

[0077] Define the candidate aggregation region with height H1 and width W1 (H1 > H, W1 > W), and calculate the width w1 and height h1 of the overlapping region of the two pseudo-image features:

[0078]

[0079] The overlapping regions are aggregated at the channel level using the Maxout operation, while retaining feature data from non-overlapping regions.

[0080] Fill the blank areas in the candidate aggregation region with zero elements to generate an aggregation feature of size (H1, W1, C);

[0081] The size (H1, W1, C) of the aggregated feature is compressed back to (H, W, C) using 3×3 convolution.

[0082] like Figure 4 As shown, the multi-resolution feature extraction submodule is used to capture target details and global semantic information, comprehensively perceiving obstacle targets of different sizes. The network includes four rounds of extraction operations, extracting shallow features F1 and deep features F2, F3, and F4 respectively, specifically:

[0083] The first round of extraction operation adjusts the number of channels through one Conv2D(256, 128, 3, 1) convolutional layer, and then extracts shallow features F1 through three Conv2D(128, 128, 3, 1) convolutional layers;

[0084] In the last three rounds of extraction, each round of operation uses a lightweight CoordinateAttention (CA) module, a pooling layer with a stride of 2, and four Conv2D (256, 256, 3, 1) convolutional layers to obtain deep features F2, F3, and F4, with sizes of 1 / 2, 1 / 4, and 1 / 8 of the original features, respectively.

[0085] like Figure 4 As shown, the Coordinating Attention (CA) module is used to simultaneously process the relationships and spatial location information between channels, specifically:

[0086] Pooling is performed along the horizontal and vertical directions respectively to obtain the embedding features in the c-th channel of the corresponding direction. and

[0087]

[0088] in, Let h represent the feature set along the horizontal direction of the c-th channel at height h. Let x represent the feature set along the vertical direction of the c-th channel at width w. c (h, j) represents the input feature along the horizontal direction in the c-th channel at height h and width j, x c (j, w) represents the input feature along the vertical direction in the c-th channel at height j and width w;

[0089] Embed features F in the horizontal and vertical directions h and F w The data are spliced ​​together to generate global receptive fields for horizontal and vertical features;

[0090] The number of channels is reduced by using 1×1 convolution, followed by batch normalization and non-linear activation, to output the intermediate feature map f:

[0091] f = ReLU[δ(contact(F)] h F w ))]

[0092] Where, f∈R 1×(W+H)×C / r r is the reduction parameter, ReLU is the non-linear activation function, δ represents 1×1 convolution, and contact represents the concatenation operation.

[0093] The intermediate feature map f is segmented using a 1×1 convolution to obtain a feature f of size (H, 1, C / r). H and features f with dimensions (1, W, C / r) W ;

[0094] The number of feature channels is restored using 3×3 convolution, and the horizontal and vertical attention weight vector sets g are calculated using the sigmoid activation function. H and g W ;

[0095] Generate the feature y of the c-th channel with attention weights. c :

[0096]

[0097] Where, x c (i, j), y c (i, j) represent the input and output features of the c-th channel at height i and width j, respectively. This represents the attention weight along the horizontal direction of the c-th channel at height i. This represents the attention weight along the vertical direction at the c-th channel with width .

[0098] like Figure 5 As shown, the multi-scale feature fusion submodule combines the advantages of shallow and deep features to fuse multi-scale features, thereby obtaining feature representations with rich semantics. Specifically:

[0099] Perform a 1×1 convolution operation on each of the deep features F2, F3, and F4 to obtain features F2′, F3′, and F4′ with the same number of channels as F1.

[0100] A two-layer bidirectional pyramid feature extraction network is constructed to transform the feature vector (F1′, F2′, F3′, F4′) into a feature fusion vector (D1, D2, D3, D4).

[0101] The bidirectional pyramid feature extraction network captures the correlation between two layers of features through lateral connections. When directly transmitting semantic information between shallow and deep features, in order to avoid the loss of semantic information due to frequent upsampling operations, each layer of the network b combines high-level semantic features and low-level positional information from bottom to top for downsampling. At the same time, skip connections enable the input features to participate in the fusion process.

[0102] Weights are added to features at different resolutions through a fast normalized convolutional fusion operation, enabling network b to adapt to learning differences in feature importance. The convolutional fusion calculation is as follows:

[0103]

[0104] Among them, Ii O represents the input features, and O represents the output value after weighted feature fusion. i and w j For learnable weights, ε = 0.001;

[0105] The detection head performs classification and regression tasks on the extracted multi-resolution features (D1, D2, D3, D4), and outputs parameters such as vehicle target confidence, bounding box center point coordinates, geometric dimensions, and heading angle.

[0106] This invention addresses existing problems in the field of collaborative perception, such as low communication transmission efficiency and the asynchrony between the transmitted spatiotemporal state of the agent and the actual state at the current moment in complex intersection scenarios. It designs a time feature delay compensation module with spatiotemporal sequence prediction capabilities to extract target state spatial information, enabling CAV1 to predict future features based on received vehicle historical features, reducing the lag effect caused by network communication transmission delays and improving the vehicle's perception capability of dynamic scenes. Furthermore, it constructs a spatial feature information fusion compensation module composed of three sub-modules: cross-vehicle feature information aggregation, multi-resolution feature extraction, and multi-scale feature fusion. This module fuses environmental point cloud features from surrounding vehicles, effectively enhancing the network's understanding of local fine-grained information and global contextual information in the point cloud.

[0107] Finally, it should be noted that the above specific embodiments are only used to illustrate the method and core idea of ​​the present invention, and are not intended to limit the present invention. For those skilled in the art, several equivalent modifications or substitutions can be made without departing from the spirit and principles of the present invention, and these equivalent modifications or substitutions are all included within the scope of patent protection of the present invention.

Claims

1. A cooperative perception method for intelligent connected vehicles based on 3D point clouds in an intersection environment, characterized in that, The method includes the following steps: Step S1: Establish the coordinate system of each vehicle and transform the point cloud coordinates collected by the surrounding vehicles CAV2 to the coordinate system of the vehicle CAV1. Step S2: Use a volumetric feature extraction network to process the original point cloud data of each vehicle and extract local spatial features of the point cloud; Step S3: Use V2X communication to transmit the features collected by CAV2 to CAV1; Step S4: Capture the time dependence of point cloud feature sequences and predict feature state changes through the time feature delay compensation module to reduce the feature lag effect caused by network communication transmission delay. Step S5: The spatial feature information fusion and compensation module fuses multi-resolution spatial feature information from multiple vehicle perspectives to expand the environmental perception area. Step S6: Output the vehicle target confidence, bounding box center point coordinates, geometric dimensions, and heading angle parameters through the fully connected layer; In step S4, the historical frame feature sequence of CAV2 point cloud is used as input, and the features of future moments are predicted by the ConvLSTM algorithm to achieve time synchronization of the point cloud status of the two vehicles. In step S5, the spatial feature information fusion compensation module includes: a cross-vehicle feature information aggregation submodule, a multi-resolution feature extraction submodule, and a multi-scale feature fusion submodule; The cross-vehicle feature information aggregation submodule is used to align and stitch together the pseudo-image features of two vehicles; the multi-resolution feature extraction submodule is used to capture target details and global semantic information, and comprehensively perceive obstacle targets of different sizes; the multi-scale feature fusion submodule combines the advantages of shallow and deep features to perform multi-scale feature fusion and generate feature expressions with rich semantic information. The process of cross-vehicle feature information aggregation is as follows: Determine the height H 1. Width W Candidate aggregation region of 1, , The overlapping region of the features of the two pseudo-images is calculated, and the overlapping region is aligned and aggregated at the channel level using the Maxout operation. The multi-resolution feature extraction submodule includes four rounds of extraction operations, which extract shallow features respectively. F 1 and deep features F 2. F 3. F 4; The multi-scale feature fusion submodule uses a bidirectional multi-scale information feature pyramid network to fuse features at different resolutions. The convolutional fusion calculation is as follows: in, For input features, O The output value after weighted feature fusion. and For learnable weights, ε =0.

001.

2. The intelligent connected vehicle cooperative perception method based on 3D point cloud in an intersection environment according to claim 1, characterized in that, In step S1, the coordinate transformation process is as follows: the local Cartesian coordinate systems of the vehicle CAV1 and the surrounding vehicles CAV2 are respectively converted using... and This means that when a nearby vehicle CAV2 enters the effective communication range, the vehicle CAV1 sends its pose status information to CAV2, and the transformation matrix is ​​calculated by combining the relative pose information between the two vehicles. T ,use T Point cloud acquired by CAV2 Projecting onto the CAV1 coordinate system yields the transformed point cloud. This synchronizes the point cloud state space of the two vehicles.

3. The intelligent connected vehicle cooperative perception method based on 3D point cloud in an intersection environment according to claim 1, characterized in that, In step S2, the geometric center of each vehicle is taken as the center point, and the direction of movement of CAV1 is... Horizontal direction Divide the region of interest into a uniformly sized grid, with the grid size set to [value missing]. m The PointPillar method is used to extract volumetric point cloud features without vertical height constraints within the mesh, generating a size of [missing information]. Pseudo-image features f .

4. The intelligent connected vehicle cooperative perception method based on 3D point cloud in an intersection environment according to claim 1, characterized in that, In step S3, the feature transfer process involves compressing the CAV2 feature map size to... It then sends the received features to CAV1 via the V2X communication network. f v Size projection back .

5. The intelligent connected vehicle cooperative perception method based on 3D point cloud in an intersection environment according to claim 1, characterized in that, The loss function for the ConvLSTM network is: in, and Represent t Real-time features and predicted features.

Citation Information

Patent Citations

  • Multi-vehicle cooperative sensing method and device based on laser radar and storage medium

    CN116893424A

  • Intelligent network connection automobile cooperative perception and data fusion method

    CN117956507A

  • Multi-mode-based automatic driving perception method and device, equipment and medium

    CN115879060A

  • Intelligent networked automobile cooperative sensing method based on model and data dual drive

    CN117372994A