Vehicle intention identification method and device, vehicle and medium

Abstract

The invention relates to a vehicle intention recognition method and device, a vehicle and a medium, and the method comprises the steps: obtaining target sensing data collected for a target vehicle, and extracting a target three-dimensional detection frame of the target vehicle from the target sensing data; according to the target three-dimensional detection frame, determining a target lamp language of a tail lamp of the target vehicle; obtaining target lane line information of a lane where the target vehicle is located currently from the target sensing data; and determining the driving intention of the target vehicle according to the target three-dimensional detection frame, the target lamp language and the target lane line information. According to the embodiment of the invention, the driving intention of the target vehicle can be identified by integrating different influence factors, so that the real driving intention of the target vehicle can be effectively identified. In addition, for the scene that the target vehicle is partially shielded or cut off, and consequently the vehicle cannot collect the information of the tail lamp of the target vehicle, the driving intention of the target vehicle can still be effectively recognized through the embodiment of the invention.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle control, and in particular to a method, device, vehicle and medium for identifying vehicle intention. Background Art

[0002] With the rapid development of autonomous driving technology, the recognition of vehicle intentions has become an important part of realizing intelligent transportation systems. Accurately understanding the behavior and intentions of other vehicles is crucial to ensuring road safety and improving traffic efficiency.

[0003] In practical applications, the intention of a vehicle can be identified based on its direction, speed, etc. Although this method can identify the vehicle's driving intention to a certain extent, this level of recognition is not enough for autonomous driving. The inability to accurately identify the intention of other vehicles will affect the safety of the vehicle's autonomous driving. Summary of the invention

[0004] In view of the above problems, a method, device, vehicle and medium for identifying vehicle intention are proposed to overcome the above problems or at least partially solve the above problems, including:

[0005] A method for identifying vehicle intention, the method comprising:

[0006] Acquire target perception data collected for a target vehicle, and extract a target three-dimensional detection frame of the target vehicle from the target perception data;

[0007] Determining a target light language of a taillight of the target vehicle according to the target three-dimensional detection frame;

[0008] Acquiring target lane line information of the lane where the target vehicle is currently located from the target perception data;

[0009] The driving intention of the target vehicle is determined according to the target three-dimensional detection frame, the target light language, and the target lane line information.

[0010] Optionally, extracting a target three-dimensional detection frame of the target vehicle from the target perception data includes:

[0011] The target perception data is input into a preset target three-dimensional detection network to obtain a target three-dimensional detection frame of the target vehicle.

[0012] Optionally, the feature extraction module of the target three-dimensional detection network includes:

[0013] Full-dimensional dynamic convolution and adaptive pooling layers.

[0014] Optionally, the regression parameters of the target three-dimensional detection network include:

[0015] A first heat map of the projection position of the center point of the three-dimensional detection frame on the image of the target perception data, a first offset, depth, uncertainty, length, width, height and angle of the target vehicle projected on the image of the target perception data by the center point of the three-dimensional detection frame;

[0016] The auxiliary learning parameters of the target three-dimensional detection network include:

[0017] A second heat map of the two-dimensional projection points of the corner points and the center point of the three-dimensional detection frame, a second offset of the two-dimensional projection points of the corner points and the center point of the three-dimensional detection frame to the center point of the two-dimensional frame, an intersection-and-union ratio, and a sine and cosine representation of the rotation angle of the three-dimensional detection frame;

[0018] The target three-dimensional detection network uses a focal loss function based on Gaussian kernel weighting to calculate the loss of the first heat map and the second heat map;

[0019] The target three-dimensional detection network uses the Laplace arbitrary uncertainty loss function to calculate the loss of depth and uncertainty;

[0020] The target three-dimensional detection network uses an absolute value loss function to calculate the loss of the length, width, height, first offset and second offset of the target vehicle;

[0021] The target three-dimensional detection network uses a cross entropy loss function to calculate the loss of the angle of the target vehicle;

[0022] The target three-dimensional detection network uses an intersection-over-union loss function to calculate the loss of the intersection-over-union ratio;

[0023] The target 3D detection network uses a smooth absolute value loss function to calculate the loss represented by the sine and cosine of the 3D detection box rotation angle.

[0024] Optionally, the target three-dimensional detection frame includes at least one of the following information:

[0025] The center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the heading angle of the target vehicle.

[0026] Optionally, determining the driving intention of the target vehicle according to the target three-dimensional detection frame, the target light language, and the target lane line information includes:

[0027] Determine a target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located according to the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information;

[0028] Determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs;

[0029] The driving intention of the target vehicle is determined according to the target lane line information, the target angle interval, the target distance interval, and the target light language.

[0030] Optionally, when the target light language is empty, determining the driving intention of the target vehicle according to the target three-dimensional detection frame, the target light language, and the target lane line information includes:

[0031] Determine a target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located according to the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information;

[0032] Determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs;

[0033] The driving intention of the target vehicle is determined according to the target lane line information, the target angle interval, and the target distance interval.

[0034] Optionally, determining the target light language of the taillight of the target vehicle according to the target three-dimensional detection frame includes:

[0035] Generating a target tracking sequence of the target vehicle according to the target three-dimensional detection frame;

[0036] Cutting the target tracking sequence to obtain a target tail sequence of the target vehicle;

[0037] The target tail sequence is input into a preset light language detection network to obtain the target light language of the taillight of the target vehicle.

[0038] Optionally, the target three-dimensional detection frame includes at least a first three-dimensional detection frame at a current moment and a second three-dimensional detection frame at a previous moment, and generating a target tracking sequence of the target vehicle according to the target three-dimensional detection frame includes:

[0039] Predicting a third three-dimensional detection frame at a next moment of the second three-dimensional detection frame according to the second three-dimensional detection frame;

[0040] Fusing the first three-dimensional detection frame and the third three-dimensional detection frame to obtain a fourth three-dimensional detection frame;

[0041] When the object in the fourth three-dimensional detection frame matches the object in the first three-dimensional detection frame, the tracking sequence of the target vehicle is updated according to the first three-dimensional detection frame to obtain the target tracking sequence.

[0042] The present invention also provides a vehicle intention recognition device, the device comprising:

[0043] A three-dimensional detection frame extraction module is used to obtain target perception data collected for a target vehicle, and extract a target three-dimensional detection frame of the target vehicle from the target perception data;

[0044] A light language determination module, used to determine the target light language of the taillight of the target vehicle according to the target three-dimensional detection frame;

[0045] A lane line acquisition module, used to acquire target lane line information of the lane where the target vehicle is currently located from the target perception data;

[0046] The intention recognition module is used to determine the driving intention of the target vehicle based on the target three-dimensional detection frame, the target light language, and the target lane line information.

[0047] Optionally, the three-dimensional detection frame extraction module is used to input the target perception data into a preset target three-dimensional detection network to obtain a target three-dimensional detection frame of the target vehicle.

[0048] Optionally, the feature extraction module of the target three-dimensional detection network includes: full-dimensional dynamic convolution and adaptive pooling layer.

[0049] Optionally, the regression parameters of the target three-dimensional detection network include:

[0050] A first heat map of the projection position of the center point of the three-dimensional detection frame on the image of the target perception data, a first offset, depth, uncertainty, length, width, height and angle of the target vehicle projected on the image of the target perception data by the center point of the three-dimensional detection frame;

[0051] The auxiliary learning parameters of the target three-dimensional detection network include:

[0052] A second heat map of the two-dimensional projection points of the corner points and the center point of the three-dimensional detection frame, a second offset of the two-dimensional projection points of the corner points and the center point of the three-dimensional detection frame to the center point of the two-dimensional frame, an intersection-and-union ratio, and a sine and cosine representation of the rotation angle of the three-dimensional detection frame;

[0053] The target three-dimensional detection network uses a focal loss function based on Gaussian kernel weighting to calculate the loss of the first heat map and the second heat map;

[0054] The target three-dimensional detection network uses the Laplace arbitrary uncertainty loss function to calculate the loss of depth and uncertainty;

[0055] The target three-dimensional detection network uses an absolute value loss function to calculate the loss of the length, width, height, first offset and second offset of the target vehicle;

[0056] The target three-dimensional detection network uses a cross entropy loss function to calculate the loss of the angle of the target vehicle;

[0057] The target three-dimensional detection network uses an intersection-over-union loss function to calculate the loss of the intersection-over-union ratio;

[0058] The target 3D detection network uses a smooth absolute value loss function to calculate the loss represented by the sine and cosine of the 3D detection box rotation angle.

[0059] Optionally, the target three-dimensional detection frame includes at least one of the following information:

[0060] The center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the heading angle of the target vehicle.

[0061] Optionally, the intention recognition module is used to determine the target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located based on the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information; determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs; determine the driving intention of the target vehicle based on the target lane line information, the target angle interval, the target distance interval, and the target light language.

[0062] Optionally, the intention recognition module is also used to determine, when the target light language is empty, the target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located according to the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information; determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs; determine the driving intention of the target vehicle according to the target lane line information, the target angle interval, and the target distance interval.

[0063] Optionally, the light language determination module is used to generate a target tracking sequence of the target vehicle based on the target three-dimensional detection frame; crop the target tracking sequence to obtain a target tail sequence of the target vehicle; input the target tail sequence into a preset light language detection network to obtain the target light language of the taillights of the target vehicle.

[0064] Optionally, the target three-dimensional detection frame includes at least a first three-dimensional detection frame at a current moment and a second three-dimensional detection frame at a previous moment, and the light language determination module is used to predict a third three-dimensional detection frame of the second three-dimensional detection frame at the next moment based on the second three-dimensional detection frame; fuse the first three-dimensional detection frame and the third three-dimensional detection frame to obtain a fourth three-dimensional detection frame; when an object in the fourth three-dimensional detection frame matches an object in the first three-dimensional detection frame, update the tracking sequence of the target vehicle according to the first three-dimensional detection frame to obtain the target tracking sequence.

[0065] The present invention also provides a vehicle, characterized in that it includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, and when the computer program is executed by the processor, the method for identifying vehicle intention as described above is implemented.

[0066] The present invention also provides a computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the method for identifying vehicle intention as described above is implemented.

[0067] Beneficial effects of the present invention:

[0068] In the present invention, target perception data collected for the target vehicle is obtained, and a target three-dimensional detection frame of the target vehicle is extracted from the target perception data; the target light language of the taillights of the target vehicle is determined based on the target three-dimensional detection frame; the target lane line information of the lane where the target vehicle is currently located is obtained from the target perception data; the driving intention of the target vehicle is determined based on the target three-dimensional detection frame, the target light language, and the target lane line information. Through the embodiment of the present invention, different influencing factors can be combined to identify the driving intention of the target vehicle, thereby effectively identifying the true driving intention of the target vehicle. In addition, for the scenario where the target vehicle is partially blocked or truncated, resulting in the inability of the vehicle to collect information about the taillights of the target vehicle, the embodiment of the present invention can still effectively identify the driving intention of the target vehicle. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 is a flow chart of steps of a method for identifying vehicle intention according to an embodiment of the present invention;

[0070] Figure 2a is a flowchart of another method for identifying vehicle intention according to an embodiment of the present invention;

[0071] Figure 2b is a schematic diagram of the structure of a three-dimensional target detection network according to an embodiment of the present invention;

[0072] Figure 2cThis is a flowchart of steps for analyzing the light language of the taillights of a target vehicle according to an embodiment of the present invention;

[0073] Figure 2d is a flowchart of steps for identifying vehicle intentions according to an embodiment of the present invention;

[0074] Figure 3 is a flowchart of the steps of another method for identifying vehicle intention according to an embodiment of the present invention;

[0075] Figure 4 It is a structural schematic diagram of a vehicle intention recognition device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] Reference Figure 1 , shows a flowchart of a method for identifying vehicle intention according to an embodiment of the present invention, which may include the following steps:

[0078] Step 101: Acquire target perception data collected for a target vehicle, and extract a target three-dimensional detection frame of the target vehicle from the target perception data.

[0079] Among them, the target perception data may refer to the perception data collected from the target vehicle; illustratively, the target perception data may be image data of the target perception data, or may be point cloud data, which is not limited in the embodiment of the present invention.

[0080] In actual applications, during the autonomous driving process, the vehicle needs to identify the driving intentions of other surrounding vehicles in order to ensure the safety of autonomous driving; then, the autonomous driving decision is generated based on the driving intentions of other surrounding vehicles; in order to identify the driving intentions of other surrounding vehicles, the vehicle can first obtain target perception data collected for other surrounding vehicles from a laser radar or an image acquisition device for target perception data; among them, other surrounding vehicles can refer to target vehicles.

[0081] After obtaining the target perception data collected for the target vehicle, the vehicle can extract the target three-dimensional detection frame for the target vehicle from the target perception data; wherein the target three-dimensional detection frame can refer to the 3D (Three-Dimensional) detection frame for the target vehicle; the 3D detection frame is a geometric representation method used to represent the position, size and direction of an object in three-dimensional space. Unlike a two-dimensional detection frame (usually used for object detection in an image of target perception data), a 3D detection frame provides three-dimensional information of an object in the real world. Exemplarily, the target three-dimensional detection frame can be used to provide three-dimensional information of the target vehicle in the real world.

[0082] Step 102: Determine the target light language of the taillight of the target vehicle according to the target three-dimensional detection frame.

[0083] After obtaining the target three-dimensional detection frame for the target vehicle, the target three-dimensional detection frame can be analyzed to determine the target light language of the taillight of the target vehicle. The target light language may include: go straight, brake, turn left, turn right and double flash, etc., which is not limited in the embodiment of the present invention.

[0084] Step 103: Obtain target lane line information of the lane in which the target vehicle is currently located from the target perception data.

[0085] After obtaining the target perception data, the vehicle can also obtain the target lane line information of the lane in which the target vehicle is currently located from the target perception data; wherein the target lane line information may include the coordinate information of the lane line, the semantic information of the lane line, etc., and the embodiment of the present invention does not limit this.

[0086] Step 104: Determine the driving intention of the target vehicle based on the target three-dimensional detection frame, the target light language, and the target lane line information.

[0087] After obtaining the target three-dimensional detection frame, target light language, and target lane line information, the vehicle can predict the target vehicle's driving intention based on the target vehicle's three-dimensional information in the target three-dimensional detection frame in the real world, as well as the target light language and target lane line information; the driving intention may include going straight, braking, turning left, turning right, etc., and the embodiment of the present invention does not limit this. Compared to the method of identifying the vehicle's intention based on the vehicle's direction, speed, etc., the embodiment of the present invention combines different influencing factors to identify the target vehicle's driving intention, and can effectively identify the target vehicle's true driving intention.

[0088] In some feasible embodiments, when the target vehicle and the host vehicle are traveling in the same direction, the target light language of the target vehicle's taillights can be determined based on the target three-dimensional detection frame; when the target vehicle and the host vehicle are traveling in the opposite direction, the target three-dimensional detection frame may not include the target vehicle's taillights, and at this time the target light language of the target vehicle's taillights cannot be obtained based on the target three-dimensional detection frame.

[0089] The target 3D detection frame provides the precise position and posture information of the vehicle in 3D space, the target light language (such as brake lights, turn signals, etc.) provides direct driving intention information, and the target lane line information provides the position and motion information of the vehicle relative to the lane line. Fusion of these different modal data helps to fully understand the behavior and intention of the vehicle. This combination can improve the redundancy and reliability of the system. Even if a data parsing error occurs, other information can still provide effective supplement and verification. By fusing multi-source data, we can better cope with complex traffic scenarios and improve the safety of the system. Specifically:

[0090] When the target light language of the taillights of the target vehicle can be determined based on the target three-dimensional detection frame, the vehicle can determine the driving intention of the target vehicle based on the target three-dimensional detection frame, the target light language, and the target lane line information; and when the target light language of the taillights of the target vehicle cannot be determined based on the target three-dimensional detection frame, the vehicle can also determine the driving intention of the target vehicle based on the target three-dimensional detection frame and the target lane line information. For the scenario where the target vehicle is partially blocked or cut off, so that the vehicle cannot collect the information of the taillights of the target vehicle, the embodiment of the present invention can still effectively identify the driving intention of the target vehicle.

[0091] In an embodiment of the present invention, target perception data collected for a target vehicle is obtained, and a target three-dimensional detection frame of the target vehicle is extracted from the target perception data; the target light language of the taillights of the target vehicle is determined based on the target three-dimensional detection frame; the target lane line information of the lane in which the target vehicle is currently located is obtained from the target perception data; the driving intention of the target vehicle is determined based on the target three-dimensional detection frame, the target light language, and the target lane line information. Through the embodiment of the present invention, different influencing factors can be combined to identify the driving intention of the target vehicle, thereby effectively identifying the true driving intention of the target vehicle. In addition, for scenarios where the target vehicle is partially blocked or truncated, resulting in the inability of the vehicle to collect information about the taillights of the target vehicle, the embodiment of the present invention can still effectively identify the driving intention of the target vehicle.

[0092] Reference Figure 2a , shows a flowchart of another method for identifying vehicle intention according to an embodiment of the present invention, which may include the following steps:

[0093] Step 201: Acquire target perception data collected for a target vehicle, and input the target perception data into a preset target three-dimensional detection network to obtain a target three-dimensional detection frame of the target vehicle.

[0094] In actual applications, in order to ensure the safety of autonomous driving, the vehicle that is driving automatically needs to identify the driving intentions of surrounding target vehicles during the process; then, the autonomous driving decision is generated based on the driving intentions of the target vehicles; in order to identify the driving intentions of the target vehicles, the vehicle can first obtain target perception data collected for the target vehicle from a laser radar or an image acquisition device for target perception data.

[0095] After obtaining the target perception data collected for the target vehicle, the vehicle can input the target perception data into a preset target three-dimensional detection network; the target three-dimensional detection network can output a target three-dimensional detection frame of the target vehicle based on the target perception data.

[0096] In one embodiment of the present invention, the feature extraction module of the target three-dimensional detection network may include: full-dimensional dynamic convolution and adaptive pooling layer.

[0097] In some feasible embodiments, in order to enhance the feature extraction capability of the target three-dimensional detection network and to better retain the details in the perceived data, the present invention can design a target three-dimensional detection network; specifically, a feature extraction module of the target three-dimensional detection network can be designed, and the feature extraction module of the target three-dimensional detection network can include a full-dimensional dynamic convolution and an adaptive pooling layer.

[0098] Among them, the full-dimensional dynamic convolution can be ODConvl (Fully Dynamic Convolution); full-dimensional dynamic convolution is a technology in convolutional neural networks that improves the flexibility and adaptability of the model by dynamically adjusting the weights of the convolution kernel. Since the full-dimensional dynamic convolution allows the model to dynamically adjust the convolution kernel according to the different characteristics of the input data, designing the full-dimensional dynamic convolution in the feature extraction module of the target three-dimensional detection network can enhance the adaptability of the model; in addition, by dynamically generating the convolution kernel, the model can more effectively extract the features of the input data; in addition, compared with the traditional fixed convolution kernel, the full-dimensional dynamic convolution can reduce the number of parameters in the model while maintaining performance, which helps to reduce overfitting and reduce computational costs. The dynamic convolution kernel can be adjusted according to the characteristics of the input data, which helps the model to show better generalization ability on unseen data. The full-dimensional dynamic convolution can be easily integrated into the existing convolutional neural network architecture, providing greater flexibility and scalability for model design. In addition, the full-dimensional dynamic convolution uses a dynamic attention mechanism to selectively focus on important features, and extends the dynamic convolution to multiple dimensions, which can more finely adapt to the characteristics of the input data, thereby improving the effect of feature extraction. At the same time, a parallel strategy is adopted to simultaneously learn attention in different dimensions. This strategy allows the network to process the features of each dimension more efficiently while ensuring complementarity and synergy between dimensions.

[0099] The adaptive pooling layer can be AdaPool (Adaptive Pooling Layer); the adaptive pooling layer is a pooling technology in deep learning, which can automatically adjust the size of the pooling area to ensure that the output feature map has a fixed size. In practical applications, the adaptive pooling layer can ensure that the size of the output feature map is a pre-set fixed size regardless of the size of the input feature map. This has good adaptability for network architectures that require fixed-size outputs (such as fully connected layers). Unlike traditional fixed-size pooling layers, the adaptive pooling layer does not require the input feature map to have a specific size, which provides greater flexibility for the design of the target 3D detection network. Since the adaptive pooling layer can automatically adjust to inputs of different sizes, it is not necessary to adjust the network structure for input data of different sizes when designing the target 3D detection network. In addition, the adaptive pooling layer can distribute the pooling operation more evenly when processing inputs of different sizes, which helps to reduce information loss.

[0100] As an example, the target three-dimensional detection network is described below as a DLA (Deep Layer Aggregation) 34 network: the DLA 34 network is composed of a ResNet 34 network based on a multi-level fusion structure, which can be divided into two modules: iterative deep aggregation and hierarchical deep aggregation.

[0101] Among them, the iterative deep aggregation module aggregates features from the shallowest scale, introduces connections and jump connections between different levels and iterative stages to effectively utilize multi-level information, so that shallow features can also be passed to the deep layer of the network.

[0102] The hierarchical deep aggregation module aggregates features from different layers by introducing connections between different layers of the network, which can take advantage of the features extracted at different layers and fuse multi-scale and multi-level information. The 3×3 convolution of the DLA34 network is replaced with ODConv, i.e., full-dimensional dynamic convolution, and the pooling layer structure is replaced with AdaPool, i.e., adaptive pooling layer structure.

[0103] In practical applications, ODConv can be used to enhance the feature extraction capability of the DLA34 network: specifically, ODConv utilizes a multi-dimensional attention mechanism and a parallel strategy to learn different attentions in the four dimensions of the convolution kernel.

[0104] Among them, the four dimensions include the input channel of the convolution kernel, the output channel of the convolution kernel, the spatial dimension of the convolution kernel, and the number of convolution kernels.

[0105] The four attentions complement each other and are multiplied with the convolution kernels in the order of the number of kernels, input channels, output channels, and spatial dimensions, which significantly enhances the feature extraction capability of the DLA34 network.

[0106] like Figure 2b As shown in the figure, firstly, GAP (Global Average Pooling) is performed on the input feature x, then the FC layer (Fully Connected Layer) is used to reduce the dimension, and then the ReLU (Rectified Linear Unit) activation function is used. Then, the FC layer is used to increase the dimension on the four dimensions of the convolution kernel and then input to the Sigmoid activation function. Finally, the output results on the four dimensions are multiplied, and the output results of all convolution kernels are added and then dot multiplied with the input feature to obtain the output feature y. The formula for ODConv output is as follows:

[0107]

[0108] Among them, x represents the input feature, y represents the output feature, and w i represents the convolution kernel, α wi Represents the convolution kernel w i Attention in the number dimension, α fi Represents the convolution kernel w i Output channel dimension attention, α ci Represents the convolution kernel w iAttention in the input channel dimension, α si Represents the convolution kernel w i Attention in spatial dimensions, ⊙ represents multiplication operations in different dimensions.

[0109] In the DLA34 network, AdaPool can be used to retain more detailed features: AdaPool can be divided into two parts:

[0110] One part is eDSCPool (Efficient Dense Semantic Correspondence Pooling), which is based on the Dyess similarity coefficient index to learn the image features of the target perception data for a specific area. The Dyess similarity coefficient is a distance measurement method, and the calculation formula is as follows:

[0111]

[0112] in, is the output based on the DSC coefficient in the region, C is the region for measuring distance, c is the coordinate index of C, is the average value in c, a i，c is the i-th value in c.

[0113] At the same time, considering that zero-value coefficients will be assigned zero weights, the exponential is used to learn the features. The calculation formula is as follows:

[0114]

[0115] in, is the output eigenvalue, R is the pooling core area, i is the coordinate index of R, j is the coordinate index of R, and a i is the activation value, is the average activation value of R, is the i-th weight coefficient, is the jth weight coefficient.

[0116] eDSCPool average pooling uses equal weights for all input vectors within the kernel region. Therefore, the merged output is affected by outliers within the region. By improving the calculation of the regional average, the impact of outliers in creating the pooled image during forward propagation and in the gradient calculation during backpropagation can be limited. The basic assumption of this method is that geometrically close observations are more similar than geometrically farther observations. In addition, since distance methods may produce artifacts when applied directly to the input region, the use of a similarity measure is more suitable for the regional nature of pooling, so an exponent (e) of the similarity between the activation vector and the average activation is used. This makes the pooling method differentiable during backpropagation, because at least the minimum gradient is calculated for each position, ensuring that a non-zero weight is assigned, and also reduces the possibility of the gradient vanishing problem, thereby improving the preservation of activations.

[0117] The other part is eMPool (Efficient Multi-scale Pooling), which learns the image features of the target perception data for a specific area based on the natural index, ensuring that larger activations will have a greater impact on the final output, while also ensuring that the minimum weight value is assigned to the lowest activation. The calculation formula is as follows:

[0118]

[0119] in, is the output eigenvalue, R is the pooling core area, i is the coordinate index of R, a i is the activation value, is the weight coefficient.

[0120] Next, the two parts can be combined using a learnable parameter β:

[0121]

[0122] in, is the output feature value of AdaPool, and β is the weight mask.

[0123] eMPool is based on the natural exponent (e), ensuring that larger activations will have a greater impact on the final output, while also ensuring that the lowest activations are assigned the smallest weight value. The weights in eMPool are used for nonlinear transformations based on the corresponding activation values, and activations with higher values ​​will dominate over those with lower values. Since most pooling operations are performed on high-dimensional feature spaces, highlighting activations with greater influence is more balanced than selecting only the largest activations. In the latter case, discarding most activations risks losing important information. The output of eMPool is produced by summing all weighted activations within the kernel region R. The result is based on a probability distribution proportional to the values ​​of neighboring activations within the kernel region.

[0124] The two pooling kernel properties can complement each other to discover the most informative features in the kernel region. Using a learnable weight mask β to balance the two methods has the advantage of creating a comprehensive pooling strategy that provides a good balance between preserving informative details and local image structure.

[0125] In one embodiment of the present invention, the regression parameters of the target three-dimensional detection network include: a first heat map of the projection position of the center point of the three-dimensional detection frame on the image of the target perception data, a first offset, depth, uncertainty, length, width, height and angle of the target vehicle projected on the image of the target perception data by the center point of the three-dimensional detection frame;

[0126] The auxiliary learning parameters of the target 3D detection network include: a second heat map of the 2D projection points of the corner points and the center point of the 3D detection frame, a second offset of the 2D projection points of the corner points and the center point of the 3D detection frame to the center point of the 2D frame, an intersection-over-union ratio, and sine and cosine representations of the rotation angle of the 3D detection frame;

[0127] Among them, the target three-dimensional detection network uses the focal loss function based on Gaussian kernel weighting to calculate the loss of the first heat map and the second heat map; the target three-dimensional detection network uses the Laplace arbitrary uncertainty loss function to calculate the loss of depth and uncertainty; the target three-dimensional detection network uses the absolute value loss function to calculate the loss of the length, width, height, first offset and second offset of the target vehicle; the target three-dimensional detection network uses the cross entropy loss function to calculate the loss of the angle of the target vehicle; the target three-dimensional detection network uses the intersection-and-union loss function to calculate the loss of the intersection-and-union ratio; the target three-dimensional detection network uses the smoothed absolute value loss function to calculate the loss represented by the sine and cosine of the rotation angle of the three-dimensional detection box.

[0128] In some feasible embodiments, regression parameters and auxiliary learning parameters of the target three-dimensional detection network may be designed to improve the accuracy of the target three-dimensional detection network.

[0129] Specifically, the regression parameters of the target three-dimensional detection network may refer to parameters directly used to regress the three-dimensional detection frame, which may include: a first heat map of the projection position of the center point of the three-dimensional detection frame on the image of the target perception data, a first offset of the projection of the center point of the three-dimensional detection frame on the image of the target perception data, depth, uncertainty, length, width, height and angle of the target vehicle.

[0130] Among them, the first heat map may refer to a heat map of the projection position of the center point of the three-dimensional detection frame on the image of the target perception data; the first offset of the projection of the center point of the three-dimensional detection frame on the image of the target perception data may refer to the offset from the first heat map to the actual center point; the depth may refer to the distance from the target vehicle to the laser radar or image acquisition device of the vehicle; the uncertainty may refer to the estimate of the uncertainty of the depth prediction; the angle of the target vehicle may refer to the orientation or rotation angle of the target vehicle.

[0131] Regression parameters can help the network predict the location information of three-dimensional objects more accurately, including the coordinates of the center point, width, height, and depth of the object. It can also understand the direction and orientation of the object in three-dimensional space, thereby more comprehensively understanding the state of the object. In addition, the sizes of different objects vary greatly, and regression parameters can dynamically adjust the size of the detection box to better adapt to objects of different scales and improve the robustness and accuracy of detection. By optimizing the regression parameters, the error between the detection box and the boundary of the real object can be reduced, and the accuracy of detection can be improved.

[0132] Auxiliary learning parameters of the target three-dimensional detection network may include: a second heat map of the two-dimensional projection points of the corner points and center point of the three-dimensional detection box, a second offset of the two-dimensional projection points of the corner points and center point of the three-dimensional detection box to the center point of the two-dimensional box, an intersection-and-union ratio, and sine and cosine representations of the three-dimensional detection box rotation angle.

[0133] Among them, the second heat map can be used to assist in locating the projection of the three-dimensional detection frame on the image of the target perception data; the second offset can refer to the offset of the two-dimensional projection points of the eight corner points and the center point of the three-dimensional detection frame to the center point of the two-dimensional frame; the intersection and union ratio can refer to the intersection and union ratio of the three-dimensional detection frame projected onto the image of the target perception data and the real two-dimensional frame, and the intersection and union ratio can be used to evaluate the overlap between the predicted frame and the real frame. The sine and cosine representations of the three-dimensional detection frame rotation angle can refer to the rotation angle of the target vehicle represented by trigonometric functions.

[0134] In some feasible embodiments, in order to more accurately locate the target, a Gaussian kernel weighted focal loss function suitable for handling class imbalance problems can be used to calculate the loss of the first heat map and the second heat map; a Laplace arbitrary uncertainty loss function suitable for estimating the uncertainty of data can be used to calculate the loss of depth and uncertainty; an absolute value loss function can be used to calculate the loss of the length, width, height, first offset, and second offset of the target vehicle; in order to more accurately classify the posture of the target vehicle, a cross entropy loss function can be used to calculate the loss of the angle of the target vehicle; an intersection-over-union loss function can be used to calculate the loss of the intersection-over-union ratio; in order to reduce the sensitivity of the model to the rotation angle, so that the model can maintain good performance when facing different rotation angles. The loss represented by the sine and cosine of the rotation angle of the three-dimensional detection box can be represented by the smooth absolute value loss function.

[0135] In addition, by training auxiliary learning parameters to predict 2D projection features associated with 3D targets, the network's understanding of object geometry can be enhanced. In standard target 3D detection boxes, usually only 3D bounding box annotations are available. By leveraging auxiliary learning, these annotations can be converted into more supervisory signals, which helps the network learn richer and more fine-grained feature representations. Auxiliary learning can help the network avoid overfitting when learning the main task. By introducing additional learning objectives, the generalization ability of the network can be increased, making it more robust when processing unseen data.

[0136] In one embodiment of the present invention, the target three-dimensional detection frame includes at least one of the following information: the coordinates of the center point of the target vehicle, the length, width and height of the target vehicle, and the heading angle of the target vehicle.

[0137] In some feasible embodiments, the target three-dimensional detection frame may include the center point coordinates (x, y, z) of the target vehicle, the length l of the target vehicle, the width w of the target vehicle, the height h of the target vehicle, and the heading angle θ of the target vehicle.

[0138] In some feasible embodiments, for the constructed 3D detection network, the KITTI dataset can be divided into a training set and a test set according to a:b (which can be 7:3, 4:1 or 9:1), and the data can be preprocessed (for example: standardization of image and point cloud data, data enhancement, format conversion, etc.). Relevant training parameters are set, including learning rate, batch size, optimizer, etc.

[0139] Next, the three-dimensional detection network can be trained: first, the training set is input into the three-dimensional detection network, the output is calculated, and the loss is calculated with the true label. Next, the model parameters are updated by the back propagation algorithm to minimize the loss. Then, after each round of training, the model performance is evaluated using the test set, and the loss or other performance indicators are recorded. Finally, the model with the lowest loss on the test set is selected as the optimal model and saved to obtain the target three-dimensional detection network. Exemplarily, the number of iterations can be 100, the batch size can be 8, the initial learning rate can be 0.001, the optimizer can be the stochastic gradient descent method, and the momentum factor can be 0.937, which is not limited in the embodiment of the present invention.

[0140] Step 202: Generate a target tracking sequence for the target vehicle according to the target three-dimensional detection frame.

[0141] Among them, the target tracking sequence may refer to a collection of position and state information of a target vehicle in a plurality of consecutive frames. The target tracking sequence may be used to ensure that the target vehicle can be continuously identified and tracked in a video or a series of images, thereby recording the movement trajectory of the target vehicle.

[0142] In some feasible embodiments, the target 3D detection frame may include multiple 3D detection frames, which may be 3D detection frames generated for the target vehicle at different times. After obtaining the target 3D detection frame, a target tracking sequence for the target vehicle may be generated based on the multiple 3D detection frames.

[0143] In one embodiment of the present invention, step 202 may be implemented by the following sub-steps:

[0144] Sub-step 11: the target 3D detection frame includes at least a first 3D detection frame at a current moment and a second 3D detection frame at a previous moment, and a third 3D detection frame of the second 3D detection frame at a next moment is predicted based on the second 3D detection frame.

[0145] In some feasible embodiments, the target three-dimensional detection frame may include at least a first three-dimensional detection frame obtained at the current moment and a second three-dimensional detection frame at a previous moment relative to the current moment; wherein the first three-dimensional detection frame and the second three-dimensional detection frame are measurement quantities.

[0146] After obtaining the first three-dimensional detection frame and the second three-dimensional detection frame, a third three-dimensional detection frame at a time point of downward movement relative to a time point when the second three-dimensional detection frame is generated can be predicted based on the second three-dimensional detection frame.

[0147] Sub-step 12: Fusing the first three-dimensional detection frame and the third three-dimensional detection frame to obtain a fourth three-dimensional detection frame.

[0148] In practical applications, both the measured quantity and the predicted quantity may not be accurate; therefore, the first three-dimensional detection frame and the third three-dimensional detection frame can be fused to obtain a fourth three-dimensional detection frame to combine the predicted quantity and the measured quantity to ensure the accuracy and reliability of the fourth three-dimensional detection frame.

[0149] Sub-step 13: When the object in the fourth three-dimensional detection frame matches the object in the first three-dimensional detection frame, the tracking sequence of the target vehicle is updated according to the first three-dimensional detection frame to obtain a target tracking sequence.

[0150] After obtaining the fourth three-dimensional detection frame, if it is determined that the object in the fourth three-dimensional detection frame matches the object in the first three-dimensional detection frame, the tracking sequence of the target vehicle can be updated according to the first three-dimensional detection frame to obtain a target tracking sequence.

[0151] For example, a 3D Kalman filter algorithm may be used to predict the state based on the second three-dimensional detection frame, and a constant velocity model may be used to approximate the displacement between frames to obtain a prediction value:

[0152] Among them, the prediction stage formula is:

[0153]

[0154] in, represents the state at time k-1 (i.e., the second 3D detection frame), They represent the state at time k (i.e., the third 3D detection frame), A k represents the state transfer matrix, Indicates A k The transpose of B k Represents the control matrix, which is used to control external input; represents the external input at time k, represents the estimated covariance matrix of the k-1 time error, represents the estimated covariance matrix of the error at time k, Q k Represents the covariance matrix of the process noise.

[0155] Then, the Kalman gain is used to perform weighted fusion of the measured and predicted quantities to obtain the fourth three-dimensional detection frame, and the covariance matrix is ​​updated to reflect the uncertainty of the state estimation; specifically:

[0156]

[0157] Among them, K k represents the Kalman gain, represents the predicted value, H k represents the transformation matrix, Indicates H k The transpose of R k represents the covariance matrix of the measurement noise, represents the final estimated state at time k (i.e., the fourth 3D detection box), Represents the estimated covariance matrix of the final error at time k, z k represents the measured value, and I represents the unit matrix.

[0158]

[0159] Among them, K k represents the Kalman gain, H k represents the transformation matrix, represents the estimated covariance matrix of the error at time k, and I represents the unit matrix. The final result is is the updated covariance matrix, which describes the error and uncertainty of the state estimate at the current moment. The diagonal elements represent the uncertainty (variance) of each state variable, such as position, velocity, etc. The off-diagonal elements represent the covariance between the state variables, that is, the correlation between them. ). Wherein, the state variables are T = {x, y, z, θ, h, w, l, v x , v y , v z}. (center point coordinates (x, y, z), length l, width w, height h, heading angle θ, v x ,v y ,v z is the speed of the target along the x, y, z coordinate system).

[0160] Next, the IoU3D between the first three-dimensional detection frame and the fourth three-dimensional detection frame can be calculated, a cost matrix can be established, and the cost matrix can be processed using the Hungarian algorithm to obtain the matching results of objects between different frames, thereby obtaining the target tracking sequence of the target vehicle.

[0161] Among them, IoU3D is an indicator used to evaluate the accuracy of detection results in 3D object detection, which is used to compare the degree of overlap between the 3D detection box and the 3D real box in 3D space. The content of each row and column of the cost matrix is ​​the similarity between the measured quantity and the predicted quantity, which is used to solve the data association problem, that is, how to pair the objects in adjacent frames. The IoU3D calculation formula is:

[0162]

[0163] Among them, v i is the volume of the 3D detection box of the i-th detected target, v j is the volume of the 3D detection box of the j-th predicted target, v ij is the intersection volume between the 3D detection box of the i-th detected target and the 3D detection box of the j-th predicted target. If there is a detected target and a predicted target, each with a corresponding 3D box, then the two boxes may overlap partially or completely. The intersection volume is the volume of the overlapping part of the two boxes, which can be used to evaluate the accuracy of the prediction, that is, the degree of overlap between the predicted box and the box of the actual detected target.

[0164] When the Hungarian algorithm is used to process the cost matrix to obtain the matching results of objects between different frames and obtain the target tracking sequence of the target vehicle, it can be specifically performed in the following manner:

[0165] 1. For the cost matrix, first find the minimum value of each row, then subtract the minimum value of the corresponding row from each element, then find the minimum value of each column, subtract the minimum value of the corresponding column from each element, so that each row or column of the matrix has at least one zero value, thus obtaining a new cost matrix.

[0166] 2. In the new cost matrix, horizontal and vertical lines are drawn to cover as many zero elements as possible using the least number of horizontal and vertical lines.

[0167] 3. If the zero elements have been completely covered, and the number of lines covering all zeros is equal to the number of rows or columns of the matrix, the matrix is ​​considered to have been successfully solved. At this point, the optimal allocation scheme can be obtained based on the position of the zero elements.

[0168] 4. If the zero element cannot be completely covered, find the smallest element in the matrix that is not covered. Subtract this element from all elements not covered by the line, and add this element to the data at all line intersections.

[0169] 5. Repeat the above steps 2-4.

[0170] The target vehicle is tracked using the 3D Kalman filter algorithm and the Hungarian algorithm. The target vehicle will have an identifier corresponding to it. The identifiers of the target vehicle in different frames are recorded together to obtain the target tracking sequence of the target vehicle.

[0171] Step 203: crop the target tracking sequence to obtain a target tail sequence of the target vehicle.

[0172] In some feasible embodiments, after obtaining the target tracking sequence of the target vehicle, the target tracking sequence may be cropped; specifically, the area of ​​the taillight of the target vehicle may be cropped to obtain the target tail sequence of the target vehicle.

[0173] Exemplarily, the heading angle of the target vehicle can be determined based on the target three-dimensional detection frame; when the heading angle indicates that the target vehicle is traveling in the same direction as the vehicle, a sliding window with a step size of 1 can be used to crop the target tracking sequence to obtain a target tail sequence.

[0174] Step 204: input the target tail sequence into a preset light language detection network to obtain the target light language of the taillight of the target vehicle.

[0175] like Figure 2c As shown, after obtaining the target tail sequence, the target tail sequence can be input into a preset light language detection network; the preset light language detection network can identify the light language of the target vehicle taillights based on the target tail sequence and output the target light language. The light language detection network can be a pre-trained three-dimensional convolutional neural network, and the light language of the target vehicle taillights can be straight ahead, brake, turn left, turn right, double flash, etc., which is not limited in the embodiment of the present invention.

[0176] Step 205: Obtain target lane line information of the lane in which the target vehicle is currently located from the target perception data.

[0177] After obtaining the target perception data, the vehicle can also obtain the target lane line information of the lane where the target vehicle is currently located from the target perception data.

[0178] Step 206: Determine a target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located based on the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information.

[0179] After obtaining the target three-dimensional detection frame, target lane line information, and target light language, the target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located can be determined based on the center point coordinates of the target vehicle in the target three-dimensional detection frame, the length, width, and height of the target vehicle, and the target lane line information.

[0180] Step 207: Determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs.

[0181] Then, the target angle interval to which the heading angle of the target vehicle belongs can be determined, and the target distance interval to which the target distance belongs can be determined. Exemplarily, multiple angle intervals and multiple distance intervals can be pre-set; then, the target angle interval to which the heading angle of the target vehicle belongs is determined from the multiple angle intervals, and the target distance interval to which the target distance belongs is determined from the multiple distance intervals.

[0182] Step 208: Determine the driving intention of the target vehicle according to the target lane line information, the target angle interval, the target distance interval, and the target light language.

[0183] After determining the target angle interval and the target distance interval, the driving intention of the target vehicle can be determined based on the target lane line information, the target angle interval, the target distance interval, and the target light language of the taillights of the target vehicle.

[0184] For example, within the continuous frames that meet the requirements, the target light language preliminarily indicates the driving intention of the target vehicle, and the heading angle and the distance between the target vehicle and the lane line of the current lane as comprehensive judgment conditions can improve the accuracy of the judgment.

[0185] Specifically, if the turn signal or the turn signal and the brake light are on at the same time, the target angle interval is the first preset angle range, and the target distance interval is the first preset distance range, it can be determined that the state of the target vehicle is turning or changing lanes.

[0186] If the brake light is on, the target angle interval is the second preset angle range, and the target distance interval is the second preset distance range, it can be determined that the state of the target vehicle is braking.

[0187] If the double flashes or the double flashes and brake lights are on at the same time, the target angle interval is the fourth preset angle range, and the target distance interval is the fourth preset distance range, it can be determined that the target vehicle is in an emergency.

[0188] Among them, the first preset angle range, the second preset angle range, the fourth preset angle range, the first preset distance range, the second preset distance range, and the fourth preset distance range can be set according to actual conditions; the first preset angle range, the second preset angle range, the fourth preset angle range, the first preset distance range, the second preset distance range, and the fourth preset distance range can be related to the target lane line information, and the embodiments of the present invention do not limit this.

[0189] For example, Figure 2d As shown, in the embodiment of the present invention, a target three-dimensional detection network can be designed first to obtain a target three-dimensional detection frame of a target vehicle.

[0190] After obtaining the target 3D detection frame according to the target 3D detection network, the target 3D detection frame can be tracked to obtain a target tracking sequence of the target vehicle. After obtaining the target tracking sequence, the target tracking sequence can be cropped to obtain a target tail sequence of the target vehicle.

[0191] Next, the light language detection network can be used to extract features from the target tail sequence to obtain the target light language of the taillights of the target vehicle.

[0192] After extracting the target lane line information, the target vehicle's driving intention can be identified by combining the target three-dimensional detection frame, target light language and target lane line information.

[0193] In an embodiment of the present invention, target perception data collected for a target vehicle is obtained, and the target perception data is input into a preset target three-dimensional detection network to obtain a target three-dimensional detection frame of the target vehicle; a target tracking sequence of the target vehicle is generated according to the target three-dimensional detection frame; the target tracking sequence is clipped to obtain a target tail sequence of the target vehicle; the target tail sequence is input into a preset light language detection network to obtain the target light language of the taillight of the target vehicle; the target lane line information of the lane where the target vehicle is currently located is obtained from the target perception data; the target distance between the target vehicle and the lane line of the lane where the target vehicle is currently located is determined according to the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information; the target angle interval to which the heading angle of the target vehicle belongs is determined, and the target distance interval to which the target distance belongs is determined; the driving intention of the target vehicle is determined according to the target lane line information, the target angle interval, the target distance interval, and the target light language. Through the embodiment of the present invention, different influencing factors can be combined to identify the driving intention of the target vehicle, thereby effectively identifying the true driving intention of the target vehicle.

[0194] Reference Figure 3 , shows a flowchart of another method for identifying vehicle intention according to an embodiment of the present invention, which may include the following steps:

[0195] Step 301: Acquire target perception data collected for a target vehicle, and input the target perception data into a preset target three-dimensional detection network to obtain a target three-dimensional detection frame of the target vehicle.

[0196] In actual applications, in order to ensure the safety of autonomous driving, the vehicle that is driving automatically needs to identify the driving intentions of surrounding target vehicles during the process; then, the autonomous driving decision is generated based on the driving intentions of the target vehicles; in order to identify the driving intentions of the target vehicles, the vehicle can first obtain target perception data collected for the target vehicle from a laser radar or an image acquisition device for target perception data.

[0197] After obtaining the target perception data collected for the target vehicle, the vehicle can input the target perception data into a preset target three-dimensional detection network; the target three-dimensional detection network can output a target three-dimensional detection frame of the target vehicle based on the target perception data.

[0198] Step 302: Generate a target tracking sequence for the target vehicle according to the target three-dimensional detection frame.

[0199] In some feasible embodiments, the target 3D detection frame may include multiple 3D detection frames, which may be 3D detection frames generated for the target vehicle at different times. After obtaining the target 3D detection frame, a target tracking sequence for the target vehicle may be generated based on the multiple 3D detection frames.

[0200] Step 303: crop the target tracking sequence to obtain a target tail sequence of the target vehicle.

[0201] After obtaining the target tracking sequence of the target vehicle, the target tracking sequence may be cropped; specifically, the area of ​​the taillight of the target vehicle may be cropped to obtain the target tail sequence of the target vehicle.

[0202] Step 304: input the target tail sequence into a preset light language detection network to obtain the target light language of the taillights of the target vehicle.

[0203] After obtaining the target tail sequence, the target tail sequence can be input into a preset light language detection network; the preset light language detection network can recognize the light language of the target vehicle taillights based on the target tail sequence and output the target light language.

[0204] Step 305: Obtain target lane line information of the lane in which the target vehicle is currently located from the target perception data.

[0205] After obtaining the target perception data, the vehicle can also obtain the target lane line information of the lane where the target vehicle is currently located from the target perception data.

[0206] Step 306: When the target light signal is empty, determine the target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located according to the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information.

[0207] In some feasible embodiments, if the target vehicle is traveling in the opposite direction to the host vehicle, the target light language of the target vehicle's taillights is not available to the host vehicle; in this case, the target light language is empty. In view of this situation, the embodiment of the present invention can identify the driving intention of the target vehicle only based on the target three-dimensional detection frame and the target lane line information.

[0208] Specifically, the target distance between the target vehicle and the lane line of the lane where the target vehicle is currently located can be determined based on the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information.

[0209] Step 307: Determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs.

[0210] Then, the target angle interval to which the heading angle of the target vehicle belongs can be determined, and the target distance interval to which the target distance belongs can be determined. Exemplarily, multiple angle intervals and multiple distance intervals can be pre-set; then, the target angle interval to which the heading angle of the target vehicle belongs is determined from the multiple angle intervals, and the target distance interval to which the target distance belongs is determined from the multiple distance intervals.

[0211] Step 308: Determine the driving intention of the target vehicle according to the target lane line information, the target angle interval, and the target distance interval.

[0212] After the target angle interval and the target distance interval are determined, the driving intention of the target vehicle can be determined according to the target lane line information, the target angle interval, and the target distance interval.

[0213] For example, the target light language of the target vehicle's taillights can provide important clues about the target vehicle's driving intentions. For example, if the brake lights are on, it can mean that the target vehicle is slowing down or stopping; if the turn signal is on, it can mean that the target vehicle intends to or is turning.

[0214] The target lane line information can provide information about the target vehicle's current lane and driving direction. Based on the target lane line information, the position and direction of the lane line of the target vehicle's current lane can be determined; further, based on the target angle interval and target distance interval, it can be determined whether the vehicle remains in the current lane or intends to switch to another lane.

[0215] Exemplarily, if the target angle interval is a third preset angle range and the target distance interval is a third preset distance range, it can be determined that the state of the target vehicle is normal driving. The specific values ​​of the third preset angle range and the third preset distance range can be set according to actual conditions. The third preset angle range and the third preset distance range can be related to the target lane line information.

[0216] For example, when performing intent recognition, three-dimensional target detection can be performed first. In the feature extraction module of the target three-dimensional detection network, full-dimensional dynamic convolution and adaptive pooling layers are integrated to improve the flexibility and adaptability of the model. The weight of the convolution kernel is allowed to be adjusted according to the dynamic characteristics of the input data. The model can adaptively optimize the convolution kernel according to the different attributes of the input, thereby significantly enhancing the feature extraction capability. In the target three-dimensional detection task, the application of full-dimensional dynamic convolution can effectively improve the adaptability and expressiveness of the model to the input data, thereby improving the ability to understand complex scenes. Compared with the traditional fixed convolution kernel method, ODConv reduces the number of model parameters, reduces the risk of overfitting, and optimizes the computational cost while maintaining performance. The dynamically generated convolution kernel can capture the characteristics of the input data more accurately, so that the model can show better generalization ability on unseen data. In addition, ODConv can be flexibly integrated into the existing convolutional neural network architecture, providing greater scalability and adaptability for model design. The size of the pooling area is automatically adjusted to ensure that the output feature map has a pre-set fixed size. This feature is particularly important when processing input feature maps of different sizes, as it avoids the complexity of adjusting the network structure due to inconsistent input sizes. Compared with the traditional fixed-size pooling layer, the adaptive pooling layer automatically adjusts the pooling area, evenly distributes the pooling operation, and reduces information loss. In the design of the target 3D detection network, the combination of the adaptive pooling layer can effectively process input data of different sizes without sacrificing information.

[0217] The DLA34 layer network is built on the ResNet34 framework with a multi-level fusion structure. It is divided into two modules: iterative deep aggregation and hierarchical deep aggregation. The iterative deep aggregation module aggregates features from the shallowest level and introduces connections and jump connections between different levels and iterative stages to make full use of multi-level information. This method ensures that shallow features can be effectively transmitted to the deep part of the network, thereby improving the overall feature representation capability. The hierarchical deep aggregation module effectively fuses features from each level by establishing connections between different levels, combining the advantages of multi-scale information extraction. The 3×3 convolution in the DLA34 network is replaced by ODConv, and the pooling layer is replaced by AdaPool. Specifically, ODConv learns different attentions on the four dimensions of the convolution kernel, including the input channel, output channel, spatial dimension, and number of convolution kernels, through a multi-dimensional attention mechanism and parallel strategy. And multiply the convolution kernel in the order of the number of convolution kernels, input channels, output channels and spatial dimensions. In the implementation of AdaPool, this layer can be divided into two parts: efficient dense semantic correspondence pooling (eDSCPool) and efficient multi-scale pooling (eMPool). eDSCPool is based on the Dyss similarity coefficient index to learn the image features of target perception data for a specific area, while eMPool uses a natural index method to ensure that larger activation values ​​have a significant impact on the final output and assign the lowest weight to smaller activation values. In addition, the regression parameters of the target 3D detection network include the center point position, depth, uncertainty, length, width, height and angle of the target vehicle of the 3D detection box.

[0218] The auxiliary learning parameters include important information such as the heat map, offset and intersection-over-union ratio of the two-dimensional projection points of the corner points and center points of the three-dimensional detection box. The design of these regression parameters and auxiliary learning parameters is to improve the accuracy and robustness of the target three-dimensional detection network. During the training process, the network uses multiple loss functions such as the focal loss function weighted by Gaussian kernel, the Laplace arbitrary uncertainty loss function and the cross entropy loss function to optimize multiple parameters. In the specific implementation, the target three-dimensional detection network is first divided into training and test sets through the Kitti dataset, and data preprocessing is performed, including standardization of image and point cloud data, data enhancement and format conversion. Through the back-propagation algorithm, the network continuously updates the model parameters to minimize the loss, and uses the test set to evaluate the model performance after each round of training, so as to select the model with the lowest loss on the test set as the final target three-dimensional detection network. This series of refined designs and training strategies effectively improves the accuracy and reliability of target three-dimensional detection.

[0219] For taillight detection, the spatial position information and geometric features (such as length, width, height and orientation angle) of the target vehicle provided by the 3D detection network can be used as the basis for the initial prediction of the taillight position. The approximate position of the taillight in three-dimensional space can be derived by using the position and direction information of the vehicle center point in the 3D detection results. Specifically, the possible area of ​​the taillight can be predicted based on the structural model of the vehicle and the relative position relationship between the taillight and the center point of the vehicle. This preliminary positioning based on the geometric model can significantly reduce the search space of the subsequent detection algorithm and improve the accuracy and efficiency of taillight detection. In addition, the depth information obtained by 3D detection can further optimize the detection process of the taillight, eliminate interference with the background through depth constraints, and improve the model's recognition ability of the taillight.

[0220] In terms of lane detection, environmental information (such as road structure, lane width and position) obtained by the three-dimensional detection network is also of great value. Lanes usually have certain geometric features, and their projection in three-dimensional space is closely related to the vehicle's motion trajectory. By combining the movement direction and position information of the target vehicle obtained by three-dimensional detection, direction constraints are introduced into the lane detection algorithm, thereby limiting the search range of the detection results and improving the detection efficiency. Specifically, the movement direction and position of the vehicle are first obtained through three-dimensional detection, and then the projection model is used to deduce the possible position of the lane, and the lane is further extracted by combining edge detection and other methods. In addition, the introduction of depth information also helps to identify the boundaries of lanes and reduce noise interference.

[0221] In terms of target tracking, the target vehicle is continuously identified and tracked by comprehensively analyzing the position and state information of the target vehicle in multiple consecutive frames to record its motion trajectory. First, the target three-dimensional detection frame can be composed of multiple three-dimensional detection frames generated at different time points, which provide spatial position and state information about the target vehicle at different times. After obtaining the target three-dimensional detection frame, the system will use this information to generate a tracking sequence for the target vehicle. Specifically, the first three-dimensional detection frame at the current moment is combined with the second three-dimensional detection frame at the previous moment, and the system will predict the third three-dimensional detection frame at the next moment based on the second three-dimensional detection frame. This prediction can be based on the Kalman filter algorithm, and the state of the target vehicle at the next moment can be obtained by approximating the inter-frame displacement through a constant velocity model. At this time, there may be a certain error between the predicted value and the measured value, so it is necessary to fuse the first three-dimensional detection frame and the third three-dimensional detection frame to obtain a more accurate fourth three-dimensional detection frame, thereby updating the tracking sequence of the target vehicle. In particular, when the object in the fourth three-dimensional detection frame matches the object in the first three-dimensional detection frame, the tracking sequence of the target vehicle can be updated according to the first three-dimensional detection frame.

[0222] In this process, the Kalman gain is used to weight the measured quantity and the predicted quantity to form a fourth three-dimensional detection frame, and the covariance matrix is ​​updated to reflect the uncertainty of the state estimation. Subsequently, the IoU3D (intersection over union) between the first three-dimensional detection frame and the fourth three-dimensional detection frame is calculated to evaluate the accuracy of the detection result. By establishing a cost matrix and processing the matrix using the Hungarian algorithm, the system can obtain the matching results of the targets between different frames, thereby completing the generation of the tracking sequence of the target vehicle. After obtaining the tracking sequence of the target vehicle, the system can crop it to extract the tail sequence of the target vehicle, with special attention to the cropping of the taillight area. At the same time, the heading angle of the target vehicle is determined based on the target three-dimensional detection frame. When the heading angle indicates that the target vehicle is traveling in the same direction as the vehicle, the sliding window can be used to crop it in the target tracking sequence. Then, the obtained target tail sequence is input into the preset light language detection network to identify the light language of the target vehicle's taillights, including signals such as straight, brake, turn left, turn right, and double flash.

[0223] A method for identifying the target vehicle's driving intention by comprehensively analyzing its heading angle and target distance information. First, the system divides the target vehicle's heading angle into multiple preset target angle intervals, and divides the target distance into multiple target distance intervals, so as to accurately judge the target vehicle's behavior. In the specific implementation, the system will predefine multiple angle intervals and distance intervals, and determine the specific interval to which it belongs based on the target vehicle's real-time heading angle and distance information. Next, combined with the target lane line information, target angle interval, target distance interval, and target light language, the system can infer the target vehicle's driving intention. For example, within the continuous frames that meet the requirements, the state of the target light language preliminarily indicates the target vehicle's driving intention, and the heading angle and the distance between the target vehicle and the current lane line serve as important comprehensive judgment conditions to improve the accuracy of intention judgment. If the turn signal or the turn signal and brake light are on at the same time, and the target angle interval and the target distance interval are both within the preset range, the system determines that the target vehicle is turning or changing lanes; if only the brake light is on and meets the second preset condition, the system determines that the vehicle is braking; and if the hazard lights or the hazard lights and brake lights are on at the same time, and the target angle and distance are within the fourth preset range, the system determines that the vehicle is in an emergency.

[0224] Finally, the present invention obtains the three-dimensional detection frame of the target vehicle based on the target three-dimensional detection network, and then tracks the frame to generate a tracking sequence of the target vehicle. After the tracking sequence is clipped, the tail sequence of the target vehicle will be obtained. The target tail sequence is feature extracted through the light language detection network to obtain the target light language information of the taillight. At the same time, the lane line information of the current lane is extracted from the target perception data, and the target three-dimensional detection frame, target light language and target lane line information are combined to further identify the driving intention of the target vehicle.

[0225] In summary, the present invention provides an effective method to identify the real driving intention of the target vehicle by comprehensively considering the three-dimensional detection frame, light language, heading angle and lane line information of the target vehicle. This method not only improves the accuracy of driving intention judgment, but also provides strong technical support for the development of intelligent transportation systems.

[0226] As shown in Table 1, the judgment conditions, judgment results, weight influence, verification strategy, data support indicators and possible influencing factors corresponding to some driving behaviors in the embodiments of the present invention are shown:

[0227] Table 1:

[0228]

[0229]

[0230] In an embodiment of the present invention, target perception data collected for a target vehicle is obtained, and the target perception data is input into a preset target three-dimensional detection network to obtain a target three-dimensional detection frame of the target vehicle; a target tracking sequence of the target vehicle is generated according to the target three-dimensional detection frame; the target tracking sequence is clipped to obtain a target tail sequence of the target vehicle; the target tail sequence is input into a preset light language detection network to obtain the target light language of the taillight of the target vehicle; the target lane line information of the lane where the target vehicle is currently located is obtained from the target perception data; when the target light language is empty, the target distance between the target vehicle and the lane line of the lane where the target vehicle is currently located is determined according to the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information; the target angle interval to which the heading angle of the target vehicle belongs is determined, and the target distance interval to which the target distance belongs is determined; the driving intention of the target vehicle is determined according to the target lane line information, the target angle interval, and the target distance interval. For the scene where the target vehicle is partially blocked or truncated, resulting in the inability of the vehicle to collect the information of the taillight of the target vehicle, the embodiment of the present invention can still effectively identify the driving intention of the target vehicle.

[0231] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0232] Reference Figure 4 , shows a schematic diagram of the structure of a vehicle intention recognition device according to an embodiment of the present invention, which may include the following modules:

[0233] The three-dimensional detection frame extraction module 401 is used to obtain target perception data collected for the target vehicle and extract the target three-dimensional detection frame of the target vehicle from the target perception data;

[0234] A light language determination module 402 is used to determine a target light language of the taillight of the target vehicle according to the target three-dimensional detection frame;

[0235] The lane line acquisition module 403 is used to acquire the target lane line information of the lane where the target vehicle is currently located from the target perception data;

[0236] The intention recognition module 404 is used to determine the driving intention of the target vehicle based on the target three-dimensional detection frame, target light language, and target lane line information.

[0237] In one embodiment of the present invention, a three-dimensional detection frame extraction module is used to input target perception data into a preset target three-dimensional detection network to obtain a target three-dimensional detection frame of a target vehicle.

[0238] In one embodiment of the present invention, the feature extraction module of the target three-dimensional detection network includes: full-dimensional dynamic convolution and adaptive pooling layer.

[0239] In one embodiment of the present invention, the regression parameters of the target three-dimensional detection network include:

[0240] A first heat map of the projection position of the center point of the three-dimensional detection frame on the image of the target perception data, a first offset, depth, uncertainty, length, width, height and angle of the target vehicle projected on the image of the target perception data by the center point of the three-dimensional detection frame;

[0241] The auxiliary learning parameters of the target 3D detection network include:

[0242] A second heat map of the two-dimensional projection points of the corner points and the center point of the three-dimensional detection frame, a second offset of the two-dimensional projection points of the corner points and the center point of the three-dimensional detection frame to the center point of the two-dimensional frame, an intersection-and-union ratio, and a sine and cosine representation of the rotation angle of the three-dimensional detection frame;

[0243] Among them, the target three-dimensional detection network uses a focal loss function based on Gaussian kernel weighting to calculate the loss of the first heat map and the second heat map;

[0244] The target 3D detection network uses the Laplace arbitrary uncertainty loss function to calculate the loss of depth and uncertainty;

[0245] The target 3D detection network uses an absolute value loss function to calculate the loss of the length, width, height, first offset, and second offset of the target vehicle;

[0246] The target 3D detection network uses the cross entropy loss function to calculate the loss of the target vehicle’s angle;

[0247] The target 3D detection network uses the intersection-over-union loss function to calculate the loss of the intersection-over-union ratio;

[0248] The target 3D detection network uses a smoothed absolute value loss function to calculate the loss represented by the sine and cosine of the rotation angle of the 3D detection box.

[0249] In one embodiment of the present invention, the target three-dimensional detection frame includes at least one of the following information:

[0250] The coordinates of the center point of the target vehicle, the length, width and height of the target vehicle, and the heading angle of the target vehicle.

[0251] In one embodiment of the present invention, an intention recognition module is used to determine the target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located based on the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information; determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs; determine the driving intention of the target vehicle based on the target lane line information, the target angle interval, the target distance interval, and the target light language.

[0252] In one embodiment of the present invention, the intention recognition module is also used to determine the target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located according to the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information when the target light language is empty; determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs; determine the driving intention of the target vehicle according to the target lane line information, the target angle interval, and the target distance interval.

[0253] In one embodiment of the present invention, a light language determination module is used to generate a target tracking sequence of a target vehicle based on a target three-dimensional detection frame; crop the target tracking sequence to obtain a target tail sequence of the target vehicle; and input the target tail sequence into a preset light language detection network to obtain the target light language of the taillights of the target vehicle.

[0254] In one embodiment of the present invention, the target three-dimensional detection frame includes at least a first three-dimensional detection frame at a current moment and a second three-dimensional detection frame at a previous moment, and the light language determination module is used to predict a third three-dimensional detection frame of the second three-dimensional detection frame at the next moment based on the second three-dimensional detection frame; fuse the first three-dimensional detection frame and the third three-dimensional detection frame to obtain a fourth three-dimensional detection frame; when an object in the fourth three-dimensional detection frame matches an object in the first three-dimensional detection frame, update the tracking sequence of the target vehicle according to the first three-dimensional detection frame to obtain a target tracking sequence.

[0255] In an embodiment of the present invention, target perception data collected for a target vehicle is obtained, and a target three-dimensional detection frame of the target vehicle is extracted from the target perception data; the target light language of the taillights of the target vehicle is determined based on the target three-dimensional detection frame; the target lane line information of the lane in which the target vehicle is currently located is obtained from the target perception data; the driving intention of the target vehicle is determined based on the target three-dimensional detection frame, the target light language, and the target lane line information. Through the embodiment of the present invention, different influencing factors can be combined to identify the driving intention of the target vehicle, thereby effectively identifying the true driving intention of the target vehicle. In addition, for scenarios where the target vehicle is partially blocked or truncated, resulting in the inability of the vehicle to collect information about the taillights of the target vehicle, the embodiment of the present invention can still effectively identify the driving intention of the target vehicle.

[0256] An embodiment of the present invention further provides a vehicle, which may include a processor, a memory, and a computer program stored in the memory and capable of running on the processor, and when the computer program is executed by the processor, the above-mentioned vehicle intention recognition method is implemented.

[0257] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method for identifying vehicle intention as described above is implemented.

[0258] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0259] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0260] Those skilled in the art will appreciate that the embodiments of the present invention may be provided as methods, devices, or computer program products. Therefore, the embodiments of the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the embodiments of the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0261] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0262] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0263] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0264] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0265] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0266] The above provides a detailed introduction to a method for identifying vehicle intention, a device for identifying vehicle intention, a vehicle and a computer-readable storage medium. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for a general technician in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for identifying vehicle intention, characterized in that: The method comprises: Acquire target perception data collected for a target vehicle, and extract a target three-dimensional detection frame of the target vehicle from the target perception data; Determining a target light language of a taillight of the target vehicle according to the target three-dimensional detection frame; Acquiring target lane line information of the lane where the target vehicle is currently located from the target perception data; The driving intention of the target vehicle is determined according to the target three-dimensional detection frame, the target light language, and the target lane line information.

2. The method according to claim 1, characterized in that The step of extracting a target three-dimensional detection frame of the target vehicle from the target perception data includes: The target perception data is input into a preset target three-dimensional detection network to obtain a target three-dimensional detection frame of the target vehicle.

3. The method according to claim 2, characterized in that The feature extraction module of the target three-dimensional detection network includes: Full-dimensional dynamic convolution and adaptive pooling layers.

4. The method according to claim 2, characterized in that: The regression parameters of the target three-dimensional detection network include: A first heat map of the projection position of the center point of the three-dimensional detection frame on the image of the target perception data, a first offset, depth, uncertainty, length, width, height and angle of the target vehicle projected on the image of the target perception data by the center point of the three-dimensional detection frame; The auxiliary learning parameters of the target three-dimensional detection network include: A second heat map of the two-dimensional projection points of the corner points and the center point of the three-dimensional detection frame, a second offset of the two-dimensional projection points of the corner points and the center point of the three-dimensional detection frame to the center point of the two-dimensional frame, an intersection-and-union ratio, and a sine and cosine representation of the rotation angle of the three-dimensional detection frame; The target three-dimensional detection network uses a focal loss function based on Gaussian kernel weighting to calculate the loss of the first heat map and the second heat map; The target three-dimensional detection network uses the Laplace arbitrary uncertainty loss function to calculate the loss of depth and uncertainty; The target three-dimensional detection network uses an absolute value loss function to calculate the loss of the length, width, height, first offset and second offset of the target vehicle; The target three-dimensional detection network calculates the loss of the angle of the target vehicle using a cross entropy loss function; The target three-dimensional detection network uses an intersection-over-union loss function to calculate the loss of the intersection-over-union ratio; The target 3D detection network uses a smooth absolute value loss function to calculate the loss represented by the sine and cosine of the 3D detection box rotation angle.

5. The method according to claim 1, characterized in that The target 3D detection frame includes at least one of the following information: The center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the heading angle of the target vehicle.

6. The method according to claim 5, characterized in that The determining the driving intention of the target vehicle according to the target three-dimensional detection frame, the target light language, and the target lane line information includes: Determine a target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located according to the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information; Determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs; The driving intention of the target vehicle is determined according to the target lane line information, the target angle interval, the target distance interval, and the target light language.

7. The method according to claim 5, characterized in that When the target light language is empty, determining the driving intention of the target vehicle according to the target three-dimensional detection frame, the target light language, and the target lane line information includes: Determine a target distance between the target vehicle and the lane line of the lane in which the target vehicle is currently located according to the center point coordinates of the target vehicle, the length, width and height of the target vehicle, and the target lane line information; Determine the target angle interval to which the heading angle of the target vehicle belongs, and determine the target distance interval to which the target distance belongs; The driving intention of the target vehicle is determined according to the target lane line information, the target angle interval, and the target distance interval.

8. The method according to claim 1, characterized in that: The step of determining the target light language of the taillight of the target vehicle according to the target three-dimensional detection frame includes: Generating a target tracking sequence of the target vehicle according to the target three-dimensional detection frame; Cutting the target tracking sequence to obtain a target tail sequence of the target vehicle; The target tail sequence is input into a preset light language detection network to obtain the target light language of the taillight of the target vehicle.

9. The method according to claim 8, characterized in that The target three-dimensional detection frame at least includes a first three-dimensional detection frame at a current moment and a second three-dimensional detection frame at a previous moment, and generating a target tracking sequence of the target vehicle according to the target three-dimensional detection frame includes: Predicting a third three-dimensional detection frame at a next moment of the second three-dimensional detection frame according to the second three-dimensional detection frame; Fusing the first three-dimensional detection frame and the third three-dimensional detection frame to obtain a fourth three-dimensional detection frame; When the object in the fourth three-dimensional detection frame matches the object in the first three-dimensional detection frame, the tracking sequence of the target vehicle is updated according to the first three-dimensional detection frame to obtain the target tracking sequence.

10. A vehicle intention recognition device, characterized in that: The device comprises: A three-dimensional detection frame extraction module is used to obtain target perception data collected for a target vehicle, and extract a target three-dimensional detection frame of the target vehicle from the target perception data; A light language determination module, used to determine the target light language of the taillight of the target vehicle according to the target three-dimensional detection frame; A lane line acquisition module, used to acquire target lane line information of the lane where the target vehicle is currently located from the target perception data; The intention recognition module is used to determine the driving intention of the target vehicle based on the target three-dimensional detection frame, the target light language, and the target lane line information.

11. A vehicle, characterized in that: The invention comprises a processor, a memory and a computer program stored in the memory and capable of running on the processor, wherein when the computer program is executed by the processor, the method for recognizing vehicle intention as claimed in any one of claims 1 to 9 is implemented.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for identifying vehicle intention as described in any one of claims 1 to 9 is implemented.

Images