A vehicle speed prediction method and system fusing vehicle end perception and V2X information, a device, and a storage medium

Abstract

The application relates to the technical field of intelligent networked vehicles and vehicle-road cooperation, and discloses a vehicle speed prediction method and system fusing vehicle-end perception and V2X information, equipment and a storage medium. The method comprises the following steps: acquiring vehicle-end perception data and V2X cooperation data; performing time alignment, abnormal value elimination and expired message filtering on the data; performing credibility evaluation on different data sources based on time freshness, stability and consistency; determining the fusion weight of vehicle-end perception features and V2X cooperation features according to a current traffic scene and the credibility, generating a fusion feature sequence; obtaining an initial speed prediction result based on the fusion feature sequence; and performing analytical expression correction on the initial speed prediction result according to road speed limit, safe vehicle distance, signal light passing window and vehicle dynamics boundary, so as to obtain a target speed prediction result. Compared with the prior art, the application can balance speed prediction accuracy, result executability and multi-source information utilization robustness in a complex traffic scene.

Description

Technical Field

[0001] This invention relates to the fields of intelligent connected vehicles, vehicle-road cooperation, and vehicle motion state prediction technology, and in particular to a method, system, device, and storage medium for predicting vehicle speed by integrating vehicle-side perception and V2X information. Background Technology

[0002] This section provides only background information relevant to this disclosure and is not necessarily prior art.

[0003] With the development of intelligent connected vehicles and vehicle-road cooperative technologies, vehicles no longer rely solely on cameras, millimeter-wave radar, lidar, wheel speedometers, and inertial sensors for local environmental perception. They can also obtain information such as the status of surrounding vehicles, traffic light status, map topology, lane connectivity, road events, and roadside guidance information through communication methods such as V2V, V2I, and V2N.

[0004] Existing vehicle speed prediction methods mostly use historical vehicle speed sequences or single-vehicle perception information as the main input. Although they can fit short-term speed change trends under simple working conditions, they do not make sufficient use of external collaborative information in scenarios such as signalized intersections, complex intersections, traffic congestion, and event guidance.

[0005] On the other hand, vehicle-side perception information and V2X information differ significantly in terms of timeliness, stability, and consistency. For example, radar provides high real-time observation of vehicles ahead, but is prone to target jitter in obstructed and cut-in / cut-out scenarios; V2I messages can provide traffic light status and map connectivity, but their reference value decreases significantly when messages are delayed, lost, or expired. Existing methods typically lack a unified modeling of these differences.

[0006] Furthermore, many speed prediction models directly output numerical sequences without a posterior correction mechanism that integrates with speed limits, safe distances, traffic light windows, and vehicle dynamics. As a result, although the prediction results may have small statistical errors, they may not be feasible in engineering applications. Summary of the Invention

[0007] The purpose of this invention is to provide a method, system, device and storage medium for vehicle speed prediction that integrates vehicle-side perception and V2X information, so as to solve the problems of crude multi-source heterogeneous data fusion mechanism, lack of traffic constraints and dynamic constraints in speed prediction and insufficient adaptability to complex traffic scenarios in the prior art.

[0008] Firstly, the solution provides a vehicle speed prediction method that integrates vehicle-side perception and V2X information, employing the following technical approach: acquiring vehicle-side perception data and V2X collaborative data; preprocessing the multi-source data and calculating the credibility of each data source from three dimensions: time freshness, stability, and consistency; adaptively weighting and fusing vehicle-side perception features and V2X collaborative features in conjunction with the traffic scenario; generating an initial speed prediction result based on historical fusion feature sequences; then correcting the initial speed prediction result by incorporating road speed limit constraints, safe following distance constraints, traffic light passage window constraints, and dynamic smoothing constraints to obtain the target speed prediction result; finally, updating the scene correction parameters and fusion weights online based on the error between the actual speed and the predicted speed.

[0009] The specific steps include:

[0010] Step 1: Collect vehicle-side perception data and V2X collaborative data and perform data preprocessing; calculate the comprehensive reliability of the preprocessed data from three dimensions: time freshness, stability, and consistency.

[0011] Step 2: Based on the current traffic scenario and the credibility obtained in Step 1, calculate the fusion weight of vehicle-side perception features and V2X collaborative features to generate fused features;

[0012] Step 3: Input the fused historical feature sequence into the velocity prediction model to obtain the initial velocity prediction value within the preset time window in the future;

[0013] Step 4: Correct the initial prediction values ​​based on speed limits, safe following distances, traffic light passage windows, and vehicle dynamics boundaries;

[0014] Step 5: Output the final target velocity prediction result, and update the scene correction parameters and fusion weights online according to the prediction error. The update of the fusion weights depends on the prediction error and the comprehensive confidence obtained in Step 1.

[0015] In some embodiments, step 1 includes the following steps:

[0016] Step 1-1: Collect vehicle-side perception data and V2X collaborative data respectively. The vehicle-side perception data includes, but is not limited to, vehicle speed, vehicle acceleration, distance to the vehicle in front, relative speed of the vehicle in front, vehicle position, heading angle, and vehicle control status. The V2X collaborative data includes, but is not limited to, traffic light phase status, phase remaining time, road speed limit, map topology information, road event information, roadside guidance information, and perception sharing information. The V2X collaborative data is obtained by receiving and parsing collaborative information from V2V, V2I, or roadside units. The perception sharing information may include information about surrounding vehicles, but its scope is broader than that of surrounding vehicles, and also includes target-level and event-level environmental perception results generated and shared by roadside perception devices or other networked traffic participants.

[0017] Step 1-2: Perform data preprocessing on the collected multi-source data, aiming to unify asynchronous, multi-frequency, and different quality data into analyzable input. The preprocessing specifically includes at least one of time alignment, missing value compensation, outlier removal, and expired message filtering, and forms vehicle-side perception feature vector and V2X collaborative feature vector respectively.

[0018] After data collection, it is uniformly converted into a computable vehicle-side perception state vector, specifically:

[0019] Define the vehicle-side perception state vector :

[0020]

[0021] In the formula: Indicates time The vehicle-side perception state vector; This indicates the vehicle's speed, expressed in m / s, with a range of values ​​of [missing value]. The corresponding range is... arrive km / h, which can cover the typical speed range of urban roads, elevated roads and highways; This represents the longitudinal acceleration of the vehicle, measured in m / s². 2 The range of values ​​is Negative values ​​indicate deceleration, while positive values ​​indicate acceleration. This range covers both normal driving and emergency braking scenarios. This represents the longitudinal distance between the vehicle and the vehicle in front at the current moment, in meters (m), and its value range is... When the distance exceeds 200m, the direct impact of the preceding vehicle on short-term speed prediction is usually significantly reduced. The relative speed with respect to the vehicle in front is defined as follows: The unit is m / s, and the recommended value range is [value range missing]. . The vehicle location can be represented by two-dimensional projected coordinates or high-precision map reference coordinates. There is no fixed range of values, but a unified coordinate system should be used throughout the system. This represents the vehicle's heading angle, measured in rad. A recommended value range is [range missing]. . This indicates vehicle control state characteristics, which may include throttle opening, brake pedal opening, gear position, and drive mode. After normalization, the range is... .

[0022] Define V2X cooperative state vector :

[0023]

[0024] In the formula: Indicates time V2X cooperative state vector; This indicates the phase state of the traffic light, using integer encoding, such as red, yellow, green, and unknown state. The set of possible values ​​is... ; This represents the remaining time of the current signal phase, in seconds, and its value range is... For common urban intersections, this range is sufficient to cover the vast majority of actual control phases; This displays the speed limit information for the current road segment or lane, in m / s, with a value range of [value missing]. ; Indicates the intensity of a road event or the level of an obstacle, normalized to an interval. If discrete levels are used, the following can be taken: , respectively representing no event, minor event, moderate event, and serious event. The map topology and lane connectivity features, typically derived from MAP messages or high-precision maps, can be encoded into vectors and then normalized. . This indicates roadside guidance information, which may include suggested speed, target lane, or coordinated traffic signs, and covers a range of... m / s;

[0025] Define the total state vector :

[0026]

[0027] In the formula, Indicates time The global input state. This formula places the vehicle-side perception state and the V2X collaborative state under the same analytical framework, serving as the unified input basis for subsequent credibility assessment, adaptive fusion, and speed prediction.

[0028] Steps 1-3: Calculate the overall reliability of the preprocessed data from three dimensions: time freshness, stability, and consistency.

[0029] In some embodiments, steps 1-3 include the following steps:

[0030] Step 1-3-1: Calculate the data's time freshness, stability factor, and consistency factor respectively:

[0031] The first is calculated using the following formula. Class data source at time Time freshness :

[0032]

[0033] in, For the first The difference between the current time and the most recent valid update time of the data source. Indicates the first The time decay coefficient of the data source has a range of values. ; The range of values ​​is When the message arrived, ,at this time As the news becomes outdated, Monotonically decrease; larger values ​​can be used for high-time-sensitivity data such as traffic light status and broadcast messages from vehicles ahead, while smaller values ​​can be used for low-change data such as static maps; This represents the difference between the current time and the most recent valid update time of the data source, expressed in seconds, and satisfies... .

[0034] The first is calculated using the following formula. Stability factor of data source :

[0035]

[0036] in, This indicates the normalized variance of the data source within the most recent time window, with values ​​falling within the range of... Within the range, to avoid outliers dominating the market excessively; The range of values ​​is Using rational functions The reasons are as follows: First, this form is always bounded and will not diverge; second, when the variance of fluctuation is small, the penalty for stable data is lighter, which is conducive to preserving effective information; third, when the variance of fluctuation increases, the weight will drop rapidly, which can suppress data sources with severe jitter.

[0037] The first is calculated using the following formula. Consistency factor of data source :

[0038]

[0039] in, This represents the consistency adjustment coefficient, with a value range of [value missing]. , Indicates the first The normalization deviation between the class data source and the reference source satisfies ; The range of values ​​is For example, it can be defined as the deviation between the vehicle-to-vehicle (V2V) speed detected and the speed broadcast by the vehicle, or the deviation between the roadside recommended speed and the locally perceived congestion state. The consistency factor means that if a data source severely conflicts with other reliable data sources, the reliability of that data source at the current moment should be reduced. The reason for using an exponential function is that it is more sensitive to the amplification of deviations and can quickly reduce the corresponding weights when multi-source conflicts are significant, thereby reducing the interference of erroneous collaborative information on prediction.

[0040] Step 1-3-2: Calculate the comprehensive credibility based on the obtained time freshness, stability factor, and consistency factor using weighted average. :

[0041]

[0042] in , , The weights for freshness, stability, and consistency are respectively, all satisfying the following conditions. And the sum of the three is 1.

[0043] The weights for time freshness, stability, and consistency can be quantitatively determined based on historical calibration data. First, according to traffic scenario labels... The historical samples are grouped, and the optimal weights for the corresponding scenario are obtained with the objective of minimizing the initial velocity prediction error, provided that all weights are non-negative and the sum of the weights is 1.

[0044]

[0045] The constraints are:

[0046]

[0047] In the formula, Indicates the first Number of calibration samples in traffic-like scenarios Indicates the first The actual speed corresponding to each sample This represents the initial velocity prediction value obtained after applying candidate confidence weights. The above optimization can be implemented using grid search, projective gradient descent, or other optimization algorithms with simplex constraints.

[0048] When sufficient calibration samples are lacking, the initial weights for time freshness, stability, and consistency can be set to 0.4, 0.3, and 0.3, respectively. The weights are then calibrated based on the scenario; for example, increasing the weight for time freshness in an intersection scenario and increasing the weight for stability in a low-speed congestion following scenario. Once the number of valid samples accumulated for a particular scenario reaches a preset threshold, the weights are recalibrated using historical data from that scenario. This ensures that the weight allocation has a clear quantitative basis and can adapt to the data quality characteristics of different traffic scenarios.

[0049] In some embodiments, step 2 includes the following steps:

[0050] Step 2-1: Calculate the vehicle perception weights based on the combined reliability of the collected vehicle-side perception data and V2X collaborative data. and V2X weights Specifically:

[0051]

[0052]

[0053] in This represents the overall reliability of the vehicle-side sensing sources. Represents the overall credibility of V2X message sources;

[0054] Obtain the fused integrated feature vector

[0055]

[0056] Step 2-2: Obtain fused features based on the weights obtained in Step 2-1 and the traffic scene labels. Explicitly incorporating scene labels into the fusion process significantly improves model interpretability and practical adaptability, specifically:

[0057]

[0058] To represent traffic scene labels, it is recommended to use a finite set. These correspond to scenarios such as free driving, following other vehicles, intersections, traffic congestion, and event guidance. This represents the normalized vehicle-side perception feature vector. This represents the normalized V2X collaborative feature vector. and These represent the scene-related vehicle perception weights and V2X weights, respectively, both located in... Within the range.

[0059] Before feature fusion, the vehicle-side perception features and V2X collaborative features are first mapped to a common feature space of the same dimension:

[0060]

[0061]

[0062] In the formula, , These represent the projection matrices of the vehicle-side perception features and the V2X collaborative features, respectively. , Let each represent the corresponding bias vector, such that... and They have the same dimensions.

[0063] For traffic scenarios Pre-set or calibrate the prior coefficients of the vehicle-side perception scenario using historical data. Prior coefficients for V2X scenarios The final fusion weights are calculated using the following formula:

[0064]

[0065]

[0066] In the formula, A positive number is used to prevent the denominator from being zero. The fusion feature is represented as:

[0067]

[0068] In an optional embodiment, the prior coefficients for each scenario may use the following initial values:

[0069]

[0070] The above values ​​are only initial calibration values ​​and do not constitute a limitation on the protection scope. In free driving and following scenarios, the vehicle's status, distance to the vehicle in front, and relative speed have a significant impact on short-term speed changes, therefore, the vehicle-side perception prior coefficient is increased; in intersection scenarios, traffic light phase, remaining time, and map lane connectivity are more critical, therefore, the V2X prior coefficient is increased; in event-guided scenarios, road events and roadside perception shared information have a strong impact on vehicle speed changes, therefore, the V2X prior coefficient is further increased.

[0071] In some embodiments, step 3 includes the following steps:

[0072] Step 3-1: Arrange the fusion features generated in Step 2 in chronological order to construct a historical fusion feature sequence; define the historical fusion feature sequence:

[0073]

[0074] In the formula: Indicates the end time Historical fusion feature sequence; Indicates the length of the history window, satisfying The range is If the system sampling frequency is 10Hz, then the preferred frequency can be... to The fixed-length sliding window corresponds to a historical window of 1 to 3 seconds. Vehicle speed prediction relies more on recent dynamics; an excessively long historical window introduces outdated information and redundant noise, while also increasing the computational burden. Conversely, an excessively short window cannot fully reflect the speed change trend and control intent. Therefore, the fixed-length sliding window achieves a better balance between information sufficiency and real-time performance.

[0075] Step 3-2: Input the historical fusion feature sequence into the pre-trained speed prediction model;

[0076] Step 3-3: Using the velocity prediction model, perform time-series recursion or direct mapping to output the initial velocity prediction values ​​for each moment within a preset future time window. This yields a multi-step prediction sequence:

[0077]

[0078] In the formula: This represents the initial prediction velocity in one step before analytical constraint correction, in m / s, and ranges from [value missing]. ; Indicates the future The initial predicted velocity sequence of the step. Indicates the prediction step size, satisfying , range .

[0079] In some embodiments, the velocity prediction model is constructed using a gated recurrent unit network. The model input is the value up to the current time. Historical fusion feature sequence:

[0080]

[0081] Among them, the historical window length It can be set to 10 to 30, corresponding to the most recent 1 to 3 seconds of historical information at a sampling frequency of 10Hz; the fusion features at each moment. All are located in the same Dimensional public feature space, It can be set to 16 to 128, therefore the dimension of the model input tensor is... .

[0082] The model consists of an input normalization layer, at least one GRU temporal coding layer, and a fully connected output layer. The GRU temporal coding layer recursively processes the historical fused features to obtain the latent state at the final time step. In an optional embodiment, the time series prediction function can be specifically written as:

[0083]

[0084]

[0085] In the formula: Indicates time The hidden state vector can be set as follows: , among which dimension The range is . This represents the hidden state update function, which can be implemented by an LSTM unit, a GRU unit, or other gated structures. and These represent the output layer weight matrix and bias vector, respectively.

[0086] In a preferred embodiment, the GRU network is configured with two layers, with the hidden state dimensions of the first and second layers being 64 and 32, respectively, and a random deactivation layer with a drop-off ratio of 0.1 to 0.3 can be set between the two layers.

[0087] For one-step velocity prediction, the fully connected output layer contains one output node and generates the initial predicted velocity for the next time step based on the hidden state:

[0088]

[0089] For multi-step velocity prediction, the fully connected output layer includes Each output node directly generates the future. Initial velocity prediction sequence at each time step:

[0090]

[0091] in, It can be set from 1 to 30. The Softplus function is used to ensure that the model output is non-negative. The model output is the initial predicted speed before adjustments for road speed limits, safe following distances, traffic light windows, and vehicle dynamics boundaries.

[0092] The model training data consists of historical vehicle operation data with completion time alignment and outlier handling. A length of [length missing] is used. The sliding window is used to construct training samples, and the previous The fused features at each time step are used as model input, and then... The measured vehicle speed at each moment is used as the supervision label. The training data can be divided into training set, validation set and test set in chronological order to avoid data from the same continuous journey appearing in both the training set and the test set.

[0093] During training, continuous features are standardized using the mean and standard deviation obtained from the training set statistics, while discrete features such as signal phase and scene labels are processed using cross-entropy encoding or one-hot encoding. The model uses a joint loss function. Conduct training:

[0094]

[0095] in, This represents the prediction error term; This represents the first-order difference smoothing term of the predicted sequence, used to suppress high-frequency jitter; This represents a consistency constraint term used to penalize over-rate predictions during the training phase, specifically:

[0096]

[0097] in, Represents the number of training samples, satisfying ; Indicates the first The true speed of each sample; Indicates the first Prediction speed per sample;

[0098]

[0099]

[0100] in, Indicates the first The speed limit value corresponding to each sample.

[0101] Model parameters Updates are performed via error backpropagation. The optimizer can be the Adam optimizer, and the initial learning rate can be set to... to The batch size can be set from 32 to 256, and the number of training rounds can be determined by an early stopping strategy based on the validation set loss.

[0102] After the model completes offline training and reaches the preset error requirement, the model parameters are fixed. The model is deployed to in-vehicle or edge computing devices to perform real-time inference. During the online operation phase of the vehicle, the main updates are to the scene correction parameters and multi-source fusion weights. It does not require retraining the entire time-series prediction model in each sampling period, thus ensuring real-time computing performance and model stability.

[0103] In some embodiments, step 4 includes the following steps:

[0104] Step 4-1: Calculate the upper bound of the safe speed based on the safe time interval. and the average speed required to satisfy the current passable window :

[0105]

[0106]

[0107] This represents the minimum static safety distance, with a range of [missing information]. , Indicates the current spacing. Indicates the safe time interval, in seconds, with a range of... In urban road following scenarios, a duration of 1.5 to 2.5 seconds can be selected; This indicates the distance from the vehicle's current position to the stop line or target control point. Indicates the remaining time of the currently passable window;

[0108] If a range-based access window is used, it can be further represented as:

[0109] Formula (22):

[0110]

[0111] In the formula: This indicates the distance from the vehicle's current position to the stop line or target control point, in meters (m). A suggested range is... ; This represents the remaining time in seconds for the currently passable window, satisfying the following conditions: ; This represents the average speed required to pass through the intersection within the current traffic window, expressed in m / s; and These represent the minimum and maximum guiding speed boundaries allowed through the window, respectively.

[0112] Step 4-2: Combining the speed limit correction, safe speed upper limit correction, and traffic guidance speed boundary correction, take the minimum value among them as the speed after traffic constraint correction. ;

[0113]

[0114] in Indicates speed limit This represents the predicted initial velocity value;

[0115] Step 4-3: Based on the vehicle's current speed Maximum permissible acceleration Maximum permissible deceleration and sampling period The speed, after traffic constraint correction, is then subjected to dynamic smoothing correction to obtain the final target speed prediction value. :

[0116]

[0117] in, Indicates will Limited to the range Inside. The range of values ​​is ; The range of values ​​is . The sampling period is expressed in seconds and satisfies the following conditions: .

[0118] In some embodiments, step 4 includes the following steps:

[0119] Step 5-1: Obtain the vehicle's actual speed at the next moment, and calculate the prediction error between the actual speed and the final target speed prediction value output in Step 4. :

[0120]

[0121] in, Indicates time The prediction error is expressed in m / s. The range is... However, under normal operating conditions, its absolute value is usually much smaller than this upper limit; Indicates time The actual measured speed; Indicates time The final predicted speed.

[0122] Step 5-2: Based on the prediction error, adjust the scene correction parameters corresponding to the current scene. Update:

[0123]

[0124] in, Indicates the online learning rate, ranging from to ;

[0125] Step 5-3: Based on the prediction error and the combined reliability of the vehicle-side perception data source and the V2X collaborative data source obtained in Step 1. Calculate the V2X fusion weights for the next time step. And determine the vehicle-side perception fusion weight for the next moment based on the V2X fusion weight. This enables online updates of the fusion weights;

[0126]

[0127]

[0128] in, Initial V2X weights;

[0129] Step 5-4: Output the final target velocity prediction result and use the updated scene correction parameters and fusion weights for velocity prediction in subsequent time steps.

[0130] Secondly, this solution provides a vehicle speed prediction system that integrates vehicle-side perception and V2X information, including: a data acquisition module, a preprocessing module, a credibility assessment module, a scene recognition module, an adaptive fusion module, a speed prediction module, and a constraint correction module;

[0131] The credibility assessment module is used to calculate the credibility of each data source based on time freshness, stability and consistency. The adaptive fusion module is used to determine the fusion weight of vehicle-side perception features and V2X collaborative features according to the current traffic scenario and the credibility. The constraint correction module is used to correct the initial speed prediction results based on road speed limit constraints, safe distance constraints, traffic light passage window constraints and dynamic smoothing constraints.

[0132] Thirdly, this solution provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0133] Fourthly, this solution provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor using the above-described method. This solution has the following beneficial effects:

[0134] (1) By incorporating time freshness, stability and consistency into the credibility assessment process, the interference of expired messages, jitter messages and conflict messages on speed prediction results can be reduced, and the robustness of multi-source data utilization can be improved.

[0135] (2) By explicitly introducing traffic scenarios into the fusion weight determination process, the vehicle-side perception features and V2X collaborative features can adopt differentiated contribution ratios in different scenarios, thereby enhancing the adaptability in complex traffic scenarios.

[0136] (3) By adding speed limit, safe distance, traffic light passage window and dynamic smoothing constraints after the initial speed prediction result, the output result can be further improved from "predictable" to "executable", which is convenient for implementation in assisted driving or vehicle control links.

[0137] (4) This invention is compatible with specific time series prediction models. Its innovations are focused on the credibility-driven scenario-related fusion and constraint correction link, which is easy to deploy on different vehicle models and different computing platforms. Attached Figure Description

[0138] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0139] Figure 1 The system architecture diagram shows the relationship between the vehicle-side perception unit, V2X communication unit, credibility assessment module, adaptive fusion module, speed prediction module, and constraint correction module.

[0140] Figure 2 The process flowchart shows the complete process from multi-source data acquisition to final speed output and online updates.

[0141] Figure 3 A diagram illustrating credibility assessment, showing how time freshness, stability, and consistency contribute to overall credibility. Detailed Implementation

[0142] A vehicle speed prediction method integrating vehicle-side perception and V2X information employs the following technical solution: acquiring vehicle-side perception data and V2X collaborative data; preprocessing the multi-source data and calculating the credibility of each data source from three dimensions: time freshness, stability, and consistency; adaptively weighting and fusing vehicle-side perception features and V2X collaborative features in conjunction with traffic scenarios; generating an initial speed prediction result based on historical fusion feature sequences; then correcting the initial speed prediction result by incorporating road speed limit constraints, safe following distance constraints, traffic light passage window constraints, and dynamic smoothing constraints to obtain the target speed prediction result; finally, updating the scene correction parameters and fusion weights online based on the error between the actual speed and the predicted speed.

[0143] like Figure 1-2 As shown, the specific steps include:

[0144] Step 1: Collect vehicle-side perception data and V2X collaborative data and perform data preprocessing; calculate the comprehensive reliability of the preprocessed data from three dimensions: time freshness, stability, and consistency.

[0145] Step 2: Based on the current traffic scenario and the credibility obtained in Step 1, calculate the fusion weight of vehicle-side perception features and V2X collaborative features to generate fused features;

[0146] Step 3: Input the fused historical feature sequence into the velocity prediction model to obtain the initial velocity prediction value within the preset time window in the future;

[0147] Step 4: Correct the initial prediction values ​​based on speed limits, safe following distances, traffic light passage windows, and vehicle dynamics boundaries;

[0148] Step 5: Output the final target velocity prediction result, and update the scene correction parameters and fusion weights online according to the prediction error. The update of the fusion weights depends on the prediction error and the comprehensive confidence obtained in Step 1.

[0149] Step 1 includes the following steps:

[0150] Step 1-1: Collect vehicle-side perception data and V2X collaborative data respectively. The vehicle-side perception data includes, but is not limited to, vehicle speed, vehicle acceleration, distance to the vehicle in front, relative speed of the vehicle in front, vehicle position, heading angle, and vehicle control status. The V2X collaborative data includes, but is not limited to, traffic light phase status, phase remaining time, road speed limit, map topology information, road event information, roadside guidance information, and perception sharing information. The V2X collaborative data is obtained by receiving and parsing collaborative information from V2V, V2I, or roadside units. The perception sharing information may include information about surrounding vehicles, but its scope is broader than that of surrounding vehicles, and also includes target-level and event-level environmental perception results generated and shared by roadside perception devices or other networked traffic participants.

[0151] Step 1-2: Perform data preprocessing on the collected multi-source data, aiming to unify asynchronous, multi-frequency, and different quality data into analyzable input. The preprocessing specifically includes at least one of time alignment, missing value compensation, outlier removal, and expired message filtering, and forms vehicle-side perception feature vector and V2X collaborative feature vector respectively.

[0152] After data collection, it is uniformly converted into a computable vehicle-side perception state vector, specifically:

[0153] Define the vehicle-side perception state vector :

[0154]

[0155] In the formula: Indicates time The vehicle-side perception state vector; This indicates the vehicle's speed, expressed in m / s, with a range of values ​​of [missing value]. The corresponding range is... arrive km / h, which can cover the typical speed range of urban roads, elevated roads and highways; This represents the longitudinal acceleration of the vehicle, measured in m / s². 2 The range of values ​​is Negative values ​​indicate deceleration, while positive values ​​indicate acceleration. This range covers both normal driving and emergency braking scenarios. This represents the longitudinal distance between the vehicle and the vehicle in front at the current moment, in meters (m), and its value range is... When the distance exceeds 200m, the direct impact of the preceding vehicle on short-term speed prediction is usually significantly reduced. The relative speed with respect to the vehicle in front is defined as follows: The unit is m / s, and the recommended value range is [value range missing]. . The vehicle location can be represented by two-dimensional projected coordinates or high-precision map reference coordinates. There is no fixed range of values, but a unified coordinate system should be used throughout the system. This represents the vehicle's heading angle, measured in rad. A recommended value range is [range missing]. . This indicates vehicle control state characteristics, which may include throttle opening, brake pedal opening, gear position, and drive mode. After normalization, the range is... .

[0156] Define V2X cooperative state vector :

[0157]

[0158] In the formula: Indicates time V2X cooperative state vector; This indicates the phase state of the traffic light, using integer encoding, such as red, yellow, green, and unknown state. The set of possible values ​​is... ; This represents the remaining time of the current signal phase, in seconds, and its value range is... For common urban intersections, this range is sufficient to cover the vast majority of actual control phases; This displays the speed limit information for the current road segment or lane, in m / s, with a value range of [value missing]. ; Indicates the intensity of a road event or the level of an obstacle, normalized to an interval. If discrete levels are used, the following can be taken: , respectively representing no event, minor event, moderate event, and serious event. The map topology and lane connectivity features, typically derived from MAP messages or high-precision maps, can be encoded into vectors and then normalized. . This indicates roadside guidance information, which may include suggested speed, target lane, or coordinated traffic signs, and covers a range of... m / s;

[0159] Define the total state vector :

[0160]

[0161] In the formula, Indicates time The global input state. This formula places the vehicle-side perception state and the V2X collaborative state under the same analytical framework, serving as the unified input basis for subsequent credibility assessment, adaptive fusion, and speed prediction.

[0162] Steps 1-3: Calculate the overall reliability of the preprocessed data from three dimensions: time freshness, stability, and consistency.

[0163] like Figure 3 As shown, steps 1-3 include the following steps:

[0164] Step 1-3-1: Calculate the data's time freshness, stability factor, and consistency factor respectively:

[0165] The first is calculated using the following formula. Class data source at time Time freshness :

[0166]

[0167] in, For the first The difference between the current time and the most recent valid update time of the data source. Indicates the first The time decay coefficient of the data source has a range of values. ; The range of values ​​is When the message arrived, ,at this time As the news becomes outdated, Monotonically decrease; larger values ​​can be used for high-time-sensitivity data such as traffic light status and broadcast messages from vehicles ahead, while smaller values ​​can be used for low-change data such as static maps; This represents the difference between the current time and the most recent valid update time of the data source, expressed in seconds, and satisfies... .

[0168] The first is calculated using the following formula. Stability factor of data source :

[0169]

[0170] in, This indicates the normalized variance of the data source within the most recent time window, with values ​​falling within the range of... Within the range, to avoid outliers dominating the market excessively; The range of values ​​is Using rational functions The reasons are as follows: First, this form is always bounded and will not diverge; second, when the variance of fluctuation is small, the penalty for stable data is lighter, which is conducive to preserving effective information; third, when the variance of fluctuation increases, the weight will drop rapidly, which can suppress data sources with severe jitter.

[0171] The first is calculated using the following formula. Consistency factor of data source :

[0172]

[0173] in, This represents the consistency adjustment coefficient, with a value range of [value missing]. , Indicates the first The normalization deviation between the class data source and the reference source satisfies ; The range of values ​​is For example, it can be defined as the deviation between the vehicle-to-vehicle (V2V) speed detected and the speed broadcast by the vehicle, or the deviation between the roadside recommended speed and the locally perceived congestion state. The consistency factor means that if a data source severely conflicts with other reliable data sources, the reliability of that data source at the current moment should be reduced. The reason for using an exponential function is that it is more sensitive to the amplification of deviations and can quickly reduce the corresponding weights when multi-source conflicts are significant, thereby reducing the interference of erroneous collaborative information on prediction.

[0174] Step 1-3-2: Calculate the comprehensive credibility based on the obtained time freshness, stability factor, and consistency factor using weighted average. :

[0175]

[0176] in , , The weights for freshness, stability, and consistency are respectively, all satisfying the following conditions. And the sum of the three is 1.

[0177] The weights for time freshness, stability, and consistency can be quantitatively determined based on historical calibration data. First, according to traffic scenario labels... The historical samples are grouped, and the optimal weights for the corresponding scenario are obtained with the objective of minimizing the initial velocity prediction error, provided that all weights are non-negative and the sum of the weights is 1.

[0178]

[0179] The constraints are:

[0180]

[0181] In the formula, Indicates the first Number of calibration samples in traffic-like scenarios Indicates the first The actual speed corresponding to each sample This represents the initial velocity prediction value obtained after applying candidate confidence weights. The above optimization can be implemented using grid search, projective gradient descent, or other optimization algorithms with simplex constraints.

[0182] When sufficient calibration samples are lacking, the initial weights for time freshness, stability, and consistency can be set to 0.4, 0.3, and 0.3, respectively. The weights are then calibrated based on the scenario; for example, increasing the weight for time freshness in an intersection scenario and increasing the weight for stability in a low-speed congestion following scenario. Once the number of valid samples accumulated for a particular scenario reaches a preset threshold, the weights are recalibrated using historical data from that scenario. This ensures that the weight allocation has a clear quantitative basis and can adapt to the data quality characteristics of different traffic scenarios.

[0183] Step 2 includes the following steps:

[0184] Step 2-1: Calculate the vehicle perception weights based on the combined reliability of the collected vehicle-side perception data and V2X collaborative data. and V2X weights Specifically:

[0185]

[0186]

[0187] in This represents the overall reliability of the vehicle-side sensing sources. Represents the overall credibility of V2X message sources;

[0188] Obtain the fused integrated feature vector

[0189]

[0190] Step 2-2: Obtain fused features based on the weights obtained in Step 2-1 and the traffic scene labels. Explicitly incorporating scene labels into the fusion process significantly improves model interpretability and practical adaptability, specifically:

[0191]

[0192] To represent traffic scene labels, it is recommended to use a finite set. These correspond to scenarios such as free driving, following other vehicles, intersections, traffic congestion, and event guidance. This represents the normalized vehicle-side perception feature vector. This represents the normalized V2X collaborative feature vector. and These represent the scene-related vehicle perception weights and V2X weights, respectively, both located in... Within the range.

[0193] Before feature fusion, the vehicle-side perception features and V2X collaborative features are first mapped to a common feature space of the same dimension:

[0194]

[0195]

[0196] In the formula, , These represent the projection matrices of the vehicle-side perception features and the V2X collaborative features, respectively. , Let each represent the corresponding bias vector, such that... and They have the same dimensions.

[0197] For traffic scenarios Pre-set or calibrate the prior coefficients of the vehicle-side perception scenario using historical data. Prior coefficients for V2X scenarios The final fusion weights are calculated using the following formula:

[0198]

[0199]

[0200] In the formula, A positive number is used to prevent the denominator from being zero. The fusion feature is represented as:

[0201]

[0202] The above values ​​are only initial calibration values ​​and do not constitute a limitation on the protection scope. In free driving and following scenarios, the vehicle's status, distance to the vehicle in front, and relative speed have a significant impact on short-term speed changes, therefore, the vehicle-side perception prior coefficient is increased; in intersection scenarios, traffic light phase, remaining time, and map lane connectivity are more critical, therefore, the V2X prior coefficient is increased; in event-guided scenarios, road events and roadside perception shared information have a strong impact on vehicle speed changes, therefore, the V2X prior coefficient is further increased.

[0203] Step 3 includes the following steps:

[0204] Step 3-1: Arrange the fusion features generated in Step 2 in chronological order to construct a historical fusion feature sequence; define the historical fusion feature sequence:

[0205]

[0206] In the formula: Indicates the end time Historical fusion feature sequence; Indicates the length of the history window, satisfying The range is If the system sampling frequency is 10Hz, then the preferred frequency can be... to The fixed-length sliding window corresponds to a historical window of 1 to 3 seconds. Vehicle speed prediction relies more on recent dynamics; an excessively long historical window introduces outdated information and redundant noise, while also increasing the computational burden. Conversely, an excessively short window cannot fully reflect the speed change trend and control intent. Therefore, the fixed-length sliding window achieves a better balance between information sufficiency and real-time performance.

[0207] Step 3-2: Input the historical fusion feature sequence into the pre-trained speed prediction model;

[0208] Step 3-3: Using the velocity prediction model, perform time-series recursion or direct mapping to output the initial velocity prediction values ​​for each moment within a preset future time window. This yields a multi-step prediction sequence:

[0209]

[0210] In the formula: This represents the initial prediction velocity in one step before analytical constraint correction, in m / s, and ranges from [value missing]. ; Indicates the future The initial predicted velocity sequence of the step. Indicates the prediction step size, satisfying , range .

[0211] The velocity prediction model is constructed using a gated recurrent unit network. The model input is the value up to the current time. Historical fusion feature sequence:

[0212]

[0213] Among them, the historical window length It can be set to 10 to 30, corresponding to the most recent 1 to 3 seconds of historical information at a sampling frequency of 10Hz; the fusion features at each moment. All are located in the same Dimensional public feature space, It can be set to 16 to 128, therefore the dimension of the model input tensor is... .

[0214] The model consists of an input normalization layer, at least one GRU temporal coding layer, and a fully connected output layer. The GRU temporal coding layer recursively processes the historical fused features to obtain the latent state at the final time step. The time series prediction function can be specifically written as:

[0215]

[0216]

[0217] In the formula: Indicates time The hidden state vector can be set as follows: , among which dimension The range is . This represents the hidden state update function, which can be implemented by an LSTM unit, a GRU unit, or other gated structures. and These represent the output layer weight matrix and bias vector, respectively.

[0218] The GRU network is set up with two layers, with the first and second layers having hidden state dimensions of 64 and 32 respectively, and a random deactivation layer with a drop-off ratio of 0.1 to 0.3 can be set between the two layers.

[0219] For one-step velocity prediction, the fully connected output layer contains one output node and generates the initial predicted velocity for the next time step based on the hidden state:

[0220]

[0221] For multi-step velocity prediction, the fully connected output layer includes Each output node directly generates the future. Initial velocity prediction sequence at each time step:

[0222]

[0223] in, It can be set from 1 to 30. The Softplus function is used to ensure that the model output is non-negative. The model output is the initial predicted speed before adjustments for road speed limits, safe following distances, traffic light windows, and vehicle dynamics boundaries.

[0224] The model training data consists of historical vehicle operation data with completion time alignment and outlier handling. A length of [length missing] is used. The sliding window is used to construct training samples, and the previous The fused features at each time step are used as model input, and then... The measured vehicle speed at each moment is used as the supervision label. The training data can be divided into training set, validation set and test set in chronological order to avoid data from the same continuous journey appearing in both the training set and the test set.

[0225] During training, continuous features are standardized using the mean and standard deviation obtained from the training set statistics, while discrete features such as signal phase and scene labels are processed using cross-entropy encoding or one-hot encoding. The model uses a joint loss function. Conduct training:

[0226]

[0227] in, This represents the prediction error term; This represents the first-order difference smoothing term of the predicted sequence, used to suppress high-frequency jitter; This represents a consistency constraint term used to penalize over-rate predictions during the training phase, specifically:

[0228]

[0229] in, Represents the number of training samples, satisfying ; Indicates the first The true speed of each sample; Indicates the first Prediction speed per sample;

[0230]

[0231]

[0232] in, Indicates the first The speed limit value corresponding to each sample.

[0233] Model parameters Updates are performed via error backpropagation. The optimizer can be the Adam optimizer, and the initial learning rate can be set to... to The batch size can be set from 32 to 256, and the number of training rounds can be determined by an early stopping strategy based on the validation set loss.

[0234] After the model completes offline training and reaches the preset error requirement, the model parameters are fixed. The model is deployed to in-vehicle or edge computing devices to perform real-time inference. During the online operation phase of the vehicle, the main updates are to the scene correction parameters and multi-source fusion weights. It does not require retraining the entire time-series prediction model in each sampling period, thus ensuring real-time computing performance and model stability.

[0235] Step 4 includes the following steps:

[0236] Step 4-1: Calculate the upper bound of the safe speed based on the safe time interval. and the average speed required to satisfy the current passable window :

[0237]

[0238]

[0239] This represents the minimum static safety distance, with a range of [missing information]. , Indicates the current spacing. Indicates the safe time interval, in seconds, with a range of... In urban road following scenarios, a duration of 1.5 to 2.5 seconds can be selected; This indicates the distance from the vehicle's current position to the stop line or target control point. Indicates the remaining time of the currently passable window;

[0240] If a range-based access window is used, it can be further represented as:

[0241] Formula (22):

[0242]

[0243] In the formula: This indicates the distance from the vehicle's current position to the stop line or target control point, in meters (m). A suggested range is... ; This represents the remaining time in seconds for the currently passable window, satisfying the following conditions: ; This represents the average speed required to pass through the intersection within the current traffic window, expressed in m / s; and These represent the minimum and maximum guiding speed boundaries allowed through the window, respectively.

[0244] Step 4-2: Combining the speed limit correction, safe speed upper limit correction, and traffic guidance speed boundary correction, take the minimum value among them as the speed after traffic constraint correction. ;

[0245]

[0246] in Indicates speed limit This represents the predicted initial velocity value;

[0247] Step 4-3: Based on the vehicle's current speed Maximum permissible acceleration Maximum permissible deceleration and sampling period The speed, after traffic constraint correction, is then subjected to dynamic smoothing correction to obtain the final target speed prediction value. :

[0248]

[0249] in, Indicates will Limited to the range Inside. The range of values ​​is ; The range of values ​​is . The sampling period is expressed in seconds and satisfies the following conditions: .

[0250] Step 4 includes the following steps:

[0251] Step 5-1: Obtain the vehicle's actual speed at the next moment, and calculate the prediction error between the actual speed and the final target speed prediction value output in Step 4. :

[0252]

[0253] in, Indicates time The prediction error is expressed in m / s. The range is... However, under normal operating conditions, its absolute value is usually much smaller than this upper limit; Indicates time The actual measured speed; Indicates time The final predicted speed.

[0254] Step 5-2: Based on the prediction error, adjust the scene correction parameters corresponding to the current scene. Update:

[0255]

[0256] in, Indicates the online learning rate, ranging from to ;

[0257] Step 5-3: Based on the prediction error and the combined reliability of the vehicle-side perception data source and the V2X collaborative data source obtained in Step 1. Calculate the V2X fusion weights for the next time step. And determine the vehicle-side perception fusion weight for the next moment based on the V2X fusion weight. This enables online updates of the fusion weights;

[0258]

[0259]

[0260] in, Initial V2X weights;

[0261] Step 5-4: Output the final target velocity prediction result and use the updated scene correction parameters and fusion weights for velocity prediction in subsequent time steps.

[0262] Example 1: Intersection Green Light Remaining Time Constraint Scenario

[0263] Scenario Setting: The target vehicle is located at the entrance lane of a city intersection, currently within the green light window. Vehicle-side perception is effective, and V2X roadside information is of high freshness. The vehicle needs to predict the target speed in the next moment and make corrections based on speed limits, safe distance, signal clearance window, and dynamic boundaries.

[0264]

[0265] (1) Calculation of the overall credibility of vehicle-side and V2X

[0266] Let the time decay coefficient of the vehicle-side sensing source be... Update time difference Fluctuation variance Normalization bias compared to other sources Consistency coefficient Calculate according to formulas (4) to (7):

[0267]

[0268]

[0269]

[0270]

[0271] Let the time decay coefficient of the V2X message source be... Update time difference Fluctuation variance Normalization bias compared to other sources Consistency coefficient ,but:

[0272]

[0273]

[0274]

[0275]

[0276] Further calculate the fusion weights using formulas (10) and (11):

[0277]

[0278]

[0279] Therefore, in this embodiment, the overall reliability of vehicle-side perception and V2X messages is similar, but vehicle-side perception is slightly higher, so the weight of vehicle-side perception is slightly greater than that of V2X. This result is consistent with reality: radar and vehicle status updates are fast, and although roadside messages are highly structured, they have a certain communication delay.

[0280] (2) Initial velocity prediction and constraint correction

[0281] Assume the time series prediction module outputs the initial prediction rate for the next time step based on the historical window. The output can be obtained from formula (14), and the model output result is directly substituted here for subsequent constraint correction.

[0282] First, adjust the speed limit according to formula (18):

[0283]

[0284] Next, the upper limit of the safe speed is calculated according to formulas (19) and (20):

[0285]

[0286] Then, calculate the upper bound of the average speed required to satisfy the current green light passage window according to formula (21):

[0287]

[0288] According to formula (23), the traffic constraint correction result can be obtained:

[0289]

[0290] Finally, dynamic boundaries are applied according to formula (24). The executable speed range can be obtained from the current speed and sampling period:

[0291]

[0292]

[0293] Therefore, the final predicted speed in this embodiment is 13.40 m / s. This result is lower than the initial output of 15.2 m / s, indicating that when approaching an intersection and with a limited remaining green light time, the signal passage window and dynamic boundary together suppress excessively high predicted speeds.

[0294] (3) Online update example

[0295] Let the actual velocity observed at the next sampling time be... And take the current scene correction parameters Online learning rate Then, online updates are performed according to formulas (25) to (28).

[0296]

[0297]

[0298]

[0299]

[0300] Since the actual speed was slightly higher than the corrected predicted speed, the V2X weights were slightly increased after the online update, indicating that roadside information still has a positive guiding role in this intersection sample.

[0301] Example 2: V2X Deprecation Suppression in Vehicle-to-Vehicle Scenarios

[0302] To illustrate that the credibility mechanism of this invention does not mechanically increase V2X weight in all scenarios, a following scenario is given where V2X messages expire but vehicle-side perception is dominant.

[0303]

[0304] In this embodiment, the vehicle-side sensing source parameters are set to... , , , V2X message parameters are set to , , , .but:

[0305]

[0306]

[0307]

[0308]

[0309]

[0310]

[0311]

[0312]

[0313]

[0314] The results show that when V2X messages are clearly expired, the system automatically transfers the weights to the vehicle's perception. If a fixed 50%:50% fusion is used, expired messages will still cause significant disturbances to the output; however, this solution can automatically suppress this interference.

[0315] This invention provides a method, system, device, and storage medium for vehicle speed prediction that integrates vehicle-side perception and V2X information. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for predicting vehicle speed by integrating vehicle-side perception and V2X information, characterized in that, include: Step 1: Collect vehicle-side perception data and V2X collaborative data and perform data preprocessing; For the preprocessed data, the overall credibility is calculated from three dimensions: time freshness, stability, and consistency. Step 2: Based on the current traffic scenario and the credibility obtained in Step 1, calculate the fusion weight of vehicle-side perception features and V2X collaborative features to generate fused features; Step 3: Input the fused historical feature sequence into the velocity prediction model to obtain the initial velocity prediction value within the preset time window in the future; Step 4: Correct the initial prediction values ​​based on speed limits, safe following distances, traffic light passage windows, and vehicle dynamics boundaries; Step 5: Output the final target velocity prediction result, and update the scene correction parameters and fusion weights online according to the prediction error. The update of the fusion weights depends on the prediction error and the comprehensive confidence obtained in Step 1.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1-1: Collect vehicle-side perception data and V2X collaborative data respectively. The vehicle-side perception data includes, but is not limited to, vehicle speed, vehicle acceleration, distance to the vehicle in front, relative speed of the vehicle in front, vehicle position, heading angle, and vehicle control status. The V2X collaborative data includes, but is not limited to, traffic light phase status, phase remaining time, road speed limit, map topology information, road event information, roadside guidance information, and perception sharing information. Step 1-2: Perform data preprocessing on the collected multi-source data. The preprocessing specifically includes at least one of time alignment, missing value compensation, outlier removal, and expired message filtering, and forms vehicle-side perception feature vector and V2X collaborative feature vector respectively. Steps 1-3: Calculate the overall reliability of the preprocessed data from three dimensions: time freshness, stability, and consistency.

3. The method according to claim 2, characterized in that, Steps 1-3 include the following steps: Step 1-3-1: Calculate the data's time freshness, stability factor, and consistency factor respectively: The first is calculated using the following formula. Class data source at time Time freshness : in, For the first The difference between the current time and the most recent valid update time of the data source. Indicates the first The time decay coefficient of the data source has a range of values. ; The first is calculated using the following formula. Stability factor of data source : in, This indicates the normalized variance of the data source within the most recent time window, with values ​​falling within the range of... Within the range; The first is calculated using the following formula. Consistency factor of data source : in, This represents the consistency adjustment coefficient, with a value range of [value missing]. , Indicates the first The normalization deviation between the class data source and the reference source satisfies ; Step 1-3-2: Calculate the comprehensive credibility based on the obtained time freshness, stability factor, and consistency factor using weighted average. : in , , The weights for freshness, stability, and consistency are respectively, all satisfying the following conditions. And the sum of the three is 1.

4. The method according to claim 1, characterized in that, Step 2 includes the following steps: Step 2-1: Calculate the vehicle perception weights based on the combined reliability of the collected vehicle-side perception data and V2X collaborative data. and V2X weights Specifically: in This represents the overall reliability of the vehicle-side sensing sources. Represents the overall credibility of V2X message sources; Step 2-2: Obtain fused features based on the weights obtained in Step 2-1 and the traffic scene labels. Specifically: To represent traffic scene labels, it is recommended to use a finite set. These correspond to scenarios such as free driving, following other vehicles, intersections, traffic congestion, and event guidance. This represents the normalized vehicle-side perception feature vector. This represents the normalized V2X collaborative feature vector; Before feature fusion, the vehicle-side perception features and V2X collaborative features are mapped to a common feature space of the same dimension: In the formula, , These represent the projection matrices of the vehicle-side perception features and the V2X collaborative features, respectively. , Let each represent the corresponding bias vector, such that... and Having the same dimensions; For traffic scenarios Pre-set or calibrate the prior coefficients of the vehicle-side perception scenario using historical data. Prior coefficients for V2X scenarios The final fusion weights are calculated using the following formula: In the formula, A positive number is used to prevent the denominator from being zero. The fusion feature is represented as:

5. The method according to claim 1, characterized in that, Step 3 includes the following steps: Step 3-1: Arrange the fusion features generated in Step 2 in chronological order to construct a historical fusion feature sequence; Step 3-2: Input the historical fusion feature sequence into the pre-trained speed prediction model; Step 3-3: Perform time-series recursion or direct mapping using the speed prediction model to output the initial speed prediction values ​​for each moment within the preset future time window.

6. The method according to claim 1, characterized in that, Step 4 includes the following steps: Step 4-1: Calculate the upper bound of the safe speed based on the safe time interval. and the average speed required to satisfy the current passable window : Indicates the minimum static safety distance. Indicates the current spacing. Indicates the safe time interval; This indicates the distance from the vehicle's current position to the stop line or target control point. Indicates the remaining time of the currently passable window; Step 4-2: Combining the speed limit correction, safe speed upper limit correction, and traffic guidance speed boundary correction, take the minimum value among them as the speed after traffic constraint correction. ; in Indicates speed limit This represents the predicted initial velocity value; Step 4-3: Based on the vehicle's current speed Maximum permissible acceleration Maximum permissible deceleration and sampling period The speed, after traffic constraint correction, is then subjected to dynamic smoothing correction to obtain the final target speed prediction value. : in, Indicates will Limited to the range Inside.

7. The method according to claim 1, characterized in that, Step 4 includes the following steps: Step 5-1: Obtain the vehicle's actual speed at the next moment, and calculate the prediction error between the actual speed and the final target speed prediction value output in Step 4. ; Step 5-2: Based on the prediction error, adjust the scene correction parameters corresponding to the current scene. Update: in, Indicates the online learning rate, ranging from to ; Step 5-3: Based on the prediction error and the combined reliability of the vehicle-side perception data source and the V2X collaborative data source obtained in Step 1. Calculate the V2X fusion weights for the next time step. And determine the vehicle-side perception fusion weight for the next moment based on the V2X fusion weight. This enables online updates of the fusion weights; in, Initial V2X weights; Step 5-4: Output the final target velocity prediction result and use the updated scene correction parameters and fusion weights for velocity prediction in subsequent time steps.

8. A vehicle speed prediction system integrating vehicle-side perception and V2X information, characterized in that, include: The system includes a data acquisition module, a preprocessing module, a credibility assessment module, a scene recognition module, an adaptive fusion module, a speed prediction module, and a constraint correction module. The credibility assessment module is used to calculate the credibility of each data source based on time freshness, stability and consistency. The adaptive fusion module is used to determine the fusion weight of vehicle-side perception features and V2X collaborative features according to the current traffic scenario and the credibility. The constraint correction module is used to correct the initial speed prediction results based on road speed limit constraints, safe distance constraints, traffic light passage window constraints and dynamic smoothing constraints.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the method according to any one of claims 1 to 7.

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

Images