Track estimation method and device
By using bidirectional neural networks and nonlinear filtering techniques in trajectory estimation, the accuracy and efficiency problems of the prior art when dealing with nonlinear vehicle dynamic and high noise environments are solved, and higher trajectory prediction accuracy and robustness are achieved and optimize the use of computing resources.
Patent Information
- Application Number
- CN202411441453.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-15
- Publication Date
- 2025-02-18
AI Technical Summary
When existing trajectory estimation technologies deal with the nonlinear characteristics of vehicle dynamics, high noise environments and inefficient data processing, it is difficult to achieve accurate trajectory prediction.
Using a bidirectional neural network to predict the trajectory observation data at the current moment, combined with nonlinear filtering technology, a priori estimation and covariance prediction are performed based on the trajectory data at the previous moment, and the Kalman gain and optimal estimate value are updated to improve the trajectory data. accuracy.
By effectively capturing the front and rear information of vehicle dynamics, it improves the accuracy of trajectory prediction, it is suitable for dealing with nonlinear problems of vehicle movement in traffic environments, and improves the accuracy and robustness of position estimation, reduces computing resource consumption, and achieves faster Processing speed.
Smart Images
Figure CN119471744A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of vehicle-related technologies, and in particular to a trajectory estimation method, electronic equipment, storage medium and computer program product. Background Art
[0002] In the Vehicle-to-Everything (V2X) industry, real-time and accurate trajectory estimation is essential to achieve efficient vehicle communication and intelligent transportation systems. Key areas of V2X technology include communications between vehicles, vehicles and infrastructure, and vehicles and pedestrians, all of which rely on accurate location and movement information.
[0003] Existing trajectory estimation technologies usually use the Global Positioning System (GPS) and various sensors (such as accelerometers and gyroscopes) to obtain the vehicle position and movement status. However, these technologies have the following problems:
[0004] Nonlinear issues: The nonlinear nature of vehicle dynamics makes it difficult to accurately predict the future position of the vehicle using traditional linear filtering techniques (such as Kalman filtering), especially in highly dynamic environments such as urban driving.
[0005] High noise sensitivity: GPS signals are easily interfered with in urban canyons or tunnels, resulting in inaccurate location information. In addition, other sensors may also produce errors due to environmental factors or equipment quality issues.
[0006] Data processing efficiency: Existing technologies often require complex data fusion processing to improve the accuracy of trajectory prediction, which not only increases the computational burden but may also delay information processing and response speed. Summary of the invention
[0008] Based on this, it is necessary to provide a trajectory estimation method, electronic device, storage medium and computer program product to address the technical problems of nonlinearity, high noise sensitivity and low data processing efficiency in trajectory estimation in the prior art.
[0009] The present invention provides a trajectory estimation method, comprising:
[0010] Obtain the original trajectory data from the sensor and predict the trajectory observation data at the current moment through a bidirectional neural network;
[0011] Through nonlinear filtering, based on the optimal estimate of the trajectory data at the previous moment, the prior estimate of the trajectory data at the current moment is predicted, and based on the optimal estimate covariance at the previous moment, the prior estimate covariance at the current moment is predicted;
[0012] Based on the prior estimated covariance at the current moment, the Kalman gain at the current moment is updated; based on the trajectory observation data at the current moment, the prior estimated value of the trajectory data at the current moment and the Kalman gain, the optimal estimated value of the trajectory data at the current moment is updated; based on the Kalman gain at the current moment and the prior estimated covariance, the optimal estimated covariance at the current moment is updated.
[0013] Furthermore, the obtaining of original trajectory data from the sensor and predicting the trajectory observation data at the current moment through a bidirectional neural network includes:
[0014] Acquire multiple sets of trajectory raw data from the sensor through window sliding processing;
[0015] Multiple groups of original trajectory data processed by window sliding are input into a bidirectional neural network to obtain trajectory observation data at the current moment predicted by the bidirectional neural network.
[0016] Furthermore, the step of obtaining multiple groups of original trajectory data from the sensor through window sliding processing includes:
[0017] Obtain the time series data consisting of all the original trajectory data of the sensor from the current time to the starting time;
[0018] A window is controlled from the time series data to slide from the current time to the starting time in sequence according to a preset step length, and each time the window slides, the original trajectory data located in the window is extracted from the time series data as a group of original trajectory data.
[0019] Furthermore, the bidirectional neural network is a bidirectional long short-term memory network.
[0020] Further, the method of predicting the prior estimation value of the trajectory data at the current moment based on the optimal estimation value of the trajectory data at the previous moment through nonlinear filtering, and predicting the prior estimation covariance of the current moment based on the optimal estimation covariance at the previous moment, includes:
[0021] Through unscented Kalman filtering, the prior estimate of the trajectory data at the current moment is predicted based on the optimal estimate of the trajectory data at the previous moment, and the prior estimate covariance at the current moment is predicted based on the optimal estimate covariance at the previous moment.
[0022] Further, the updating of the optimal estimate of the trajectory data at the current moment based on the trajectory observation data at the current moment, the prior estimate of the trajectory data at the current moment, and the Kalman gain includes:
[0023] The optimal estimate of the trajectory data at the current moment is updated to where x k|k is the optimal estimate of the trajectory data at the current moment, K is the prior estimate of the trajectory data at the current moment. k is the Kalman gain at the current moment, y k is the trajectory observation data at the current moment, and H is the observation matrix.
[0024] Further, the updating of the optimal estimated covariance at the current moment based on the Kalman gain at the current moment and the prior estimated covariance includes:
[0025] Update the optimal estimated covariance at the current moment to: Among them, P k|k is the optimal estimated covariance at the current moment, K k is the Kalman gain at the current moment, H is the observation matrix, is the prior estimated covariance at the current moment.
[0026] The present invention provides an electronic device, comprising:
[0027] at least one processor; and,
[0028] a memory communicatively connected to at least one of the processors; wherein,
[0029] The memory stores instructions that can be executed by at least one of the processors. The instructions are executed by at least one of the processors to enable the at least one of the processors to perform the trajectory estimation method as described above.
[0030] The present invention provides a storage medium, wherein the storage medium stores computer instructions, and when a computer executes the computer instructions, it is used to execute all steps of the trajectory estimation method as described above.
[0031] The present invention provides a computer program product, comprising a computer program / instruction, which implements the trajectory estimation method as described above when executed by a processor.
[0032] The present invention predicts the trajectory observation data at the current moment through a bidirectional neural network, which can effectively capture the contextual information of vehicle dynamics, improve the accuracy of trajectory prediction, and is suitable for dealing with nonlinear problems of vehicle movement in traffic environments. At the same time, through nonlinear filtering, the observation data and the predicted data are combined, which can more accurately handle sensor noise and GPS signal interference, and improve the accuracy and robustness of position estimation. The present invention combines bidirectional neural networks and nonlinear filtering to estimate trajectories, which not only improves the accuracy of data processing, but also reduces the consumption of computing resources through optimization algorithms, achieves faster processing speeds, and meets the needs of real-time vehicle-road-cloud communications. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1A workflow diagram of a trajectory estimation method according to an embodiment of the present invention;
[0034] Figure 2 This is a workflow diagram of a trajectory estimation method according to another embodiment of the present invention;
[0035] Figure 3 This is a schematic diagram of the effect of an example of the present invention;
[0036] Figure 4 The figure is a schematic diagram of the hardware structure of an electronic device of the present invention. DETAILED DESCRIPTION
[0037] The specific embodiments of the present invention are further described below in conjunction with the accompanying drawings. The same components are represented by the same reference numerals. It should be noted that the words "front", "rear", "left", "right", "upper" and "lower" used in the following description refer to directions in the accompanying drawings, and the words "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.
[0038] The following is an explanation of the key terms or technical abbreviations used in this embodiment, including:
[0039] V2X communication (Vehicle-to-Everything Communication):
[0040] Refers to the communication technology between vehicles and their surroundings (including other vehicles, pedestrians, road infrastructure, etc.). This communication is a key technology for realizing intelligent transportation systems and improving road safety.
[0041] Global Positioning System (GPS):
[0042] Global Positioning System, a system that uses satellite signals to provide geographic location information. It allows users anywhere on Earth to obtain location and time information at any point in time.
[0043] Long Short-Term Memory (LSTM):
[0044] Long short-term memory networks, a special type of recurrent neural network used to process and predict important events in time series data with long intervals and delays.
[0045] Unscented Kalman Filter (UKF):
[0046] Unscented Kalman filter, a variant of Kalman filter for dealing with nonlinear systems. It avoids the linearization process in traditional Kalman filter by using Sigma points to approximate the prediction and update process.
[0047] Bidirectional Networks:
[0048] In the neural network model, the bidirectional structure can process the past and future information of the data at the same time, which is very effective for sequence data processing and is often used in fields such as speech recognition, text processing, and time series analysis.
[0049] Sigma Points:
[0050] The sample points used in the Unscented Kalman Filter are used to capture the probability distribution in the prediction and update steps. The selection of these points is critical to ensure the accuracy of the filtering process.
[0051] like Figure 1 The figure shows a workflow diagram of a trajectory estimation method according to an embodiment of the present invention, comprising:
[0052] Step S101, obtaining original trajectory data from the sensor, and predicting the trajectory observation data at the current moment through a bidirectional neural network;
[0053] Step S102, predicting the prior estimation value of the trajectory data at the current moment based on the optimal estimation value of the trajectory data at the previous moment through nonlinear filtering, and predicting the prior estimation covariance of the current moment based on the optimal estimation covariance at the previous moment;
[0054] Step S103, based on the prior estimated covariance at the current moment, the Kalman gain at the current moment is updated; based on the trajectory observation data at the current moment, the prior estimated value of the trajectory data at the current moment and the Kalman gain, the optimal estimated covariance at the current moment is updated based on the Kalman gain at the current moment and the prior estimated covariance.
[0055] Specifically, the present invention can be applied to electronic devices with processing capabilities, such as a controller of a vehicle, such as an electronic control unit (ECU) of a vehicle.
[0056] First, step S101 is executed to obtain the original trajectory data from the sensor, and the trajectory observation data at the current moment is predicted through a bidirectional neural network.
[0057] Specifically, the sensor includes GPS or other sensors. Accurate and available trajectory data is extracted from the original GPS or other sensors as the original trajectory data, and then the original trajectory data is input into the bidirectional neural network, and the bidirectional neural network predicts the trajectory data at the current moment as the trajectory observation data. The trajectory data may be the trajectory point position.
[0058] By using a bidirectional neural network to process time series data and capture the data dependencies before and after each time point, the traditional unidirectional network can only consider the limitations of past information. By introducing a bidirectional neural network structure in trajectory prediction, it is used to simultaneously learn the forward and backward dependencies of trajectory data, which improves the accuracy of prediction and the response speed to future data.
[0059] Then, step S102 is executed to predict the prior estimation value of the trajectory data at the current moment based on the optimal estimation value of the trajectory data at the previous moment through nonlinear filtering, and predict the prior estimation covariance of the current moment based on the optimal estimation covariance at the previous moment.
[0060] Specifically, step S102 is a prediction step of nonlinear filtering. Nonlinear filtering estimates the prior value and the prior estimated covariance through the state model. Then, the estimated prior value is corrected based on the observed value, so that sensor noise and GPS signal interference can be processed more accurately. At the same time, this embodiment uses nonlinear filtering to deal with the nonlinear problem of vehicle motion and improve the accuracy and robustness of position estimation.
[0061] In one embodiment, the nonlinear filter is an unscented Kalman filter.
[0062] Finally, step S103 is executed to update the Kalman gain at the current moment based on the prior estimated covariance at the current moment, update the optimal estimated value of the trajectory data at the current moment based on the trajectory observation data at the current moment, the prior estimated value of the trajectory data at the current moment, and the Kalman gain, and update the optimal estimated covariance at the current moment based on the Kalman gain at the current moment and the prior estimated covariance.
[0063] Among them, step S103 is the updating step of nonlinear filtering. This embodiment adopts step S101, based on the output of the bidirectional neural network structure as observation data, to update the optimal estimate of the trajectory data at the current moment. The combination of the bidirectional neural network structure and nonlinear filtering not only improves the accuracy of data processing, but also reduces the consumption of computing resources through the optimization algorithm, achieving faster processing speed and meeting the needs of real-time vehicle-road-cloud communication.
[0064] The present invention predicts the trajectory observation data at the current moment through a bidirectional neural network, which can effectively capture the contextual information of vehicle dynamics, improve the accuracy of trajectory prediction, and is suitable for dealing with nonlinear problems of vehicle movement in traffic environments. At the same time, through nonlinear filtering, the observation data and the predicted data are combined, which can more accurately handle sensor noise and GPS signal interference, and improve the accuracy and robustness of position estimation. The present invention combines bidirectional neural networks and nonlinear filtering to estimate trajectories, which not only improves the accuracy of data processing, but also reduces the consumption of computing resources through optimization algorithms, achieves faster processing speeds, and meets the needs of real-time vehicle-road-cloud communications.
[0065] like Figure 2 FIG. 1 is a flowchart of a trajectory estimation method in another embodiment of the present invention, comprising:
[0066] Step S201, obtaining multiple groups of original trajectory data processed by window sliding from the sensor.
[0067] In one embodiment, the step of acquiring multiple groups of raw trajectory data from the sensor through window sliding processing includes:
[0068] Obtain the time series data consisting of all the original trajectory data of the sensor from the current time to the starting time;
[0069] A window is controlled from the time series data to slide from the current time to the starting time in sequence according to a preset step length, and each time the window slides, the original trajectory data located in the window is extracted from the time series data as a group of original trajectory data.
[0070] Step S202, inputting multiple groups of trajectory original data processed by window sliding into a bidirectional neural network to obtain trajectory observation data at the current moment predicted by the bidirectional neural network, wherein the bidirectional neural network is a bidirectional long short-term memory network.
[0071] Step S203, using unscented Kalman filtering, predicting the priori estimated value of the trajectory data at the current moment based on the optimal estimated value of the trajectory data at the previous moment, and predicting the priori estimated covariance of the current moment based on the optimal estimated covariance at the previous moment.
[0072] Step S204, based on the prior estimated covariance at the current moment, the Kalman gain at the current moment is updated; based on the trajectory observation data at the current moment, the prior estimated value of the trajectory data at the current moment and the Kalman gain, the optimal estimated covariance at the current moment is updated based on the Kalman gain at the current moment and the prior estimated covariance.
[0073] In one embodiment, the updating of the optimal estimate of the trajectory data at the current moment based on the trajectory observation data at the current moment, the prior estimate of the trajectory data at the current moment, and the Kalman gain includes:
[0074] The optimal estimate of the trajectory data at the current moment is updated to where x k|k is the optimal estimate of the trajectory data at the current moment, K is the prior estimate of the trajectory data at the current moment. k is the Kalman gain at the current moment, y k is the trajectory observation data at the current moment, and H is the observation matrix.
[0075] In one embodiment, the updating of the optimal estimated covariance at the current moment based on the Kalman gain at the current moment and the prior estimated covariance includes:
[0076] Update the optimal estimated covariance at the current moment to: Among them, P k|k is the optimal estimated covariance at the current moment, K k is the Kalman gain at the current moment, H is the observation matrix, is the prior estimated covariance at the current moment.
[0077] Specifically, step S201 is first performed to obtain multiple groups of original trajectory data processed by window sliding from the sensor.
[0078] Specifically, in the data preprocessing stage, the window sliding technology is applied to generate sequence data with time windows, which are used to input the bidirectional neural network model, optimize the data input structure, and improve the processing efficiency and prediction performance. At the same time, when training the bidirectional neural network model, the window sliding technology is also used to generate sequence data with time windows for training the bidirectional neural network model.
[0079] The trajectory data is preprocessed through the window sliding technique to provide a suitable input format for the bidirectional neural network, enabling the model to understand and predict the local dynamics of the trajectory more effectively.
[0080] In one embodiment, the step of acquiring multiple groups of raw trajectory data from the sensor through window sliding processing includes:
[0081] Obtain the time series data consisting of all the original trajectory data of the sensor from the current time to the starting time;
[0082] A window is controlled from the time series data to slide from the current time to the starting time in sequence according to a preset step length, and each time the window slides, the original trajectory data located in the window is extracted from the time series data as a group of original trajectory data.
[0083] Specifically, the sensor includes GPS and other sensors, and the data output by the sensor is raw trajectory data.
[0084] In some embodiments, data preprocessing is also included.
[0085] The purpose of data preprocessing is to extract accurate and usable trajectory data from the original GPS or other sensors, laying the foundation for subsequent data analysis and model training.
[0086] Data preprocessing includes:
[0087] Data collection: Get GPS data containing timestamp and location coordinates (such as longitude and latitude).
[0088] Data cleaning: Identify and remove outliers, such as data points with speeds that do not conform to physical limits or records with sudden changes in location. These data may be caused by loss of GPS signals or sensor errors.
[0089] Data cleaning example: Suppose there is a sequence of GPS data points, and the coordinates of one point suddenly jump from (34.0522, -118.2437) to (0, 0), and then jump back to the normal range. This point is obviously erroneous data and should be removed.
[0090] Then, window sliding processing is performed on the original trajectory data obtained from the sensor.
[0091] The purpose of window sliding processing is to extract local features of time series data through window sliding technology and enhance the model's ability to understand and predict data.
[0092] The specific implementation method of the window sliding processing includes:
[0093] Set the window size: Select N time points. This means that each input sample will contain the trajectory raw data of N consecutive time points. Preferably, the trajectory raw data is position data. The trajectory raw data of N consecutive time points will form the time series data. Preferably, select 5 time points, then each input sample will contain the trajectory raw data of 5 consecutive time points.
[0094] Execute window sliding: The window starts from the starting point of the data sequence and slides at a certain step size (such as 1 time unit) until it reaches the end of the sequence. That is, the window starts from the current moment and slides forward until the start moment. Therefore, at the current moment, there will be multiple input samples input to the LSTM, and each input sample includes multiple window size values, for example, 5 values. Among them, the data sequence is the original trajectory data output by the sensor, and the current time is the starting point of the data sequence. The window slides forward, and each time it slides, a set of original trajectory data in the window is extracted as a time series sample to obtain multiple time series samples, that is, multiple sets of original trajectory data.
[0095] For example, if the position data is p(t) and the window size is 5, then the window slides to generate a time series sample of [p(t), p(t+1), p(t+2), p(t+3), p(t+4)]. By controlling the window sliding, multiple time series samples will be obtained, and each time series sample is a set of original trajectory data.
[0096] Then, step S202 is executed to input the multiple groups of original trajectory data processed by window sliding into a bidirectional neural network to obtain the trajectory observation data at the current moment predicted by the bidirectional neural network, and the bidirectional neural network is a bidirectional long short-term memory network.
[0097] Specifically, a bidirectional LSTM model is pre-established and trained.
[0098] In some embodiments, the bidirectional LSTM model structure includes:
[0099] Input layer: accepts time series samples generated by window sliding technology;
[0100] Bidirectional LSTM layer: A LSTM layer with 50 units in each direction, which can capture the forward and backward dependencies of time series data;
[0101] Output layer: A fully connected layer that outputs the predicted position information.
[0102] The training process of the bidirectional LSTM model includes:
[0103] Loss function: The mean squared error (MSE) is used to calculate the error between the predicted value and the actual value.
[0104] Optimizer: Use Adam optimizer for model training, which can automatically adjust the learning rate and speed up the training process.
[0105] Iterative training: Multiple rounds of iterations are performed until the model's performance on the validation set no longer improves to avoid overfitting.
[0106] After the training is completed, the bidirectional LSTM model is deployed. Then, multiple sets of trajectory raw data processed by window sliding are input into the trained bidirectional neural network.
[0107] Then, step S203 is executed to predict the prior estimation value of the trajectory data at the current moment based on the optimal estimation value of the trajectory data at the previous moment through unscented Kalman filtering, and predict the prior estimation covariance of the current moment based on the optimal estimation covariance at the previous moment.
[0108] And step S204, based on the prior estimated covariance at the current moment, the Kalman gain at the current moment is updated, based on the trajectory observation data at the current moment, the prior estimated value of the trajectory data at the current moment and the Kalman gain, the optimal estimated covariance at the current moment is updated based on the Kalman gain at the current moment and the prior estimated covariance.
[0109] Specifically, step S203 and step S204 are unscented Kalman filters (UKF).
[0110] The filtering process includes:
[0111] Data input: The output of the bidirectional LSTM model is used as the observation value of UKF.
[0112] State estimation and update: UKF is used to continuously update and correct the trajectory state estimate based on the prior estimate of the estimated data predicted by the model and the trajectory observations output by the bidirectional LSTM model.
[0113] Among them, UKF parameter settings include:
[0114] Process noise Q: controls the uncertainty in the prediction step of the model.
[0115] Measurement noise R: reflects the noise level of the observed data.
[0116] Initial state x_0 and covariance P_0: Set the initial state and uncertainty of the filter.
[0117] In one embodiment, the updating of the optimal estimate of the trajectory data at the current moment based on the trajectory observation data at the current moment, the prior estimate of the trajectory data at the current moment, and the Kalman gain includes:
[0118] The optimal estimate of the trajectory data at the current moment is updated to where x k|k is the optimal estimate of the trajectory data at the current moment, K is the prior estimate of the trajectory data at the current moment. k is the Kalman gain at the current moment, y kis the trajectory observation data at the current moment, and H is the observation matrix.
[0119] In one embodiment, the updating of the optimal estimated covariance at the current moment based on the Kalman gain at the current moment and the prior estimated covariance includes:
[0120] Update the optimal estimated covariance at the current moment to: Among them, P k|k is the optimal estimated covariance at the current moment, K k is the Kalman gain at the current moment, H is the observation matrix, is the prior estimated covariance at the current moment.
[0121] Specifically, the update formula is as follows:
[0122] Status Update:
[0123] Covariance update:
[0124] where x k|k is the optimal estimate of the trajectory data at the current time k, K is the prior estimate of the trajectory data at the current time k predicted based on the previous time k-1. k is the Kalman gain at the current moment, y k is the trajectory observation data at the current moment, H is the observation matrix, P k|* is the optimal estimated covariance at the current moment, P - k|k-1 is the prior estimated covariance at the current moment.
[0125] Prior state estimation
[0126] In the prediction stage of UKF, the Sigma point method is used to handle nonlinear state estimation:
[0127] 1.Sigma point generation:
[0128]
[0129] in, is the state estimate after the last update, is the covariance matrix at the previous moment.
[0130] 2. Status prediction:
[0131]
[0132] Among them, f is the state transfer function, which is used to transfer the Sigma point to the current moment.
[0133] 3. Prior state estimation:
[0134]
[0135] Among them, w i is the weight corresponding to the Sigma point.
[0136] 4. Prior covariance estimation:
[0137]
[0138] in, is the covariance weight, Q k is the process noise covariance matrix.
[0139] There are similarities and differences between the ordinary Kalman filter and the unscented Kalman filter:
[0140] There are differences in the prediction part:
[0141] Ordinary Kalman Filter (KF): In ordinary Kalman filtering, the prediction of system state is based on a linear model. In nonlinear systems, ordinary Kalman filtering requires linearization of the system, usually using Taylor expansion, which may lead to loss of accuracy.
[0142] Unscented Kalman Filter (UKF): UKF does not rely on linearized systems, but instead uses "sigma points" to propagate uncertainties in nonlinear systems. UKF can handle nonlinear problems more accurately and avoids errors caused by the linearization process. Therefore, UKF performs better when dealing with nonlinear dynamic systems.
[0143] There are similarities in the status update part:
[0144] In the state update step, the update formulas of the ordinary Kalman filter and the unscented Kalman filter are similar, and both use the Kalman gain to correct the predicted state. Specifically, they follow similar Kalman gain update, state estimate update, and covariance update processes when updating the state.
[0145] The main difference between the unscented Kalman filter (UKF) and the ordinary Kalman filter (KF) lies in the processing method of the prediction step. The UKF handles nonlinear systems by using sigma points, while the ordinary Kalman filter relies on linear approximation. In the state update step, the processes of the two are basically the same.
[0146] This embodiment applies bidirectional LSTM to trajectory prediction. By using a bidirectional LSTM network to process time series data, the data dependency before and after each time point is captured, which optimizes the limitation that the traditional unidirectional LSTM network can only consider past information. By introducing a bidirectional LSTM structure in trajectory prediction, the forward and backward dependencies of trajectory data can be learned at the same time, which improves the accuracy of prediction and the response speed to future data. At the same time, this embodiment combines window sliding technology with bidirectional LSTM, preprocesses trajectory data through window sliding technology, and provides a suitable input format for the bidirectional LSTM network, so that the model can more effectively understand and predict the local dynamics of the trajectory. In the data preprocessing stage, this embodiment applies window sliding technology to generate sequence data with time windows. These data are used to train the bidirectional LSTM model, optimize the data input structure, and improve processing efficiency and prediction performance. Finally, this embodiment integrates UKF to optimize the prediction results obtained from the bidirectional LSTM model, and improves the accuracy of trajectory estimation by effectively processing data under nonlinear and high noise conditions. This embodiment uses UKF to perform state estimation and noise elimination on the output of the bidirectional LSTM. This integration method improves the stability and reliability of the system in complex environments. This embodiment integrates the bidirectional LSTM network and UKF technology. The bidirectional LSTM provides accurate time series predictions, while the UKF is responsible for optimizing and calibrating these predictions to ensure the accuracy of the output trajectory data in highly dynamic and high-noise environments. This integration strategy not only enhances the robustness of the prediction model, but also provides a new solution for handling application scenarios that require extremely high accuracy and real-time performance.
[0147] like Figure 3 The schematic diagram of the effect of an example of the present invention is shown, including a real trajectory 31, an observed position trajectory 32, a bidirectional LSTM predicted position trajectory 33, and a UKF position estimation trajectory 34 estimated by the trajectory estimation method of this embodiment.
[0148] in, Figure 3 The meaning of the lines and points in the figure are as follows:
[0149] The true trajectory 31 (blue dashed line) is the actual trajectory of the vehicle, representing the target trajectory that the system hopes to track and optimize. The blue dashed line represents the trajectory of the system under ideal conditions without any noise or deviation.
[0150] The observation position trajectory 32 (red dots) represents the observation data directly obtained by the sensor, which usually contains certain noise. Figure 3 As shown, the observed position trajectory 32 fluctuates randomly around the true trajectory 31, reflecting the influence of sensor noise. The deviation of the observed position 32 (red dot) from the true position 31 reveals the limitations of relying solely on sensor data.
[0151] Bidirectional LSTM predicted position trajectory 33 (green solid line) represents the prediction made by bidirectional LSTM based on historical observation data. LSTM prediction can capture the context information in the time series and generate a relatively smooth trajectory. Figure 3 As shown in Figure 3, the prediction of LSTM still has certain deviations, especially at the turning points of the trajectory or where the fluctuation is large (e.g., the parts with higher or lower positions), which shows that relying solely on LSTM cannot eliminate all noise and errors.
[0152] The UKF position estimation trajectory 34 (blue solid line) of this embodiment represents the trajectory after the unscented Kalman filter (UKF) corrects the LSTM prediction result. UKF combines the observed data and the LSTM prediction data, and obtains an optimal estimation trajectory that is closer to the actual position by processing nonlinear dynamics. It can be seen that the UKF position estimation trajectory 34 is closer to the actual trajectory 31, indicating that UKF plays a role in reducing the impact of noise and improving trajectory accuracy.
[0153] The trajectory estimation method provided in this embodiment is a trajectory optimization method based on bidirectional LSTM and unscented Kalman filter (UKF), which can significantly improve the accuracy and robustness of trajectory estimation, especially when processing trajectory data containing high noise and nonlinear dynamics. By utilizing the bidirectional LSTM structure, this embodiment can make comprehensive use of historical and future data information to predict the current state. This comprehensive use of contextual information makes the model more accurate in predicting time series data. At the same time, as an efficient nonlinear filtering method, the unscented Kalman filter avoids the linearization error in the traditional Kalman filter by using Sigma points to approximate the prediction and update process. This makes it possible for UKF to effectively estimate a more accurate state in practical applications even in the face of complex dynamic changes and severe measurement noise. This embodiment combines the advantages of LSTM and UKF. Among them, LSTM is good at processing and predicting time series data, while UKF is good at state estimation and noise processing. Combining the two not only enhances the system's resistance to noise, but also makes the prediction model insensitive to small changes in parameters, so that it can still maintain a high level of performance in a changing actual environment. From Figure 3 It can be seen that the UKF position estimation trajectory 34 estimated by the trajectory estimation method of this embodiment is very close to the real trajectory 31, and can maintain a high prediction accuracy even in a noisy environment. Finally, this embodiment applies the window sliding technology. The local features extracted by the window technology enable the model to capture the short-term change rules of the time series data more carefully, which is particularly critical for improving the prediction accuracy, especially in scenarios where the trajectory data often changes rapidly.
[0154] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.
[0155] like Figure 4 The figure shows a hardware structure diagram of an electronic device of the present invention, including:
[0156] at least one processor 401; and,
[0157] A memory 402 is communicatively connected to at least one of the processors 401; wherein,
[0158] The memory 402 stores instructions that can be executed by at least one of the processors. The instructions are executed by at least one of the processors so that the at least one of the processors can perform the trajectory estimation method as described above.
[0159] Figure 4 A processor 401 is taken as an example.
[0160] The electronic device may further include: an input device 403 and a display device 404 .
[0161] The processor 401, the memory 402, the input device 403 and the display device 404 may be connected via a bus or other means, and the figure takes the connection via a bus as an example.
[0162] The memory 402 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions / modules corresponding to the trajectory estimation method in the embodiment of the present application, for example, Figure 1 , Figure 2 The processor 401 executes various functional applications and data processing by running the non-volatile software programs, instructions and modules stored in the memory 402, that is, implements the trajectory estimation method in the above embodiment.
[0163] The memory 402 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required by at least one function; the data storage area may store data created according to the use of the trajectory estimation method, etc. In addition, the memory 402 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 402 may optionally include a memory remotely arranged relative to the processor 401, and these remote memories may be connected to the device for executing the trajectory estimation method via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.
[0164] The input device 403 can receive user clicks and generate signal inputs related to user settings and function control of the trajectory estimation method. The display device 404 can include display devices such as a display screen.
[0165] The one or more modules are stored in the memory 402 , and when executed by the one or more processors 401 , the trajectory estimation method in any of the above method embodiments is executed.
[0166] The present invention predicts the trajectory observation data at the current moment through a bidirectional neural network, which can effectively capture the contextual information of vehicle dynamics, improve the accuracy of trajectory prediction, and is suitable for dealing with nonlinear problems of vehicle movement in traffic environments. At the same time, through nonlinear filtering, the observation data and the predicted data are combined, which can more accurately handle sensor noise and GPS signal interference, and improve the accuracy and robustness of position estimation. The present invention combines bidirectional neural networks and nonlinear filtering to estimate trajectories, which not only improves the accuracy of data processing, but also reduces the consumption of computing resources through optimization algorithms, achieves faster processing speeds, and meets the needs of real-time vehicle-road-cloud communications.
[0167] An embodiment of the present invention provides a storage medium, wherein the storage medium stores computer instructions. When a computer executes the computer instructions, the computer instructions are used to execute all steps of the trajectory estimation method as described above.
[0168] In the context of the present disclosure, a storage medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. The storage medium may be a machine-readable signal medium or a machine-readable storage medium. Alternatively, the storage medium may be a non-temporary computer-readable storage medium, for example, a non-temporary computer-readable storage medium may be a ROM, a random access memory (Random Access Memory, RAM), a compact disc read-only memory (Compact Disc ROM, CD-ROM), a magnetic tape, a floppy disk, and an optical data storage device, etc.
[0169] An embodiment of the present invention provides a computer program product, including a computer program / instruction, which implements the trajectory estimation method as described above when executed by a processor.
[0170] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.
Claims
1. A trajectory estimation method, characterized in that: include: Obtain the original trajectory data from the sensor and predict the trajectory observation data at the current moment through a bidirectional neural network; Through nonlinear filtering, based on the optimal estimate of the trajectory data at the previous moment, the prior estimate of the trajectory data at the current moment is predicted, and based on the optimal estimate covariance at the previous moment, the prior estimate covariance at the current moment is predicted; Based on the prior estimated covariance at the current moment, the Kalman gain at the current moment is updated; based on the trajectory observation data at the current moment, the prior estimated value of the trajectory data at the current moment and the Kalman gain, the optimal estimated value of the trajectory data at the current moment is updated; based on the Kalman gain at the current moment and the prior estimated covariance, the optimal estimated covariance at the current moment is updated.
2. The trajectory estimation method according to claim 1, characterized in that: The method of obtaining the original trajectory data from the sensor and predicting the trajectory observation data at the current moment through a bidirectional neural network includes: Acquire multiple sets of trajectory raw data from the sensor through window sliding processing; Multiple groups of original trajectory data processed by window sliding are input into a bidirectional neural network to obtain trajectory observation data at the current moment predicted by the bidirectional neural network.
3. The trajectory estimation method according to claim 2, characterized in that: The method of acquiring multiple groups of original trajectory data from the sensor through window sliding processing includes: Obtain the time series data consisting of all the original trajectory data of the sensor from the current time to the starting time; A window is controlled from the time series data to slide from the current time to the starting time in sequence according to a preset step length, and each time the window slides, the original trajectory data located in the window is extracted from the time series data as a group of original trajectory data.
4. The trajectory estimation method according to claim 1, characterized in that: The bidirectional neural network is a bidirectional long short-term memory network.
5. The trajectory estimation method according to claim 1, characterized in that: The method of predicting the prior estimation value of the trajectory data at the current moment based on the optimal estimation value of the trajectory data at the previous moment by nonlinear filtering, and predicting the prior estimation covariance at the current moment based on the optimal estimation covariance at the previous moment, includes: Through unscented Kalman filtering, the prior estimate of the trajectory data at the current moment is predicted based on the optimal estimate of the trajectory data at the previous moment, and the prior estimate covariance at the current moment is predicted based on the optimal estimate covariance at the previous moment.
6. The trajectory estimation method according to claim 1, characterized in that: The method of updating the optimal estimate of the trajectory data at the current moment based on the trajectory observation data at the current moment, the prior estimate of the trajectory data at the current moment, and the Kalman gain includes: The optimal estimate of the trajectory data at the current moment is updated to where x k|k is the optimal estimate of the trajectory data at the current moment, K is the prior estimate of the trajectory data at the current moment. k is the Kalman gain at the current moment, y k is the trajectory observation data at the current moment, and H is the observation matrix.
7. The trajectory estimation method according to claim 1, characterized in that: The updating of the optimal estimated covariance at the current moment based on the Kalman gain at the current moment and the prior estimated covariance includes: Update the optimal estimated covariance at the current moment to: Among them, P k|k is the optimal estimated covariance at the current moment, K k is the Kalman gain at the current moment, H is the observation matrix, is the prior estimated covariance at the current moment.
8. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by at least one of the processors, and the instructions are executed by at least one of the processors to enable at least one of the processors to perform the trajectory estimation method according to any one of claims 1 to 7.
9. A storage medium, characterized in that: The storage medium stores computer instructions, and when a computer executes the computer instructions, it is used to execute all steps of the trajectory estimation method according to any one of claims 1 to 7.
10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the trajectory estimation method according to any one of claims 1 to 7 is implemented.