Low-speed starting obstacle trajectory prediction method

Abstract

The invention provides a low-speed starting obstacle trajectory prediction method. And accurate prediction of the low-speed obstacle trajectory is realized through a staged processing mechanism. Specifically, the method comprises the following steps: acquiring obstacle / vehicle state information to provide basic data input for track prediction; spatial constraints such as lane topology are introduced into associated environment information acquisition, so that the prediction reasonability is enhanced; the obstacle state is processed by a forward algorithm, so that the interference of sensor noise on a low-speed motion signal can be effectively suppressed, and the problem of track missing report caused by low signal-to-noise ratio in the starting stage is particularly solved; the noise reduction data and the environment information are fused to predict the trajectory, so that the deep learning model gets rid of the noise stripping burden and focuses on complex interaction relation modeling; a backward algorithm corrects a prediction result, track abrupt change is eliminated through reverse time sequence information transmission, and high-order kinematics consistency of the position, the speed and the acceleration is ensured. And finally, on the premise of guaranteeing the prediction timeliness, the response accuracy of the automatic driving system to the starting obstacle is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of autonomous driving, and in particular to a low-speed start obstacle trajectory prediction method. BACKGROUND

[0002] When an autonomous vehicle drives in an open road environment, it often encounters scenarios of interacting with low-speed start obstacles. The to-be-predicted traffic participants are in a state of static to low-speed motion in a short time, and interact with the autonomous vehicle. Since trajectory prediction relies on sensor perception of obstacles and the surrounding environment, and there are errors and noises in the sensor itself and the model output of the autonomous driving perception module, when the obstacle starts at low speed, the noise and the true signal often overlap together, that is, the signal-to-noise ratio is low. For the above reasons, at present, when encountering such a situation, the trajectory prediction result cannot be output in time for the low-speed start obstacle, thereby affecting the autonomous driving system to make a yielding decision for such obstacles.

[0003] In some feasible embodiments, although the historical state of the target obstacle and rich surrounding environment information are considered, the problem of large overlap between the true motion signal and the noise of the low-speed moving obstacle is ignored. When the real physical world data is input into the model for training, the model can learn to some extent how to overcome the influence of noise, but the model cannot guarantee that it can reasonably strip the noise for all sample data. In addition, the low-speed start scenario often accounts for a small proportion in the training set, and such data also belongs to small sample data. The deep learning model is trained based on the gradient method, which mainly ensures that the inference result of the model meets the expectation in most common scenarios, so it is difficult to balance the performance in most scenario data and low-speed start small sample data. Therefore, the timeliness of the trajectory prediction output is poor, which limits the intelligent level of autonomous driving. Thus, a low-speed start obstacle trajectory prediction method is needed to solve the corresponding problems. SUMMARY

[0004] The purpose of the present application is to at least solve one of the above technical defects, in particular the technical defect that the timeliness and effectiveness of the trajectory prediction output in the prior art are poor.

[0005] In a first aspect, the present application provides a low-speed start obstacle trajectory prediction method, comprising:

[0006] obstacle state information and ego vehicle state information are obtained;

[0007] environment information is obtained according to the obstacle state information and the ego vehicle state information;

[0008] The obstacle state information is processed by a forward algorithm to obtain historical position estimation information;

[0009] Obtaining a trajectory prediction result based on the historical position estimation information and the environmental information;

[0010] The trajectory prediction result is processed and corrected by a backward algorithm to obtain the target trajectory.

[0011] As an optional implementation, the method further includes:

[0012] Preprocessing the obstacle status information;

[0013] The preprocessing specifically includes:

[0014] Processing the obstacle status information according to a preset data structure to establish a mapping relationship between the obstacle status information and the obstacle identity;

[0015] Furthermore, linear interpolation is performed on the obstacle state information based on a time dimension to align the time sequence information corresponding to each piece of obstacle state information.

[0016] As an optional implementation, the environmental information includes a lane centerline, a lane boundary line, a lane type, and a lane topological connection relationship. After obtaining the environmental information based on the obstacle status information and the vehicle status information, the method further includes:

[0017] Determining, based on the environmental information, correlation parameters between each obstacle state information and the vehicle state information;

[0018] The obstacle status information is filtered through the associated parameters and the obstacle status information is updated.

[0019] As an optional implementation manner, the forward algorithm and the backward algorithm include a Kalman filter algorithm, and the method further includes:

[0020] Perform kinematic modeling based on the position information, velocity information, and acceleration information corresponding to the obstacle state information to determine the state matrix and the observation matrix;

[0021] The value of the process noise covariance matrix is ​​initialized, and noise information of the position observation and the velocity observation is determined according to the obstacle state information to obtain the observation noise covariance matrix.

[0022] As an optional implementation manner, processing the obstacle state information by a forward algorithm to obtain historical position estimation information includes:

[0023] The historical real state sequence corresponding to the obstacle information, and the state matrix, the observation matrix, the process noise covariance matrix and the observation noise covariance matrix are taken as inputs to iteratively perform Kalman filtering to obtain Kalman coefficients, an iterative state, an estimation covariance and a prediction covariance at each step, and the iterative state and the estimation covariance are taken as historical position estimation information.

[0024] As an optional implementation, the target trajectory is obtained by processing and correcting the trajectory prediction result through a backward algorithm, and the target trajectory comprises:

[0025] The target trajectory is obtained by taking the iterative state and the estimation covariance in the last step of iteration in the trajectory prediction result as inputs, and iteratively correcting the historical position estimation information obtained by the forward algorithm through a smoothing gain matrix.

[0026] The smoothing gain matrix is used to indicate the influence of historical estimation errors on future prediction errors, so as to constrain the historical position estimation information obtained in each step of iteration.

[0027] As an optional implementation, the trajectory prediction result is obtained according to the historical position estimation information and the environment information, and the trajectory prediction result comprises:

[0028] The historical position estimation information and the environment information are input into a preset encoder for encoding to obtain spatial position features, and the spatial position features are input into a decoder to obtain the trajectory prediction result.

[0029] The decoder is used to calculate the correlation information between the obstacles and the correlation information between the obstacles and the environment, so as to infer the trajectory prediction result of the obstacles.

[0030] In a second aspect, the present application provides a low-speed starting obstacle trajectory prediction device, comprising:

[0031] An acquisition module is configured to acquire obstacle state information and ego vehicle state information.

[0032] A processing module is configured to acquire environment information according to the obstacle state information and the ego vehicle state information.

[0033] The processing module is further configured to process the obstacle state information through a forward algorithm to obtain historical position estimation information.

[0034] The processing module is further configured to obtain a trajectory prediction result according to the historical position estimation information and the environment information.

[0035] The processing module is further configured to obtain a target trajectory by processing and correcting the trajectory prediction result through a backward algorithm.

[0036] In a third aspect, the present application provides a computer device, comprising one or more processors, and a memory, wherein the memory stores computer readable instructions, and the computer readable instructions are executed by the one or more processors to perform the steps of the method according to the first aspect.

[0037] In a fourth aspect, the present application provides a storage medium, wherein the storage medium stores computer readable instructions, and the computer readable instructions are executed by one or more processors to perform the steps of the method according to the first aspect.

[0038] From the above technical solutions, the embodiments of the present application have the following advantages:

[0039] Based on any of the above embodiments, the present application improves the accuracy and timeliness of low-speed obstacle trajectory prediction through a multi-stage cooperative processing mechanism. First, the original obstacle state containing noise is time-series filtered through a forward algorithm, effectively removing the observation noise in the motion signal. Especially for the low signal-to-noise ratio characteristic of the low-speed starting stage, the trajectory prediction delay problem is solved. Second, deep learning prediction is performed in combination with environmental information and denoised historical position estimation, so that the model focuses on fitting the nonlinear relationship of complex scenes rather than noise processing, improving the prediction reliability. Finally, the kinematic consistency of the predicted trajectory is corrected through a backward algorithm, and the state jump is eliminated by using the bidirectional information flow of the time series, thereby ensuring the self-consistency of the high-order parameters of the trajectory. The above process not only reduces the difficulty of training the deep learning model, but also significantly improves the timeliness and safety of the decision-making of the autonomous driving system in the starting obstacle scene. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor.

[0041] Figure 1 A flowchart of a low-speed starting obstacle trajectory prediction method provided by an embodiment of the present application is shown in the figure.

[0042] Figure 2 An internal structure diagram of a computer device provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0044] When autonomous vehicles are driving on open roads, they often encounter scenarios where they interact with obstacles that start at low speeds. For example, vehicles that are parking in or out of the roadside, vehicles that suddenly change lanes while waiting in line at the lights at an intersection, pedestrians that suddenly start at the roadside, etc. The common feature of these scenarios is that the traffic participants to be predicted all move from stationary to low-speed in a short period of time, interacting with the autonomous vehicle. Since trajectory prediction relies on the sensor's perception of obstacles and the surrounding environment, and the sensor itself, as well as the model output of the autonomous driving perception module, have errors and noise, when an obstacle starts at low speed, the noise and the real signal often overlap, that is, the signal-to-noise ratio is low. For the above reasons, when encountering such a situation, the existing technology often cannot output trajectory prediction results for obstacles that start at low speeds in a timely manner, thereby affecting the autonomous driving system's decision to give way to such obstacles. The present invention mainly solves the problem of trajectory prediction for such obstacles that start at low speeds.

[0045] Trajectory prediction is a crucial component in autonomous driving. Its upstream component is environmental perception, responsible for acquiring information about obstacles and the surrounding environment, including their location, speed, type, semantic labels, and traffic light status. Its downstream component is the decision-making and planning module, responsible for making targeted decisions based on the predicted obstacle trajectory (including maneuvers such as detouring, yielding, following, overtaking, and lane changing). Based on these decisions, the module constructs a spatiotemporal semantic corridor, within which it generates the planned trajectory for the autonomous vehicle over the next period of time.

[0046] Existing trajectory prediction technologies fall into two main categories. One is based on kinematic models, generating obstacle trajectory predictions under certain assumptions (constant speed, constant curvature). The other is based on deep learning methods. This encodes historical observations of the obstacle's state and surrounding environment, inputs them into a deep learning model, and infers the obstacle's trajectory. In actual autonomous driving systems, these two approaches complement each other.

[0047] One feasible implementation utilizes a multi-target obstacle trajectory prediction network. This network comprises an encoder and a decoder. The encoder determines scene features based on the historical states of each obstacle within a preset range, lane information within the preset range, and traffic light information within the preset range. The historical states include the speed and position of the obstacle at each sampling time prior to the current moment. The lane information includes the position, type, and heading angle of the lane centerline node, as well as the positions of the left and right lane boundaries. The traffic light information includes the position of the traffic light and its lighting information at each sampling time prior to the current moment. The decoder determines the predicted trajectory of the target obstacle at a future moment based on the scene features and the initial anchor point of the target obstacle. This achieves precise prediction of the trajectory of multiple target obstacles, improving prediction accuracy.

[0048] This feasible implementation utilizes a deep learning approach to construct a multi-target trajectory prediction network. While this approach considers the historical state of target obstacles and rich surrounding environmental information, it ignores the significant overlap between the actual motion signals and noise of slow-moving obstacles. When real-world data (with noise) is fed into the model for training, the model can learn to overcome the effects of noise to a certain extent, but it cannot guarantee that noise removal is adequate for all sample data. Furthermore, slow-speed start scenarios often account for a small proportion of the training set. Similar to driving, encountering slow-speed starts is relatively rare. Therefore, this type of data is considered small sample data. Deep learning models are trained using gradient methods, which primarily ensure that the model's inference results meet expectations in the most common scenarios. Given these facts, this feasible implementation struggles to strike a balance between data from most scenarios and small sample data from slow-speed starts, ensuring optimal performance.

[0049] To address the above issues, this application, starting from a kinematic model, uses the historical states of the target obstacle output by the perception module as observation input within the framework of Kalman smoothing, and fuses these observations to obtain a true estimate of its historical motion state. Kalman filtering is a highly efficient recursive filter that can estimate the state of a dynamic system from a series of incomplete and noisy measurements. Based on the values ​​of each observation at different times, the Kalman filter considers the joint distribution at each time and then generates an estimate of the unknown variable. The Kalman smoothing employed in this invention performs a temporal reversal based on the Kalman filter, using the states in the target obstacle's historical state sequence that are closer to the current moment to further update the previous state. This fully utilizes the information of the entire time series to estimate the true state. Inputting this result into a deep learning model can greatly reduce the model's burden. The model no longer needs to have a strong ability to derive the true signal from a noisy signal, which is particularly important in low-speed starting scenarios with low signal-to-noise ratios.

[0050] In summary, the technical concept of the present application is to improve the accuracy and timeliness of low-speed obstacle trajectory prediction through a multi-stage cooperative processing mechanism. First, the original obstacle state containing noise is filtered in time sequence through a forward algorithm, effectively removing the observation noise in the motion signal. Especially for the low signal-to-noise ratio characteristic of the low-speed starting stage, the problem of trajectory prediction delay is solved. Second, deep learning prediction is performed in combination with environmental information and denoised historical position estimation, so that the model focuses on fitting the nonlinear relationship of complex scenes rather than noise processing, improving the prediction reliability. Finally, the kinematic consistency of the predicted trajectory is corrected through a backward algorithm, using the bidirectional information flow of the time sequence to eliminate state jumps and ensure the self-consistency of the high-order parameters of the trajectory. The above process not only reduces the difficulty of deep learning model training, but also significantly improves the timeliness and safety of the decision-making of the autonomous driving system in the starting obstacle scene.

[0051] The method provided by the present application will be described in detail below according to the corresponding embodiments in some practical application scenarios.

[0052] Figure 1 The flowchart of the low-speed starting obstacle trajectory prediction method provided by an embodiment of the present application is shown in FIG. 1, and the present application provides a low-speed starting obstacle trajectory prediction method, which comprises the following steps. Figure 1

[0053] S101, obtaining obstacle state information and ego vehicle state information;

[0054] S102, obtaining environmental information according to the obstacle state information and the ego vehicle state information;

[0055] S103, processing the obstacle state information through a forward algorithm to obtain historical position estimation information;

[0056] S104, obtaining a trajectory prediction result according to the historical position estimation information and the environmental information;

[0057] S105, processing and correcting the trajectory prediction result through a backward algorithm to obtain a target trajectory.

[0058] ​The main technical route of the present application is to first obtain obstacle state information (and further obtain landmark or traffic light information) and self-vehicle state information, so that the relevant information in the high-precision map can be matched to obtain environmental information. First, the historical position estimation information is obtained through the forward algorithm step by step, and the forward trajectory prediction result can be obtained in combination with the environmental information, but the forward prediction result at this time may not be accurate enough, so the trajectory prediction result can be modified according to the state of each time point in the trajectory prediction result, step by step, and the physical constraint is reversed to obtain the target trajectory. Each specific implementation manner can be referred to the related description.

[0059] The embodiment realizes accurate prediction of low-speed obstacle trajectory through a staged processing mechanism. Specifically, the obstacle / self-vehicle state information provides basic data input for trajectory prediction; the associated environmental information introduces lane topology and other spatial constraints to enhance the rationality of the prediction; the forward algorithm processing of the obstacle state can effectively suppress the interference of sensor noise on low-speed motion signals, especially solve the trajectory missing report problem caused by low signal-to-noise ratio at the starting stage; the fusion of denoising data and environmental information prediction trajectory makes the deep learning model get rid of the noise stripping burden and focus on complex interaction modeling; the backward algorithm modifies the prediction result to eliminate trajectory mutation through reverse time sequence information transmission and ensure the high-order kinematics consistency of position, velocity and acceleration. Finally, under the premise of ensuring the timeliness of the prediction, the response accuracy of the automatic driving system to the starting obstacle is improved.

[0060] As an optional implementation manner, the method further comprises:

[0061] preprocessing the obstacle state information;

[0062] The preprocessing specifically comprises:

[0063] processing the obstacle state information according to a preset data structure to establish a mapping relationship between the obstacle state information and the obstacle identity;

[0064] and performing linear interpolation on the obstacle state information based on the time dimension to align the time sequence information corresponding to each of the obstacle state information.

[0065] The embodiment improves the data reliability through structured preprocessing. Establishing the mapping relationship of the obstacle identity ensures the independent tracking ability of the multi-target trajectory and avoids prediction errors caused by identity confusion; the time sequence information alignment processing eliminates the influence of the sensor sampling frequency difference on the state sequence and provides input data with uniform time intervals for the forward algorithm. Both of them solve the fragmentation problem of the original observation data and establish an accurate data basis for subsequent noise reduction and prediction.

[0066] As an optional implementation, the environment information comprises a lane center line, a lane boundary line, a lane type, and a lane topology connection relationship, and after the environment information is acquired according to the obstacle state information and the ego vehicle state information, the method further comprises:

[0067] According to the environment information, a correlation parameter between each of the obstacle state information and the ego vehicle state information is determined;

[0068] The obstacle state information is filtered and updated through the correlation parameter.

[0069] The correlation parameter can be used to indicate an obstacle in each of the obstacle state information that is related to the current vehicle moving track, for example, an obstacle that moves in the opposite direction, is fixed for a long time, is physically isolated by a fence or other device, and thus does not affect the current ego vehicle moving track, can be regarded as an irrelevant obstacle, and the state information of the obstacle can not be considered when the track is generated.

[0070] The embodiment realizes target intelligent filtering through environment correlation. The correlation parameter is calculated according to the lane type and the topology relationship to accurately identify an obstacle (such as a vehicle in a neighboring lane) that has a possibility of interaction with the ego vehicle; the obstacle state is updated and irrelevant targets (such as obstacles in the opposite lane) are removed based on the correlation parameter, which significantly reduces the calculation complexity of the deep learning model. While ensuring the prediction accuracy of key obstacles, the real-time performance of the system is improved.

[0071] As an optional implementation, the forward algorithm and the backward algorithm comprise a Kalman filter algorithm.

[0072] The Kalman filter itself only comprises a forward algorithm, that is, only a forward iteration operation is performed, and the function thereof is to fully utilize the state information of the previous state when estimating the current state. For example, the estimation of the nth state utilizes the information of the previous n-1 states. This is because the change in the real world is continuous and is subject to physical and kinematic constraints, for example, a person cannot be static at one time and be observed to have a position change of 10 meters after 0.1 seconds, which violates the physics, and thus the observation is largely incorrect and has a large error, and is filtered out by the Kalman filter. The backward algorithm is a reverse operation that transmits the information of the 2nd, 3rd,..., nth state to the 1st state, so that the information in the entire historical state sequence can be circulated in both directions. For example, if a person changes his position by 10 meters after 1 second, he should also be moving 1 second ago).

[0073] It should be noted that this application uses Kalman smoothing to address the low signal-to-noise ratio issue in the low-speed obstacle start scenario. Possible alternatives are to use other methods to address this issue, such as other filtering methods, such as nonlinear filtering, particle filtering, and deep learning models specifically trained for noise removal, to replace filtering methods.

[0074] This application applies a Kalman smoothing algorithm to post-process the results of deep learning methods to ensure the self-consistency of high-order trajectory information. A possible alternative is to apply a control algorithm to track the results of deep learning methods to obtain high-order trajectory information, such as the LQR (Linear Regulator) algorithm.

[0075] The method further comprises:

[0076] Perform kinematic modeling based on the position information, velocity information, and acceleration information corresponding to the obstacle state information to determine the state matrix and the observation matrix;

[0077] The value of the process noise covariance matrix is ​​initialized, and noise information of the position observation and the velocity observation is determined according to the obstacle state information to obtain the observation noise covariance matrix.

[0078] First, the obstacle is kinematically modeled. Its state includes position, velocity, and acceleration. Considering the relationship between the three, the state matrix and observation matrix can be in the form of:

[0079]

[0080]

[0081] Where dt is the time interval between states, in seconds.

[0082] From the obstacle position information data output by perception, the noise information of its position observation and the noise information of its speed observation are mined to obtain its observation noise covariance matrix R.

[0083] Since the dt interval is very short, the kinematic model can be considered to be sufficiently fidelity, so the elements in the process noise covariance matrix Q of the state recursion can take very small values, such as 0.001.

[0084] Thus, initialization is completed.

[0085] This implementation optimizes filtering through kinematic modeling. A state matrix based on position, velocity, and acceleration is constructed to establish physically correct motion constraints (such as acceleration continuity). Initialization uses noise covariance to control the model's sensitivity to sudden changes in state. Finally, noise covariance is determined based on observed data to achieve adaptive noise reduction intensity. These three factors work together to ensure the forward algorithm effectively filters observation jitter in low-speed scenes while preserving true motion trends.

[0086] As an optional implementation manner, the processing of the obstacle state information by a forward algorithm to obtain historical position estimation information includes:

[0087] The historical real state sequence corresponding to the obstacle information, as well as the state matrix, the observation matrix, the process noise covariance matrix and the observation noise covariance matrix are taken as input, and Kalman filtering is iteratively performed to obtain the Kalman coefficient, iterative state, estimated covariance and predicted covariance of each step, and the iterative state and estimated covariance are used as historical position estimation information.

[0088] According to the historical real state sequence corresponding to the obstacle information , , ,..., , and model parameters F, Q, H, R, input Kalman filter model, K represents Kalman filter parameters, P represents prediction covariance, subscript can represent the number of iterations, 0 is the initial state, and recursion is performed in the following form:

[0089]

[0090] in, , is the initial state and its estimated covariance, which can be taken as the initial value of the state sequence and its noise estimate in turn.

[0091] After obtaining the first state, execute the following formula to continue iterating until an estimate of the entire state sequence is obtained:

[0092]

[0093] The subscript j represents the index in the state sequence, starting from 2 and going up to n, where n is the length of the state sequence.

[0094] At this point, the Kalman filter (forward algorithm) obtains a forward estimate of the historical state sequence.

[0095] This implementation achieves progressive state optimization through an iterative mechanism. The Kalman filter's step-by-step iterative calculations utilize continuity constraints on the historical state sequence (for example, the previous position constrains the current displacement range) to correct observation errors frame by frame. The output iterative state and covariance provide a position estimate with confidence assessment, providing a basis for the deep learning model to distinguish high- and low-reliability features. This ultimately improves the robustness of the noise reduction process to brief signal loss.

[0096] As an optional implementation, the step of processing and correcting the trajectory prediction result by a backward algorithm to obtain the target trajectory includes:

[0097] According to the iteration state and estimated covariance in the last iteration corresponding to the trajectory prediction result, the historical position estimation information obtained by the forward algorithm iteration is corrected by reverse iteration through a smoothing gain matrix to obtain the target trajectory;

[0098] Among them, the smoothed gain matrix, that is, C in the following formula, is used to indicate the impact of historical estimation errors on future prediction errors, so as to constrain the historical position estimation information obtained in each iteration. Similarly, the subscript represents the number of iterations. Since it is a reverse iteration, each step is calculated based on the next state corresponding to the result obtained in the forward iteration.

[0099]

[0100] This implementation ensures trajectory continuity through reverse correction. Using a smoothed gain matrix, reverse iteration propagates high-confidence states from later moments back to earlier moments (e.g., current velocity is affected by future acceleration), thus overcoming the one-way information limitations of the forward algorithm. Constraining the propagation of historical estimation errors suppresses inter-frame jumps in the predicted trajectory (e.g., sudden position changes or discontinuous velocity). This enhances kinematic rationality based on deep learning trajectory prediction, meeting the smooth trajectory requirements of the decision-making and planning module.

[0101] As an optional implementation manner, obtaining a trajectory prediction result based on the historical position estimation information and the environmental information includes:

[0102] Inputting the historical position estimation information and the environmental information into a preset encoder for encoding to obtain spatial position features, and inputting the spatial position features into a decoder to obtain a trajectory prediction result;

[0103] The decoder is used to calculate the association information between obstacles and the direct association information between obstacles and the environment, so as to infer the trajectory prediction results of the obstacles.

[0104] The embodiment fuses multi-dimensional information through an encoding-decoding architecture. The encoder integrates the noise reduction position and the environment feature extraction to contain the spatial position features of the spatial-temporal relationship (such as the offset of the obstacle relative to the lane center line); the decoder calculates the obstacle-environment interaction correlation (such as the following relationship between vehicles or the position constraint of the pedestrian-road intersection) to infer the trajectory distribution conforming to the physical rules. The traditional filtering method is complementary to the deep learning advantage to output the prediction result with accuracy and explainability in the complex scene.

[0105] When the embodiments are combined with the application, the main part of the application can include:

[0106] Step one: obtain the obstacle state information output by the automatic driving perception module and the traffic light information.

[0107] Step two: obtain the ego vehicle state information output by the automatic driving positioning module.

[0108] Step three: according to the positions of the obstacles and the ego vehicle obtained in the previous two steps, query the surrounding environment information from the high-definition map, including the lane center line, the lane boundary line, the lane type (straight, left turn, right turn, U-turn, left turn waiting, etc.), and the lane topology connection relationship (neighbor lane information).

[0109] Step four: fill the state information of the obstacles into a container that can be queried by the obstacle ID.

[0110] Step five: linearly interpolate the state information of the obstacles according to the time dimension to ensure that the states of each obstacle are aligned on the time axis and have the same time interval.

[0111] Step six: according to the environment information, remove the obstacles irrelevant to the ego vehicle, and only keep the vehicles, pedestrians, and non-motor vehicles of high priority that need to be trajectory predicted.

[0112] Step seven: input the historical sequence of the state information of the obstacles into the Kalman smoothing algorithm to obtain the estimation of the historical real positions of the obstacles.

[0113] Step eight: input the estimation of the historical real positions of the obstacles obtained in step seven and the environment information obtained in step three into the encoder of the model for encoding.

[0114] Step nine: input the result of step eight into the model decoder to infer the trajectory prediction result of the obstacles by calculating the correlation information between the obstacles and the environment, and the direct correlation information between the obstacles and the environment. The result contains the motion trajectory of the obstacles in the future 6 seconds.

[0115] Step ten: input the predicted trajectory information of the obstacles obtained in step nine into the Kalman smoothing algorithm again to obtain self-consistent and high-order information, including the speed and acceleration of the obstacles in the future 6 seconds.

[0116] In summary, this application combines the Kalman smoothing algorithm and deep learning model methods commonly used in signal processing. Starting from the kinematic model followed by the real physical world, it solves the problems of low signal-to-noise ratio, difficult model learning, and untimely trajectory prediction in the low-speed starting scene of obstacles. The Kalman smoothing algorithm is good at obtaining an estimate of the real state from measurement information containing noise, and the deep learning method has advantages in fitting the nonlinear physical world. The combination of the two takes advantage of each other. This application also uses the Kalman smoothing algorithm to post-process the predicted trajectory output by the deep learning method to obtain self-consistent, high-order trajectory information. It is convenient for the use of the autonomous driving decision-making and planning module.

[0117] Existing implementation methods do not make special treatments for scenarios where obstacles start at low speeds, and it is difficult to ensure that the deep learning model can give timely trajectory prediction results in such scenarios with low signal-to-noise ratios, which often leads to untimely braking of the vehicle, resulting in traffic accidents. The present application uses the Kalman smoothing algorithm to post-process the trajectory output by the deep learning method to ensure the self-consistency of the high-order information of the trajectory. The present application also helps to ensure the inter-frame consistency of the obstacle trajectory prediction, and no sudden changes in the trajectory prediction will occur. The present application has low difficulty in on-vehicle deployment, the algorithm is simple and easy to use, and the solution is extremely efficient. It takes less than 5ms to process the historical state sequence of hundreds of obstacles at a time on the on-board chip, and it has extremely strong practicality.

[0118] The present application also provides a low-speed starting obstacle trajectory prediction device, comprising:

[0119] Acquisition module, used to obtain obstacle status information and vehicle status information;

[0120] A processing module, configured to obtain environmental information based on the obstacle status information and the vehicle status information;

[0121] The processing module is further configured to process the obstacle state information using a forward algorithm to obtain historical position estimation information;

[0122] The processing module is further configured to obtain a trajectory prediction result based on the historical position estimation information and the environmental information;

[0123] The processing module is further configured to process and correct the trajectory prediction result through a backward algorithm to obtain a target trajectory.

[0124] The embodiment realizes accurate prediction of low-speed obstacle trajectory through a staged processing mechanism. Specifically, the obstacle / ego vehicle state information is obtained to provide basic data input for trajectory prediction; the associated environment information is obtained to introduce lane topology and other spatial constraints, thereby enhancing the rationality of the prediction; the forward algorithm processes the obstacle state to effectively suppress the interference of sensor noise on low-speed motion signals, and particularly solves the trajectory missing report problem caused by low signal-to-noise ratio at the starting stage; the fusion of the denoised data and the environment information for trajectory prediction enables the deep learning model to get rid of the noise stripping burden and focus on modeling complex interaction relationships; the backward algorithm corrects the prediction results to eliminate trajectory mutations through reverse time sequence information transmission, thereby ensuring high-order kinematic consistency of position, velocity, and acceleration. Finally, the response accuracy of the autonomous driving system to the starting obstacle is improved under the premise of ensuring the timeliness of the prediction.

[0125] As an optional implementation, the processing module is further configured to:

[0126] preprocess the obstacle state information;

[0127] In the preprocessing, the obstacle state information is processed according to a preset data structure to establish a mapping relationship between the obstacle state information and the obstacle identity.

[0128] In the preprocessing, the obstacle state information is processed according to a preset data structure to establish a mapping relationship between the obstacle state information and the obstacle identity.

[0129] In the preprocessing, the obstacle state information is processed according to a preset data structure to establish a mapping relationship between the obstacle state information and the obstacle identity.

[0130] The embodiment improves data reliability through structured preprocessing. The mapping relationship between the obstacle identity ensures the independent tracking capability of the multi-target trajectory, thereby avoiding prediction errors caused by identity confusion; the time sequence information alignment processing eliminates the influence of the sensor sampling frequency difference on the state sequence, thereby providing input data with uniform time intervals for the forward algorithm. Both of them solve the fragmentation problem of the original observation data and establish an accurate data basis for subsequent noise reduction and prediction.

[0131] As an optional implementation, the environment information includes lane center line, lane boundary line, lane type, and lane topology connection relationship. After obtaining the environment information according to the obstacle state information and the ego vehicle state information, the processing module is further configured to:

[0132] According to the environment information, determine the correlation parameters between each of the obstacle state information and the ego vehicle state information.

[0133] Through the correlation parameters, the obstacle state information is filtered and updated.

[0134] The embodiment realizes target intelligent screening through environment association. According to lane type and topological relation, an association parameter is calculated to accurately identify an obstacle (such as a vehicle in a neighboring lane) that has a possibility of interaction with the ego vehicle; and based on the association parameter, an obstacle state is screened and updated to eliminate irrelevant targets (such as an obstacle in a reverse lane), thereby significantly reducing the calculation complexity of a deep learning model. While ensuring the prediction accuracy of key obstacles, the real-time performance of the system is improved.

[0135] As an optional implementation, the forward algorithm and the backward algorithm include a Kalman filtering algorithm, and the processing module is further configured to:

[0136] kinematic modeling is performed according to position information, speed information and acceleration information corresponding to the obstacle state information, to determine a state matrix and an observation matrix;

[0137] a numerical value of a process noise covariance matrix is initialized, and noise information of position observation and speed observation is determined according to the obstacle state information, to obtain an observation noise covariance matrix.

[0138] The embodiment optimizes filtering effect through kinematic modeling. A state matrix is constructed based on position / speed / acceleration to establish a motion constraint relationship (such as acceleration continuity) that conforms to physical laws; the sensitivity of the model to state mutation is controlled by initializing the process noise covariance; and adaptive noise reduction intensity adjustment is realized according to observation data. The three work together to ensure that the forward algorithm retains the true motion trend while effectively filtering observation jitter in low-speed scenarios.

[0139] As an optional implementation, the processing module processes the obstacle state information through the forward algorithm, and the specific manner of obtaining the historical position estimation information includes:

[0140] the historical real state sequence corresponding to the obstacle information, and the state matrix, the observation matrix, the process noise covariance matrix and the observation noise covariance matrix are taken as inputs, and Kalman filtering is iteratively performed to obtain Kalman coefficients, an iterative state, an estimation covariance and a prediction covariance at each step, and the iterative state and the estimation covariance are taken as the historical position estimation information.

[0141] The embodiment realizes progressive state optimization through an iterative mechanism. The step-by-step iterative calculation of Kalman filtering utilizes the continuity constraint of the historical state sequence (such as the position at the previous time limiting the current displacement range) to correct observation errors frame by frame; the output iterative state and covariance provide position estimation with confidence evaluation, which provides a basis for the deep learning model to distinguish high / low reliability features. Finally, the robustness of the noise reduction process to temporary signal loss is improved.

[0142] As an optional implementation, the processing module processes and corrects the trajectory prediction result through a backward algorithm to obtain a target trajectory in a specific manner including:

[0143] According to the iteration state and estimated covariance in the last iteration corresponding to the trajectory prediction result, the historical position estimation information obtained by the forward algorithm iteration is corrected by reverse iteration through a smoothing gain matrix to obtain the target trajectory;

[0144] The smoothing gain matrix is ​​used to indicate the influence of historical estimation errors on future prediction errors, so as to constrain the historical position estimation information obtained in each iteration.

[0145] This implementation ensures trajectory continuity through reverse correction. Using a smoothed gain matrix, reverse iteration propagates high-confidence states from later moments back to earlier moments (e.g., current velocity is affected by future acceleration), thus overcoming the one-way information limitations of the forward algorithm. Constraining the propagation of historical estimation errors suppresses inter-frame jumps in the predicted trajectory (e.g., sudden position changes or discontinuous velocity). This enhances kinematic rationality based on deep learning trajectory prediction, meeting the smooth trajectory requirements of the decision-making and planning module.

[0146] As an optional implementation manner, the processing module obtains a trajectory prediction result according to the historical position estimation information and the environmental information in a specific manner including:

[0147] Inputting the historical position estimation information and the environmental information into a preset encoder for encoding to obtain spatial position features, and inputting the spatial position features into a decoder to obtain a trajectory prediction result;

[0148] The decoder is used to calculate the association information between obstacles and the direct association information between obstacles and the environment, so as to infer the trajectory prediction results of the obstacles.

[0149] This implementation fuses multidimensional information through an encoder-decoder architecture. The encoder integrates denoised positions and environmental features to extract spatial positional features that incorporate temporal and spatial relationships (such as the offset of an obstacle relative to the lane centerline). The decoder calculates obstacle-environment interactions (such as the following relationships between vehicles or the positional constraints of pedestrians at intersections) to infer physically accurate trajectory distributions. This approach leverages the complementary strengths of traditional filtering methods and deep learning to produce accurate and interpretable predictions in complex scenarios.

[0150] It should be noted that it should be understood that the division of the various modules of the above device is merely a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. Moreover, these modules can all be implemented in the form of software called by a processing element; or they can all be implemented in the form of hardware; or some modules can be implemented in the form of software called by a processing element, and some modules can be implemented in the form of hardware. For example, the processing module can be a separately established processing element, or it can be integrated into a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called by a processing element of the above device to perform the functions of the above-mentioned module. The implementation of other modules is similar. In addition, these modules can all or partly be integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed by the hardware integrated logic circuit in the processor element or by instructions in the form of software.

[0151] Schematically, as Figure 2 As shown, Figure 2 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. The computer device 300 can be provided as a server. Figure 2 Computer device 300 includes a processing component 302, which further includes one or more processors, and a memory resource represented by memory 301 for storing instructions executable by processing component 302, such as an application. The application stored in memory 301 may include one or more modules, each corresponding to a set of instructions. In addition, processing component 302 is configured to execute the instructions to perform the method of any of the above embodiments.

[0152] The computer device 300 may further include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate based on an operating system stored in the memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or the like.

[0153] Those skilled in the art will understand that Figure 2 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0154] The embodiments of the present application provide a storage medium, the storage medium stores computer readable instructions, the computer readable instructions are executed by one or more processors, so that the one or more processors execute the method provided by any one of the embodiments.

[0155] Finally, it should be noted that in this document, the terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0156] The embodiments in the specification are described in a progressive manner, each embodiment focuses on the difference from other embodiments, and each embodiment can be combined as needed, and the same and similar parts refer to each other.

[0157] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting obstacle trajectories during low-speed starting, characterized in that: include: Obtain obstacle status information and vehicle status information; Acquiring environmental information based on the obstacle status information and the vehicle status information; Processing the obstacle state information through a forward algorithm to obtain historical position estimation information; Obtaining a trajectory prediction result based on the historical position estimation information and the environmental information; The trajectory prediction result is processed and corrected by a backward algorithm to obtain the target trajectory.

2. The method according to claim 1, characterized in that The method further comprises: Preprocessing the obstacle status information; The preprocessing specifically includes: Processing the obstacle status information according to a preset data structure to establish a mapping relationship between the obstacle status information and the obstacle identity; Furthermore, linear interpolation is performed on the obstacle state information based on a time dimension to align the time sequence information corresponding to each piece of obstacle state information.

3. The method according to claim 2, characterized in that The environmental information includes a lane centerline, a lane boundary line, a lane type, and a lane topological connection relationship. After obtaining the environmental information based on the obstacle status information and the vehicle status information, the method further includes: Determining, based on the environmental information, correlation parameters between each obstacle state information and the vehicle state information; The obstacle status information is filtered through the associated parameters and the obstacle status information is updated.

4. The method according to claim 1, wherein The forward algorithm and the backward algorithm include a Kalman filter algorithm, and the method further includes: Perform kinematic modeling based on the position information, velocity information, and acceleration information corresponding to the obstacle state information to determine the state matrix and the observation matrix; The value of the process noise covariance matrix is ​​initialized, and noise information of the position observation and the velocity observation is determined according to the obstacle state information to obtain the observation noise covariance matrix.

5. The method according to claim 4, characterized in that The process of processing the obstacle state information by a forward algorithm to obtain historical position estimation information includes: The historical real state sequence corresponding to the obstacle information, as well as the state matrix, the observation matrix, the process noise covariance matrix and the observation noise covariance matrix are taken as input, and Kalman filtering is iteratively performed to obtain the Kalman coefficient, iterative state, estimated covariance and predicted covariance of each step, and the iterative state and estimated covariance are used as historical position estimation information.

6. The method according to claim 4, characterized in that The step of processing and correcting the trajectory prediction result by a backward algorithm to obtain the target trajectory includes: According to the iteration state and estimated covariance in the last iteration corresponding to the trajectory prediction result, the historical position estimation information obtained by the forward algorithm iteration is corrected by reverse iteration through a smoothing gain matrix to obtain the target trajectory; The smoothing gain matrix is ​​used to indicate the influence of historical estimation errors on future prediction errors, so as to constrain the historical position estimation information obtained in each iteration.

7. The method according to claim 1, characterized in that The obtaining of a trajectory prediction result according to the historical position estimation information and the environmental information includes: Inputting the historical position estimation information and the environmental information into a preset encoder for encoding to obtain spatial position features, and inputting the spatial position features into a decoder to obtain a trajectory prediction result; The decoder is used to calculate the association information between obstacles and the direct association information between obstacles and the environment, so as to infer the trajectory prediction results of the obstacles.

8. A low-speed starting obstacle trajectory prediction device, characterized in that: include: Acquisition module, used to obtain obstacle status information and vehicle status information; A processing module, configured to obtain environmental information based on the obstacle status information and the vehicle status information; The processing module is further configured to process the obstacle state information using a forward algorithm to obtain historical position estimation information; The processing module is further configured to obtain a trajectory prediction result based on the historical position estimation information and the environmental information; The processing module is further configured to process and correct the trajectory prediction result through a backward algorithm to obtain a target trajectory.

9. A computer device, characterized in that: The method comprises one or more processors and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the method according to any one of claims 1 to 7 are performed.

10. A storage medium, characterized in that: The storage medium stores computer-readable instructions, which, when executed by one or more processors, enable the one or more processors to perform the steps of the method according to any one of claims 1 to 7.

Images