Intersection trajectory prediction method fusing signal phase and intersection reachable green area

Abstract

This invention relates to a method for predicting intersection trajectories by fusing signal phase and reachable traffic areas. The method extracts historical trajectory sequences and obtains signal phase and timing information, as well as high-precision map information. M 1. Gather at the intersection where vehicles can pass; then first check and... M The method involves encoding, concatenating features, fusing them using an MLP, and then capturing the interactive features via a bidirectional LSTM and multi-head self-attention mechanism. These interactive features are then concatenated with historical trajectory features and captured again using a multi-head self-attention mechanism to obtain signal timing interactive features. The set of reachable traffic areas at the intersection is concatenated with [other features] and fused using MHA to obtain the final features. Finally, a three-layer MLP is used for decoding, and the consistency of candidate trajectories is verified and corrected using the reachable traffic areas at the intersection and signal phase timing constraints. This method improves the accuracy, interpretability, and robustness of trajectory prediction in signalized intersection scenarios.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicle environmental perception and behavior prediction technology, and relates to trajectory prediction in signal-controlled intersection scenarios. Specifically, it relates to an intersection trajectory prediction method that integrates signal phase and the reachable traffic area of ​​the intersection. Background Technology

[0002] With the development of intelligent connected vehicles and autonomous driving technologies, the ability of vehicles to perceive the environment, understand behavior, and predict trajectories in urban roads, especially at signalized intersections, has become a key foundation for achieving safe decision-making and comfortable traffic flow. Intersections involve diverse types of traffic participants with a wide range of movement intentions, and are simultaneously subject to strong constraints from road geometry, lane connectivity, and traffic light control, resulting in future trajectories exhibiting significant multimodal and highly interactive characteristics. If the prediction results fail to reflect traffic rules and scenario constraints, it will directly impact the safety and effectiveness of downstream planning and control, and may lead to high-risk decisions such as crossing boundaries, misjudging conflict points, and running red lights. Existing trajectory prediction methods mostly rely on historical trajectory sequences and surrounding interaction features to build deep learning models. While these methods can achieve good error metrics in general road scenarios, they often fail to adequately express two key prior constraints in signalized intersections by simply relying on kinematic history and local interactions: firstly, the temporal constraint of traffic light phase and timing on right-of-way; and secondly, the spatial constraint of the accessible area at the intersection on mobility. Due to the lack of explicit modeling of these spatiotemporal constraints, prediction results are prone to geometrically infeasible phenomena such as unreasonable responses to phase switching critical points and trajectories crossing lane boundaries or restricted areas. Furthermore, it is difficult to uniformly characterize the feasible domain of "topological accessibility along legal lanes within a given phase window," leading to multimodal branch distortion and risk assessment bias in intersection scenarios with high interaction and multiple conflict points.

[0003] Although some existing studies or methods attempt to constrain and optimize trajectory prediction results through drivable area features or traffic signal timing information, such as Chinese patent CN115703486A which discloses "A Vehicle Trajectory Prediction Method Based on Drivable Area Feature Extraction" and Chinese patent CN 105035090A which discloses "An Autonomous Driving Vehicle Trajectory Prediction and Control Method Based on Traffic Signals," most existing methods still improve trajectory prediction from a single dimension. That is, they only consider the spatial geometric constraints provided by the drivable area or only consider the time right-of-way constraints provided by the traffic signal phase and timing. They have not yet achieved deep synergistic integration of the two under a unified prediction framework. In addition, since the future behavior of vehicles in intersection scenarios is essentially constrained by both signal control rules and road reachability, relying solely on either type of prior information is insufficient to fully characterize the true behavioral boundary of a vehicle's "accessibility along the legal lane topology within a specific signal phase window." This can easily lead to a disconnect between the temporal rationality and spatial feasibility of the prediction results. Especially in complex intersection scenarios with high interaction and multiple conflict points, how to unify and effectively integrate traffic light signal phase and timing priors, intersection reachability constraints, and multimodal trajectory prediction mechanisms remains a key technical problem that urgently needs to be solved in the field of intersection trajectory prediction. This is of great significance for improving the accuracy, rationality, and interpretability of trajectory prediction. Therefore, it is necessary to propose a multimodal trajectory prediction method for intersections that integrates traffic light signal phase and timing priors with intersection reachability constraints. This method simultaneously introduces time-dimensional right-of-way constraints and spatial-dimensional reachability constraints during the prediction process, constructing prediction results that are consistent with road rules, geometrically feasible, and capable of expressing multiple future intentions, thus providing reliable input for subsequent path planning, risk assessment, and decision control. Summary of the Invention

[0004] In view of the shortcomings and deficiencies of existing technologies, the purpose of this invention is to propose an intersection trajectory prediction method that integrates signal phase and accessible traffic area. This method first acquires historical trajectory information of the target subject and surrounding traffic participants, and simultaneously acquires intersection map elements, accessible traffic area information, and signal phase timing information, performing spatiotemporal alignment on the above data. Further, it constructs a constraint representation of the accessible traffic area and a right-of-way time constraint representation based on SPaT, and jointly encodes them with the historical movement and interaction features of multiple subjects to form a spatiotemporal constraint condition representation for intersection scenarios. Based on this, conditional multimodal decoding is used to generate multiple future candidate trajectories and their probability distributions. The candidate trajectories are then subjected to consistency verification, probability recalibration, and necessary elimination or projection correction based on the consistency of the accessible traffic area and signal right-of-way constraints. Finally, a multimodal trajectory prediction result that satisfies road rules and geometric feasibility is output, providing reliable input for subsequent path planning, risk assessment, and decision control, thereby improving the accuracy, interpretability, and robustness of trajectory prediction in signal-controlled intersection scenarios.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for predicting intersection trajectories by fusing signal phase and the reachable traffic area of ​​the intersection, the method comprising the following steps:

[0007] Step 1. Collect multi-source environmental data and road structure information, extract historical trajectory sequences, obtain signal phase and timing information, high-precision map information and the set of lane relationships of traffic participants in the scene, determine the candidate maneuver types of vehicles, calculate the reachable area of ​​intersections based on the candidate maneuver types, and obtain the set of reachable areas of all vehicles at intersections in the scene.

[0008] Step 2. First, encode the historical trajectory sequence and high-precision map information to obtain historical trajectory features. and high-precision map features The two features are concatenated, fused using an MLP, and then sequentially passed through a bidirectional LSTM recurrent network and a multi-head self-attention mechanism to capture the interaction features. The obtained phase and timing information is then compared with historical trajectory features. The signals are spliced ​​together and captured using a multi-head self-attention mechanism to obtain the signal timing interaction features. Finally, the set of accessible traffic areas at intersections and the interaction characteristics of signal timing are analyzed. Interaction features The features are concatenated and fused using a multi-head self-attention mechanism to obtain the features. ;

[0009] Step 3. Use a three-layer MLP for decoding to generate K trajectories and their corresponding probabilities; use the reachable area of ​​the intersection and signal phase timing constraints to perform spatial consistency and signal consistency verification and correction on the candidate trajectories, output the set of candidate trajectories and their probabilities after consistency processing, and then select the Top-K trajectories from high to low probabilities as the multimodal trajectory prediction results.

[0010] As a preferred embodiment of the present invention, the multi-source environmental data and road structure information mentioned in step 1 include the perception output information of the target subject and surrounding traffic participants and road map information.

[0011] As a preferred embodiment of the present invention, the signal phase and timing information mentioned in step 1 includes the current phase state, phase remaining time, and phase sequence for each lane.

[0012] As a preferred embodiment of the present invention, the candidate maneuver types in step 1 include going straight, turning right, turning left, and making a U-turn. Different candidate maneuver types correspond to different maneuver intervals. When determining the reachable area of ​​the intersection, first determine all the candidate maneuver types corresponding to vehicle i based on the lane to which vehicle i belongs, then calculate the maneuver intervals corresponding to different candidate maneuver types respectively, and then add up the maneuver intervals corresponding to all candidate maneuver types to obtain the reachable area of ​​vehicle i at the intersection.

[0013] As a preferred embodiment of the present invention, step 2 uses a pointNet-based polyline encoder to process the historical trajectory sequence. and high-precision map information M To obtain historical trajectory characteristics and high-precision map features .

[0014] As a preferred embodiment of the present invention, the number of heads in the multi-head self-attention mechanism used in step 2 is 8.

[0015] As a preferred embodiment of the present invention, in step 3, spatial consistency and signal consistency checks are performed on each candidate trajectory. During the spatial consistency check, if there are... τ This ensures that the predicted trajectory point of vehicle i falls within its accessible area at the intersection. In addition, the out-of-bounds point is projected into the reachable area of ​​the loop intersection and probability reduction is performed; during signal consistency verification, the allowed passage direction and time window under the current phase are determined according to the signal phase and timing information corresponding to the intersection. If the candidate trajectory crosses the stop line or enters the conflict zone under the red light or non-allowed phase, the trajectory is eliminated.

[0016] As a further preferred embodiment of the present invention, in step 1, the intersection is abstracted as a square of length D, and the center point of the square is set to (0,0). When vehicle i is in the straight lane, the corresponding maneuvering interval is... for:

[0017] ;

[0018] When vehicle i is in the right-turn lane, the corresponding maneuvering area for:

[0019] ;

[0020] When vehicle i is in the left-turn lane, the corresponding maneuvering area for:

[0021] ;

[0022] When vehicle i is in the U-turn lane, the corresponding maneuvering area for:

[0023] ;

[0024] Where D is the intersection scale parameter, i.e., the length of the square at the intersection, which is read directly from the high-precision map information. , , , These represent the maneuvering ranges corresponding to the candidate maneuver types of going straight, turning right, turning left, and making a U-turn, respectively.

[0025] As a further preferred embodiment of the present invention, before the consistency check in step 3, the candidate mobility type of the vehicle is predicted based on the predicted trajectory points of the candidate trajectory. Then, it is determined whether the trajectory prediction point falls outside the reachable area of ​​the intersection corresponding to the candidate mobility type. If so, the boundary point is projected back into the reachable area of ​​the intersection.

[0026] Advantages and beneficial effects of the present invention:

[0027] (1) This invention introduces traffic light signal phase and timing priors as well as intersection reachable traffic area constraints, which can collaboratively model the future behavior of vehicles from the time and space dimensions. Compared with prediction methods that only use signal timing information or intersection reachable traffic area information, it can more completely depict the right-of-way constraints and geometric reachability constraints of vehicle movement in the intersection scenario, thereby improving the rationality and completeness of trajectory prediction results.

[0028] (2) The present invention adopts a multi-head attention mechanism to interactively fuse signal phase and timing information, intersection reachable traffic area information and trajectory historical features. It can learn the correlation between different constraint factors from multiple feature subspaces, so that the model can not only pay attention to the impact of signal control on vehicle passage timing, but also pay attention to the restrictions of the intersection reachable traffic area on vehicle movement direction, steering mode and spatial distribution range, thereby enhancing the model's ability to fuse and express multi-source heterogeneous information in complex intersection scenarios.

[0029] (3) Unlike existing technologies that construct drivable areas for general road scenarios, this invention limits the reachable area to the interior of the intersection, making its spatial scope more consistent with the control area of ​​traffic lights. Since traffic lights mainly constrain the right-of-way at intersection entrances, conflict areas within the intersection, and the release relationship for different turning directions, this limitation method enables the reachable area to better carry the spatial semantics corresponding to the signal phase and timing information, thereby achieving a tight coupling between time-dimensional right-of-way constraints and spatial-dimensional reachability constraints. As a result, it not only improves the rationality of the model's response to red light prohibition, green light release, and phase switching processes, but also reduces unreasonable phenomena such as the predicted trajectory entering the wrong direction channel, crossing areas that should not be entered, or being inconsistent with the current release direction within the intersection.

[0030] (4) Based on the lane type, lane topology connectivity and vehicle lane matching results provided by the high-precision map, this invention constructs a set of candidate mobility types for traffic participants and their corresponding intersection reachable traffic area constraints. In the process of multimodal trajectory generation and screening, the prior and feasible domain consistency judgment of the above candidate set is introduced, so that the prediction branch is guided by reachable channels and rule constraints in space, avoiding problems such as mode mixing, branch redundancy or meaningless splitting under complex geometric conditions of intersections, thereby improving the interpretability, controllability and engineering usability of multimodal prediction results.

[0031] (5) This invention splices SPaT (current phase, remaining time, phase sequence) with multi-subject motion features and encodes them through a multi-head self-attention mechanism, enabling the model to explicitly learn the impact of signal timing on traffic participants' behavioral decisions such as "passing / decelerating / waiting". Especially in key situations such as phase approach switching, queue initiation and dissipation, it can generate a prediction distribution that is more in line with actual traffic strategies, thereby improving the prediction stability and accuracy of high-interaction periods at intersections.

[0032] (6) The present invention adopts a two-stage process of first generating K candidate trajectories and their probabilities, and then performing consistency verification, projection correction and probability recalibration based on the intersection reachable traffic area constraint and SPaT right-of-way constraint. While maintaining the diversity of multimodal candidates, the output probability of non-compliant trajectories is significantly reduced through spatial and signal consistency processing, avoiding the problems of training instability or inference stage out-of-bounds residue caused by traditional loss penalty alone, thereby improving the robustness and safety redundancy of the model in engineering deployment.

[0033] (7) This invention constructs spatial consistency judgment rules and signal consistency judgment rules for candidate trajectories, which are used to check whether the candidate trajectory meets the intersection reachable traffic area constraint and whether it meets the SPaT right-of-way constraint, respectively. When a candidate trajectory has inconsistent behaviors such as crossing the boundary, crossing the stop line, or entering the conflict zone under the non-allowed phase, the candidate trajectory is directly eliminated or the boundary point is projected to correct the intersection reachable traffic area before participating in the output, so that only the set of trajectories that meet the road rules and geometric feasibility are retained for output. This explicit consistency screening mechanism makes the prediction results have clear non-compliance interpretability and auditability, which makes it easy for the downstream planning and risk assessment modules to directly use the screened trajectory set for decision fusion, reducing the safety risks caused by non-compliant trajectories to planning control.

[0034] (8) This invention mainly relies on the conventional input of existing autonomous driving systems, including multi-source perception output (camera / radar / fusion positioning), high-precision map and lane topology information, as well as SPaT information that can be inferred from roadside equipment or vehicle end. It does not require additional expensive hardware or complex modifications, and can be directly embedded into the existing trajectory prediction-planning control architecture. It is suitable for typical urban high-complexity scenarios such as signal-controlled intersections, and has good engineering implementation value and promotion and application prospects. Attached Figure Description

[0035] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings:

[0036] Figure 1 Flowchart of the intersection trajectory prediction method that integrates signal phase and reachable traffic area provided by the present invention;

[0037] Figure 2 This is a schematic diagram of an intersection. Detailed Implementation

[0038] To enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application will be described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0039] like Figure 1 , Figure 2 As shown in the figure, this embodiment provides a method for predicting intersection trajectories by fusing signal phase and the reachable traffic area of ​​the intersection. The method includes the following steps:

[0040] Step 1. Data Acquisition and Feature Extraction for Intersection Scenes:

[0041] Step 1.1. Collect multi-source environmental data and road structure information related to trajectory prediction during the vehicle's journey at the signal-controlled intersection.

[0042] Specifically, in this embodiment, the multi-source environmental data and road structure information include at least: the perception output information of the target subject and surrounding traffic participants (e.g., position, speed, acceleration, heading angle / heading angular velocity, etc.), and road map information used to constrain motion accessibility (e.g., lane centerline, lane boundary, lane topology connectivity, intersection accessible passage area boundary, polygonal corridor, stop line, pedestrian crossing, traffic island and restricted area, etc.).

[0043] Furthermore, in this embodiment, the multi-source environmental data can be acquired by vehicle-mounted cameras, millimeter-wave radar, lidar, fusion perception systems, or V2X collaborative perception; the road structure information can be acquired by high-precision maps and positioning systems, or acquired online by vision / laser lane line and curb extraction modules.

[0044] Step 1.2. Based on the data obtained in Step 1.1, extract the historical trajectory sequences of the vehicle and surrounding traffic participants within the historical time window. .

[0045] Specifically, in this embodiment, the historical trajectory sequence Specifically, this includes the vehicle's historical trajectory. Historical trajectory of surrounding traffic participants , where 𝑁 represents the number of traffic participants in the surrounding area. Represents the historical trajectory of the i-th traffic participant; each trajectory sequence consists of... The state vector is composed of state vectors at each historical moment. The state vectors contain at least two-dimensional position (x,y) and may further contain kinematic quantities such as velocity, acceleration and heading angle, which are used for subsequent interactive coding and multimodal prediction modeling.

[0046] Step 1.3. Obtain the signal phase and timing (SPaT) information corresponding to the target signalized intersection. .

[0047] Specifically, in this embodiment, the signal phase and timing (SPaT) information... This includes the current phase state for each lane (i.e., the traffic light to which the current lane belongs), the remaining time of the phase (i.e., how many seconds are left until the current phase ends), and the phase sequence (which phase to enter after the current phase ends, and the phase rotation order).

[0048] Step 1.4. Obtain high-precision map information corresponding to the target signalized intersection scenario. M The set of lane relationships between traffic participants in the scene E .

[0049] Specifically, in this embodiment, the high-precision map information M This includes lane centerline, lane boundaries, lane topology connectivity, and lane type; the set of lane relationships to which traffic participants belong in the described scenario. E It is generated by matching the vehicle's historical trajectory coordinates with high-precision map information.

[0050] Step 1.5. Based on the set of lane relationships of traffic participants in the scenario. E The mapping relationship determines the candidate mobility type of the i-th traffic participant (vehicle i). ,in Indicates going straight. Indicates a right turn. Indicates a left turn. This indicates a U-turn; based on the high-precision map information of the intersection obtained in step 1.4, the intersection scale parameters are extracted. D Calculate the maneuver intervals corresponding to different candidate maneuver types, and calculate the reachable area of ​​the intersection based on all candidate maneuver types corresponding to vehicle i. Finally, the set of all accessible traffic areas at intersections in the scene is obtained. .

[0051] Specifically, in this embodiment, the current lane of each vehicle i is first obtained from the set of lane relationships of traffic participants. Candidate maneuver types are then determined based on the lane. For example, if vehicle i is currently in a left-turn + straight lane, then the candidate maneuver types for vehicle i are left turn and straight. For the two candidate maneuver types, left turn and straight, corresponding to different maneuver intervals, the maneuver intervals for left turn and straight are calculated separately. Finally, the combination of maneuver intervals calculated based on all candidate maneuver types corresponding to vehicle i is the reachable passage area of ​​vehicle i at the intersection, expressed in the form of: = + .

[0052] Furthermore, such as Figure 2In this embodiment, the intersection (crossroads) is abstracted as a square of length D, with the center point of the square set to (0,0). The intersection scale parameter D can be directly read from high-precision map information. When vehicle i is in the straight lane, the corresponding maneuvering interval is... for:

[0053] ;

[0054] When vehicle i is in the right-turn lane, the corresponding maneuvering area for:

[0055] ;

[0056] When vehicle i is in the left-turn lane, the corresponding maneuvering area for:

[0057] ;

[0058] When vehicle i is in the U-turn lane, the corresponding maneuvering area for:

[0059] ;

[0060] Step 2. Multi-agent interaction coding:

[0061] Step 2.1. Use an encoder to process the historical trajectory sequence. and high-precision map information M Encoding yields historical trajectory features. and high-precision map features .

[0062] Specifically, in this embodiment, a pointNet-based polyline encoder is used to process the historical trajectory sequence. and high-precision map information M To obtain historical trajectory characteristics and high-precision map features , can be represented as:

[0063]

[0064]

[0065] In the formula, This indicates a pointNet polyline encoder.

[0066] Step 2.2. Transfer the historical trajectory features obtained in Step 2.1 and high-precision map features After stitching, a multilayer perceptron (MLP) is used for fusion to obtain the initial fused features. .

[0067] Specifically, in this embodiment, the two feature vectors are concatenated along the first dimension to obtain the concatenated feature vectors. Then, a simple multilayer perceptron (MLP) is used for fusion to obtain the initial fused features. , can be represented as:

[0068]

[0069]

[0070] In the formula, This means concatenating the feature vectors along the first dimension. MLP It consists of two fully connected layers and a ReLU activation function.

[0071] Step 2.3. Use the Phase and Timing (SPaT) information obtained in Step 1.3. Compared with the historical trajectory features obtained in step 2.1 By splicing the data, the phase timing features of the historical trajectories are obtained. , can be represented as:

[0072]

[0073] Step 2.4. Utilize the historical trajectory phase timing features obtained in Step 2.3. A multi-head self-attention mechanism was used to capture the impact of signal phase and timing information on traffic participant behavior, resulting in signal timing interaction features. , can be represented as:

[0074]

[0075] In the formula, MHA This represents a multi-head self-attention mechanism with 8 heads, and is the query matrix for this mechanism. Q Key matrix K Sum matrix V All come from characteristics .

[0076] Step 2.5. The initial fusion features from Step 2.2... Bidirectional LSTM is used to capture the temporal features of traffic participants' trajectories, thus obtaining historical trajectory temporal features. , can be represented as:

[0077]

[0078] In the formula, It is a bidirectional LSTM recurrent network structure.

[0079] Step 2.6. Analyze the historical trajectory temporal features from Step 2.5. A multi-head self-attention mechanism is used to capture interactions between traffic participants and between traffic participants and the map, resulting in interaction features. , can be represented as:

[0080]

[0081] In the formula, MHA This represents a multi-head self-attention mechanism with 8 heads, and is the query matrix for this mechanism. Q Key matrix K Sum matrix V All come from characteristics .

[0082] Step 2.7. Set the reachable traffic areas of the intersections obtained in Step 1.5. = The signal timing interaction features obtained in step 2.4 and the interaction features obtained in step 2.6 The features are then concatenated and then fused using an MHA to obtain the features. , can be represented as:

[0083]

[0084]

[0085] In the formula, This means concatenating the feature vectors along the first dimension. This represents the features after splicing.

[0086] Step 3. Multimodal Decoding and Consistency Constraints:

[0087] This step employs a two-stage strategy of generating first and then constraining: generating first without explicitly introducing space constraints. K The candidate future trajectories and their probabilities (or confidence levels) are obtained. Then, the consistency of the candidate trajectories is checked and corrected by the intersection reachability area constraint and the signal phase timing (SPaT) constraint, so as to output the prediction results that meet the road rules and geometric feasibility.

[0088] Step 3.1. Use a three-layer MLP for decoding to generate K trajectories and their corresponding probabilities, which can be represented as:

[0089]

[0090]

[0091] In the formula, It consists of three simple MLP layers stacked together. A simple MLP consists of two fully connected layers and a ReLU activation function. This represents the sequence of discrete trajectory points of the nth candidate trajectory in the prediction time domain. Indicates the future trajectory time step. Represents the nth candidate trajectory in The trajectory points at any given moment This represents the pattern probability or confidence level of the nth candidate trajectory.

[0092] Step 3.2. For the candidate trajectory including the nth one... Each candidate trajectory, including the one mentioned above, undergoes spatial consistency and signal consistency checks. Based on these checks, candidate trajectories are eliminated, projection corrected, and probability reduced to obtain the final trajectory. Sum of probabilities The details are as follows:

[0093] (1) Spatial consistency (intersection accessibility area constraints):

[0094] If it exists τ This ensures that the predicted trajectory point of vehicle i falls within its accessible area at the intersection. Besides ( If the boundary point is projected into the reachable area of ​​the loop entrance, then the probability is reduced (by 0.05).

[0095] Furthermore, in this embodiment, the candidate maneuver type of the vehicle can be predicted based on the predicted trajectory points of the candidate trajectory. Then, it is determined whether the predicted trajectory point falls outside the reachable area of ​​the intersection corresponding to the candidate maneuver type. If so, the boundary point is projected back into the reachable area of ​​the intersection. This method enables the predicted branch to be guided by reachable channels and rules in space, avoiding problems such as pattern aliasing, branch redundancy, or meaningless splitting under complex intersection geometry.

[0096] (2) Signal consistency (SPaT right-of-way constraint):

[0097] Based on the signal phase and timing information corresponding to the intersection Determine the permitted direction and time window under the current phase. If a candidate trajectory crosses the stop line or enters the conflict zone under a red light or non-permitted phase, the trajectory is eliminated.

[0098] Finally, the output is the set of candidate trajectories after consistency processing. Y k Sum of probabilities π kFurthermore, the top-K lines can be selected from high to low probability as the multimodal trajectory prediction results.

[0099] The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the intersection trajectory prediction method based on the fusion of signal phase and reachable traffic area described above.

[0100] The present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the intersection trajectory prediction method described above, which integrates signal phase and reachable traffic area of ​​the intersection.

[0101] Those skilled in the art will understand that all or part of the functions of the various methods / modules in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the above functions can be implemented by executing the program with a computer. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented.

[0102] In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the programs can also be stored in storage media such as servers, other computers, disks, optical discs, flash drives, or portable hard drives. They can be downloaded or copied to the memory of the local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0103] The above-described specific examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for predicting intersection trajectories by fusing signal phase and reachable traffic area, characterized in that, The method includes the following steps: Step 1. Collect multi-source environmental data and road structure information, extract historical trajectory sequences, obtain signal phase and timing information, high-precision map information and the set of lane relationships of traffic participants in the scene, determine the candidate maneuver types of vehicles, calculate the reachable area of ​​intersections based on the candidate maneuver types, and obtain the set of reachable areas of all vehicles at intersections in the scene. Step 2. First, encode the historical trajectory sequence and high-precision map information to obtain historical trajectory features. and high-precision map features The two features are concatenated, fused using an MLP, and then sequentially passed through a bidirectional LSTM recurrent network and a multi-head self-attention mechanism to capture the interaction features. The obtained phase and timing information is then compared with historical trajectory features. The signals are spliced ​​together and captured using a multi-head self-attention mechanism to obtain the signal timing interaction features. Finally, the set of accessible traffic areas at intersections and the interaction characteristics of signal timing are analyzed. Interaction features The features are concatenated and fused using a multi-head self-attention mechanism to obtain the features. ; Step 3. Use a three-layer MLP for decoding to generate K trajectories and their corresponding probabilities; use the reachable area of ​​the intersection and signal phase timing constraints to perform spatial consistency and signal consistency verification and correction on the candidate trajectories, output the set of candidate trajectories and their probabilities after consistency processing, and then select the Top-K trajectories from high to low probabilities as the multimodal trajectory prediction results; In step 3, spatial consistency and signal consistency checks are performed on each candidate trajectory. During the spatial consistency check, if there are... τ This ensures that the predicted trajectory point of vehicle i falls within its accessible area at the intersection. In addition, the out-of-bounds point is projected into the reachable area of ​​the loop intersection and probability reduction is performed; during signal consistency verification, the allowed passage direction and time window under the current phase are determined according to the signal phase and timing information corresponding to the intersection. If the candidate trajectory crosses the stop line or enters the conflict zone under the red light or non-allowed phase, the trajectory is eliminated.

2. The intersection trajectory prediction method based on the fusion of signal phase and reachable traffic area as described in claim 1, characterized in that, The multi-source environmental data and road structure information mentioned in step 1 include the perception output information of the target subject and surrounding traffic participants, as well as road map information.

3. The intersection trajectory prediction method based on the fusion of signal phase and reachable traffic area as described in claim 1, characterized in that, The signal phase and timing information mentioned in step 1 includes the current phase state, remaining phase time, and phase sequence for each lane.

4. The intersection trajectory prediction method based on the fusion of signal phase and reachable traffic area as described in claim 1, characterized in that, In step 1, the candidate maneuver types include going straight, turning right, turning left, and making a U-turn. Different candidate maneuver types correspond to different maneuver intervals. When determining the reachable area at the intersection, first determine all the candidate maneuver types corresponding to vehicle i based on the lane to which vehicle i belongs. Then calculate the maneuver intervals corresponding to different candidate maneuver types. Finally, add up the maneuver intervals corresponding to all candidate maneuver types to obtain the reachable area at the intersection for vehicle i.

5. The intersection trajectory prediction method based on the fusion of signal phase and reachable traffic area as described in claim 1, characterized in that, Step 2 uses a pointNet-based polyline encoder to process the historical trajectory sequence. and high-precision map information M To obtain historical trajectory characteristics and high-precision map features .

6. The intersection trajectory prediction method based on the fusion of signal phase and reachable traffic area as described in claim 1, characterized in that, The number of heads used in the multi-head self-attention mechanism in step 2 is 8.

7. The intersection trajectory prediction method based on the fusion of signal phase and reachable traffic area as described in claim 4, characterized in that, In step 1, the intersection is abstracted as a square of length D, and the center point of the square is set to (0,0). When vehicle i is in the straight lane, the corresponding maneuvering interval is... for: ； When vehicle i is in the right-turn lane, the corresponding maneuvering area for: ； When vehicle i is in the left-turn lane, the corresponding maneuvering area for: ； When vehicle i is in the U-turn lane, the corresponding maneuvering area for: ； Where D is the intersection scale parameter, i.e., the length of the square at the intersection, which is read directly from the high-precision map information. , , , These represent the maneuvering ranges corresponding to the candidate maneuver types of going straight, turning right, turning left, and making a U-turn, respectively.

8. The intersection trajectory prediction method based on the fusion of signal phase and reachable traffic area as described in claim 1, characterized in that, Before performing the consistency check in step 3, predict the candidate maneuver type of the vehicle based on the predicted trajectory points of the candidate trajectory. Then, determine whether the predicted trajectory point falls outside the reachable area of ​​the intersection corresponding to the candidate maneuver type. If so, project the boundary point back into the reachable area of ​​the intersection.

Images