Sparse trajectory recovery method and device, system, storage medium

Abstract

The application discloses a sparse trajectory recovery method and device, system and storage medium, comprising: collecting sparse trajectory point sequence and traffic physical constraint condition information; performing path inference and speed interpolation on the sparse trajectory point sequence to obtain a trajectory infrastructure containing space connectivity and speed continuity information; extracting traffic physical constraint features according to the trajectory infrastructure and the traffic physical constraint condition information, fusing the extracted traffic physical constraint features through unified normalization and feature splicing technology to construct a traffic physical constraint condition vector; and based on the traffic physical constraint condition vector, using a conditional diffusion generation model to gradually denoise and reconstruct the sparse trajectory to generate a time-continuous and space-rational complete trajectory point sequence. The technical scheme of the application can improve trajectory recovery accuracy, reduce illegal turning rate and optimize speed smoothness.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent transportation, and particularly relates to a sparse trajectory recovery method and device, system and storage medium. BACKGROUND

[0002] With the rapid development of intelligent transportation, urban travel management and trajectory data mining technology, trajectory data collected based on mobile devices, vehicle GPS and traffic monitoring systems has become an important basis for traffic management, path planning, travel prediction, etc. However, due to factors such as device sampling frequency, signal obstruction, energy consumption control and privacy protection, the actual collected trajectory data often has serious sparsity and missing problems. Sparse trajectory data not only causes the continuity of the trajectory to be damaged, but also affects the effect of subsequent applications such as path inference, traffic behavior identification and intelligent scheduling.

[0003] Existing trajectory recovery techniques mainly rely on time series completion methods based on interpolation, graph search or deep learning. Although traditional interpolation algorithms such as linear interpolation and spline interpolation have high computational efficiency, they are difficult to capture nonlinear motion patterns in complex traffic environments. The method based on graph search can partially utilize the road topology constraints, but lacks fine modeling of traffic physical constraints. Deep learning methods, especially time series completion techniques based on conditional generation, have improved the accuracy of trajectory recovery to some extent. However, existing deep generation methods generally have the following limitations:

[0004] Firstly, most existing methods focus on the spatio-temporal continuity and global fitting effect of trajectory points, ignoring the important constraints of traffic physical rules on trajectory generation. For example, illegal turning, red light violation, and dramatic speed fluctuations are common in existing generated trajectories, which seriously affect the traffic compliance and actual usability of trajectory recovery results.

[0005] Secondly, some existing methods use context awareness or multi-source information fusion to optimize trajectory completion, but lack explicit modeling of traffic physical constraints and hard constraint mechanism design, which makes the trajectory generation process easily deviate from the real traffic behavior.

[0006] In addition, current trajectory generation methods directly predict based on sparse trajectory points, lack reasonable prior path initialization mechanism, and make the diffusion model training efficiency low and the inference process convergence speed slow. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a sparse trajectory recovery method and device, system and storage medium, which can fully integrate traffic physical rules, support path prior completion, and generate compliant trajectories using a conditional diffusion model, effectively solving the problems of insufficient traffic compliance and unreasonable generated paths in existing trajectory recovery technology.

[0008] To achieve the above object, the application adopts the following technical solutions:

[0009] A sparse trajectory recovery method comprises:

[0010] Collecting sparse trajectory point sequences and traffic physical constraint condition information;

[0011] Performing path inference and speed interpolation on the sparse trajectory point sequences to obtain a trajectory infrastructure containing spatial connectivity and speed continuity information;

[0012] According to the trajectory infrastructure and the traffic physical constraint condition information, traffic physical constraint features are extracted, and the extracted traffic physical constraint features are fused through uniform normalization and feature splicing technology to construct a traffic physical constraint condition vector;

[0013] Based on the traffic physical constraint condition vector, a conditional diffusion generation model is used to gradually denoise and reconstruct the sparse trajectory to generate a time-continuous and spatially reasonable complete trajectory point sequence.

[0014] As a preferred, the traffic physical constraint features include: speed estimate value, spatial region code, red light stay tendency score, and turning legality mask.

[0015] As a preferred, a traffic physical constraint loss function is designed for the diffusion generation process of the conditional diffusion generation model, and the traffic physical constraint loss function contains: a speed continuity loss function, a spatial coherence loss function, a turning legality loss function, and a red light stay rationality loss function.

[0016] The application also provides a sparse trajectory recovery device, comprising:

[0017] A first processing module is configured to collect sparse trajectory point sequences and traffic physical constraint condition information;

[0018] A second processing module is configured to perform path inference and speed interpolation on the sparse trajectory point sequences to obtain a trajectory infrastructure containing spatial connectivity and speed continuity information;

[0019] A third processing module is configured to extract traffic physical constraint features according to the trajectory infrastructure and the traffic physical constraint condition information, and fuse the extracted traffic physical constraint features through uniform normalization and feature splicing technology to construct a traffic physical constraint condition vector;

[0020] A fourth processing module is configured to generate a time-continuous and spatially reasonable complete trajectory point sequence by gradually denoising and reconstructing the sparse trajectory based on the traffic physical constraint condition vector and using a conditional diffusion generation model.

[0021] As preferred, the traffic physics constraint features include: speed estimation value, spatial region code, red light stay tendency score, and turning legality mask.

[0022] As preferred, a diffusion generation process of the conditional diffusion trajectory generation model is designed with a traffic physics constraint loss function, which includes: a speed continuity loss function, a spatial coherence loss function, a turning legality loss function, and a red light stay rationality loss function.

[0023] The application also provides a sparse trajectory recovery system, comprising a memory and a processor, wherein the memory stores a computer program which is run by the processor, and the computer program performs the sparse trajectory recovery method when run by the processor.

[0024] The application also provides a storage medium, wherein the storage medium stores a computer program which performs the sparse trajectory recovery method when run.

[0025] The application has at least the following beneficial effects:

[0026] 1) The application first introduces explicit modeling of traffic physics rules in the sparse trajectory recovery task, generates compliant trajectories that conform to traffic regulations through conditional vector input and rule constraint loss optimization, and significantly reduces illegal turning, red light violations, and sudden speed changes.

[0027] 2) The application provides path initialization and kinematic interpolation through a hidden Markov model, reasonably completes the speed information of the sparse trajectory, provides a prior path skeleton for the diffusion model, and improves the generation efficiency and trajectory fitting accuracy.

[0028] 3) The application designs a traffic physics constraint condition vector that integrates speed continuity, regional consistency, red light stay tendency, and turning legality, and proposes a compound loss function based on traffic physics constraints to systematically constrain the traffic physics consistency of the trajectory generation process.

[0029] 4) The conditional diffusion trajectory generation model of the application combines traffic physics constraints to guide the sampling strategy, dynamically adjusts the rule weight in the reasoning stage, realizes flexible traffic physics constraint reinforcement sampling, and effectively improves the traffic compliance and generation quality of the generated trajectory.

[0030] 5) The application has wide application prospects in the fields of intelligent transportation, trajectory completion, traffic simulation, path optimization, etc., and provides a new trajectory generation solution that takes into account trajectory recovery accuracy and traffic regulation constraints.

[0031] 6) Compared with existing methods based on statistical interpolation, pure deep learning, or context-aware generation, this invention has significant advantages in terms of the physical rationality of the generated trajectory, compliance with traffic physical constraints, and convergence speed of the generated path. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0033] Figure 1 This is a flowchart of the sparse trajectory recovery method according to an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example 1:

[0037] like Figure 1 As shown, an embodiment of the present invention provides a sparse trajectory recovery method, including:

[0038] Step S1: Collect sparse trajectory point sequences and traffic physical constraint information.

[0039] Furthermore, trajectory point data of the target vehicle is collected in real time using trajectory sampling devices (such as vehicle-mounted GPS, mobile smart terminals, traffic monitoring equipment, etc.). The collected trajectory point data is usually represented in the form of a trajectory point sequence as follows:

[0040] ;

[0041] in, and They represent the first The spatial coordinates of the trajectory points Indicates the sampling time.

[0042] Due to the low sampling frequency of the trajectory sampling device, uneven sampling interval or data loss, the obtained trajectory point sequence has obvious sparsity, and there may be a large time interval between trajectory points, and some path segments are missing.

[0043] Further, to constrain the trajectory recovery process to comply with the actual traffic physical constraints, further obtain traffic physical constraint condition information in the target area, including:

[0044] Road topology information: including road segment number, road connection relationship, road direction attribute, used to construct the feasibility network of the trajectory path.

[0045] Road turning restriction information: including the turning types allowed at each intersection (such as straight, left turn, right turn), and the forbidden turning path, forming the legality constraint of road turning.

[0046] Traffic signal light configuration information: including traffic signal light position, periodic traffic signal timing scheme, historical red light stay statistics, used to model the red light stay tendency of the trajectory point.

[0047] Spatial region division information: including road belonging region, intersection division, special control region division (such as school region, construction region), used to support region-embedded feature extraction.

[0048] The above traffic physical constraint condition information can be obtained through various ways such as road map data provided by the traffic management department, traffic control system, historical traffic signal data, etc.

[0049] Step S2, obtaining trajectory basis structure based on path inference and kinematic interpolation

[0050] For the possible path loss between sparse trajectory points, a path inference algorithm is used to estimate the feasible road connection between adjacent trajectory points. Based on the obtained road topology structure and traffic turning rules, the most likely path between adjacent trajectory points is inferred using a hidden Markov model (HMM). The path inference target is:

[0051] ;

[0052] Wherein, represents the possible path state sequence; represents the observed sparse trajectory point sequence; represents the optimal path sequence.

[0053] Through path inference, the missing connected path between trajectory points is completed, ensuring that the overall trajectory complies with the reachability of the road topology.

[0054] On the basis of path connectivity, the velocity between trajectory points is estimated by using kinematic interpolation method based on the spatio-temporal information of adjacent observation points. For the tth trajectory point, the velocity estimation formula is:

[0055] ;

[0056] wherein, is the velocity estimation value of the tth trajectory point; is the time interval of trajectory point sampling; , , and , are the spatial positions of the t-1th and t+1th trajectory points respectively.

[0057] Through path inference and velocity interpolation, the trajectory infrastructure containing spatial connectivity and velocity continuity information is obtained.

[0058] Step S3, according to the trajectory infrastructure and the traffic physical constraint condition information, the traffic physical constraint features are extracted, and then the traffic physical constraint condition vector is constructed, specifically including:

[0059] Step S301, using the velocity sequence obtained by kinematic interpolation, the velocity estimation value of each trajectory point is taken as one of the traffic physical constraint inputs. The velocity estimation value has been obtained by the following formula:

[0060] ;

[0061] Step S302, according to the regional division result of the traffic map, the road or space region to which each trajectory point belongs is identified, and the region number is obtained. For the space region where each trajectory point is located, embedding mapping is carried out by using the region number, and the region embedding vector is obtained, which represents the regional attribute of the trajectory point in the space semantic graph and can be used for modeling the transfer rationality and local consistency between regions. The region number is mapped to a continuous vector by an embedding function, which is represented as:

[0062] ;

[0063] wherein, Emb(·) represents the embedding mapping function. The region embedding vector can capture the spatial context information of the trajectory point and constrain the regional continuity of the trajectory.

[0064] Step S303, based on the historical traffic signal data, the red light stay probability of the traffic intersection is statistically analyzed, and the red light stay tendency score of the trajectory point is defined as:

[0065] ;

[0066] wherein, represents the probability of the vehicle staying at the red light at the position The red light staying tendency score is taken as a behavior constraint input to improve signal compliance in the trajectory recovery process.

[0067] Step S304, according to the road topology and traffic physical constraint condition information, it is judged whether the steering between adjacent trajectory points is legal, and the steering legality mask is defined:

[0068] ;

[0069] wherein, represents the region where the region embedding vector at the last time (t-1) is located, i.e. the region where the trajectory point is located, the region where the region embedding vector at the current time t is located, i.e. the region where the trajectory point is located.

[0070] The steering mask is used to constrain the generated trajectory to comply with the road steering regulations and avoid illegal paths.

[0071] Through steps S301-S304, the traffic physical constraint features such as speed continuity, spatial region consistency, red light staying tendency, steering legality, etc. are comprehensively extracted, and the basis for subsequent condition vector construction and trajectory generation is provided.

[0072] Step S305, the traffic physical constraint features are fused through unified normalization and feature splicing technology to construct the traffic physical constraint condition vector.

[0073] For different types of traffic physical constraint features, the following processing methods are adopted, which specifically include:

[0074] The speed estimation value is mapped to the interval [0, 1] through normalization processing to ensure numerical stability;

[0075] The spatial region number is converted into a low-dimensional vector through embedding mapping function ;

[0076] The red light staying tendency score uses the original probability value or performs normalization processing;

[0077] The steering legality mask is directly input as a binary discrete feature;

[0078] The traffic physical constraint features are combined by feature concatenation technology according to dimensions to construct a traffic physical constraint condition vector, which is defined as:

[0079] ;

[0080] Wherein: represents the normalized speed estimation value.

[0081] The feature concatenation operation ensures that various traffic physical constraint features form a unified input in the same numerical space, supporting the conditional control of the diffusion model.

[0082] Step S4, based on the traffic physical constraint condition vector, a conditional diffusion generation model is used to gradually denoise and reconstruct the sparse trajectory, to generate a time-continuous and spatially reasonable complete trajectory point sequence, specifically including:

[0083] Step S401, first, noise is injected into the target trajectory point, and the trajectory is gradually disturbed into a Gaussian distributed noise trajectory. The diffusion forward process is defined as:

[0084] ;

[0085] Wherein: represents the trajectory point in the t-th diffusion step; is a noise proportion coefficient related to the diffusion time step; represents a Gaussian distribution, is an identity matrix, with 1 on the diagonal and 0 elsewhere, representing the case where each dimension in the multi-dimensional Gaussian distribution is independent and the variance is 1, wherein ;

[0086] The forward diffusion process converts the sparse trajectory point sequence into a Gaussian noise distribution by gradually injecting noise.

[0087] Step S402, in the trajectory generation stage, the conditional diffusion model is used to gradually denoise in reverse, and the traffic physical constraint condition vector is used to guide the denoising process at each step. The trajectory generation process is defined as:

[0088] ;

[0089] Wherein: and represent the mean and variance predicted by the model respectively; is the traffic physical constraint condition vector.

[0090] The conditional vector serves as a control input at each time step to dynamically constrain the denoising process, ensuring that the generated trajectory conforms to the traffic physical rules.

[0091] Step S403, the noise trajectory is predicted and denoised by using a time sequence neural network, and the target is to accurately fit the noise term in the diffusion process. In the training process, the noise prediction loss is optimized:

[0092] ;

[0093] wherein: represents the actual noise; represents the noise term predicted by the model.

[0094] The denoising network input includes the noise trajectory, the diffusion step length and the traffic physical constraint condition vector.

[0095] Step S404, by backward iteration of the conditional diffusion model, the dense trajectory points are gradually denoised and recovered until the complete trajectory sequence is reconstructed. The backward generation process is:

[0096] ;

[0097] wherein: is the sampled Gaussian noise.

[0098] Through multi-step iteration generation, the complete trajectory point sequence satisfying the traffic physical constraints is finally obtained.

[0099] Based on the conditional diffusion generation model, a traffic physical constraint loss function is designed for the diffusion generation process, which uses traffic physical constraints to constrain the generated trajectory, and ensures that the generated trajectory conforms to the traffic behavior rules in terms of speed continuity, spatial coherence, turning legality and red light stop rationality.

[0100] The speed continuity loss is used to constrain the smooth change of the speed of adjacent trajectory points, so as to avoid unreasonable sudden speed.

[0101] Define the speed continuity loss:

[0102] ;

[0103] The spatial coherence loss is used to constrain the spatial jump between trajectory points, so as to ensure that the trajectory points move reasonably within the continuous spatial region. Define the spatial coherence loss:

[0104] ;

[0105] wherein: represents the spatial jump indicator function, when the trajectory point jumps across the non-adjacent region, otherwise 0; represents the spatial distance between the trajectory points.

[0106] Turning legality loss is used to constrain the generated trajectory to not contain illegal turning paths that violate traffic physical constraints. The turning legality loss is defined as:

[0107] ;

[0108] When the turning between trajectory points is illegal turning, the loss value will be explicitly penalized.

[0109] Red light stay reasonability loss is used to constrain trajectory points to have reasonable stay behavior in red light areas, avoiding generated trajectories ignoring traffic signals. The red light stay reasonability loss is defined as:

[0110] ;

[0111] The traffic physical constraint loss function is defined as:

[0112] ;

[0113] wherein: , , , are the weight coefficients of each traffic physical constraint loss, used to balance the importance of different loss terms.

[0114] Embodiment 2:

[0115] The embodiment of the application also provides a sparse trajectory recovery device, comprising:

[0116] A first processing module is configured to collect sparse trajectory point sequences and traffic physical constraint condition information.

[0117] A second processing module is configured to perform path inference and speed interpolation on the sparse trajectory point sequences to obtain a trajectory infrastructure containing spatial connectivity and speed continuity information.

[0118] A third processing module is configured to extract traffic physical constraint features from the trajectory infrastructure and the traffic physical constraint condition information, fuse the extracted traffic physical constraint features through uniform normalization and feature splicing technology, and construct a traffic physical constraint condition vector.

[0119] A fourth processing module is configured to use a conditional diffusion generative model to gradually denoise and reconstruct sparse trajectories based on the traffic physical constraint condition vector, and generate a complete trajectory point sequence that is time-continuous and spatially reasonable.

[0120] As an embodiment of the application, the traffic physical constraint features include: speed estimate value, spatial region code, red light stay tendency score, and turning legality mask.

[0121] As an implementation of an embodiment of the present application, a traffic physics constraint loss function is designed for a diffusion generation process of a conditional diffusion generation model, and the traffic physics constraint loss function includes a velocity continuity loss function, a spatial coherence loss function, a turning legality loss function, and a red light stop rationality loss function.

[0122] Embodiment 3

[0123] The embodiment of the present application also provides a sparse trajectory recovery system, comprising a memory and a processor, the memory has a computer program run by the processor stored thereon, and the computer program executes the sparse trajectory recovery method when run by the processor.

[0124] Embodiment 4

[0125] The embodiment of the present application also provides a storage medium, and the storage medium has a computer program stored thereon, and the computer program executes the sparse trajectory recovery method when run.

[0126] The above-described embodiments are only descriptions of the preferred modes of the present application, and do not limit the scope of the present application, and various modifications and improvements to the technical solutions of the present application made by those skilled in the art without departing from the design spirit of the present application shall fall within the protection scope of the present application defined by the claims.

Claims

1. A sparse trajectory recovery method, characterized in that, include: Collect sparse trajectory point sequences and traffic physical constraints information; Path inference and velocity interpolation are performed on sparse trajectory point sequences to obtain the basic trajectory structure containing information on spatial connectivity and velocity continuity. Based on the trajectory infrastructure and traffic physical constraints, traffic physical constraint features are extracted. The extracted traffic physical constraint features are then fused using unified normalization and feature splicing techniques to construct a traffic physical constraint vector. Based on traffic physical constraint vectors, a conditional diffusion generation model is used to gradually denoise and reconstruct sparse trajectories, generating a complete trajectory point sequence that is temporally continuous and spatially reasonable. Construct the traffic physical constraint vector, including: Step S301: Using the velocity sequence obtained by kinematic interpolation, estimate the velocity of each trajectory point. As input to traffic physics constraints, the speed estimate is obtained using the following formula: ； in, This is the velocity estimate for the t-th trajectory point; The time interval for sampling trajectory points; and These are the spatial locations of the (t-1)th and (t+1)th trajectory points, respectively. Step S302: Based on the regional division results of the traffic map, identify the road or spatial region to which each trajectory point belongs, and obtain the region number. For each spatial region where a trajectory point is located, an embedding mapping is performed using the region number to obtain a region embedding vector. The region number is mapped to a continuous vector using an embedding function, represented as follows: ； Where Emb(·) denotes the embedding mapping function; Step S303: Based on historical traffic signal data, the probability of stopping at red lights at traffic intersections is statistically analyzed, and the red light stopping tendency score of the trajectory points is defined as follows: ； in, Indicates the vehicle's location The probability of stopping at a red light; Step S304: Based on the road topology and traffic physical constraints, determine whether the turn between adjacent trajectory points is legal, and define a turn legality mask: ； in, This represents the region where the region embedding vector of the previous time step (t-1) is located, i.e., the trajectory point. The area where it is located The region where the region embedding vector is located at the current time t, i.e., the trajectory point Location; Step S305 involves fusing traffic physical constraint features using unified normalization and feature concatenation techniques to construct a traffic physical constraint condition vector. For different types of traffic physical constraint features, the following processing methods are employed: Speed ​​estimate After normalization, it is mapped to the [0,1] interval; Spatial area numbering Transform into a low-dimensional vector using the embedding mapping function. ; Red light stopping tendency score Use the original probability values ​​or perform normalization; redirection legality mask As a direct input of binary discrete features; By using feature concatenation technology, traffic physical constraint features are combined according to dimensions to construct a traffic physical constraint condition vector, defined as: ； in, This represents the normalized velocity estimate.

2. The sparse trajectory recovery method as described in claim 1, characterized in that, Traffic physical constraint features include: speed estimate, spatial area coding, red light stop tendency score, and turning legality mask.

3. The sparse trajectory recovery method as described in claim 2, characterized in that, A traffic physical constraint loss function is designed for the diffusion generation process of the conditional diffusion generation model. The traffic physical constraint loss function includes: speed continuity loss function, spatial coherence loss function, turning legality loss function, and red light stopping rationality loss function.

4. A sparse trajectory recovery apparatus for implementing the sparse trajectory recovery method of claim 1, characterized in that, include: The first processing module is used to collect sparse trajectory point sequences and traffic physical constraint information; The second processing module is used to perform path inference and velocity interpolation on the sparse trajectory point sequence to obtain the trajectory basic structure containing spatial connectivity and velocity continuity information. The third processing module is used to extract traffic physical constraint features based on the trajectory basic structure and traffic physical constraint information, and to fuse the extracted traffic physical constraint features through unified normalization and feature splicing technology to construct a traffic physical constraint vector. The fourth processing module is used to gradually denoise and reconstruct sparse trajectories based on traffic physical constraint vectors and a conditional diffusion generation model, generating a complete trajectory point sequence that is temporally continuous and spatially reasonable.

5. The sparse trajectory recovery device as described in claim 4, characterized in that, Traffic physical constraint features include: speed estimate, spatial area coding, red light stop tendency score, and turning legality mask.

6. The sparse trajectory recovery device as described in claim 5, characterized in that, A traffic physical constraint loss function is designed for the diffusion generation process of the conditional diffusion generation model. The traffic physical constraint loss function includes: speed continuity loss function, spatial coherence loss function, turning legality loss function, and red light stopping rationality loss function.

7. A sparse trajectory recovery system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the sparse trajectory recovery method as described in any one of claims 1-3 when executed by the processor.

8. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, performs the sparse trajectory recovery method as described in any one of claims 1-3.

Images