A method for estimating travel time of spatiotemporal trajectories based on spatiotemporal aggregation of contextual topological graphs

Abstract

The present invention belongs to the field of vehicle travel time estimation, and provides a method for estimating the travel time of a spatiotemporal trajectory based on spatiotemporal aggregation of a context topology map, including: 1. constructing a camera topology map based on time series information; 2. constructing a spatiotemporal trajectory travel time estimation network taking into account the context topology map, aggregating spatiotemporal attribute information, performing graph convolution and time series convolution learning on the camera topology map output in step 1, jointly learning local trajectory time prediction and global trajectory time prediction, and obtaining network parameters for estimating the travel time of a spatiotemporal trajectory taking into account the context topology map; 3. using the spatiotemporal trajectory travel time estimation network taking into account the context topology map trained in step 2, inputting the starting camera ID of the vehicle's driving trajectory, and predicting the time required for the vehicle to pass through the trajectory. The method of the present invention takes into account the topological relationship between cameras, and realizes the spatiotemporal trajectory travel time estimation based on sparse camera positions.

Description

Technical Field

[0001] The invention belongs to the field of vehicle travel time estimation, and in particular relates to a method for estimating space-time trajectory travel time based on space-time aggregation of context topology graphs. Background Art

[0002] The estimation of travel time based on spatiotemporal trajectory is a key link in tasks such as traffic road planning, urban functional area division, and online car-hailing dispatch. Existing research on spatiotemporal trajectory mainly focuses on modeling, map matching, and pattern mining based on GPS trajectory data. However, in actual applications, it is usually difficult to directly obtain the GPS driving trajectories of non-taxis and online car-hailing vehicles. The vehicle travel time can only be estimated by building a spatiotemporal trajectory model based on the spatiotemporal information of the cameras along the way. The travel estimation method generally used for dense GPS sampling point data is difficult to directly apply to camera trajectory data. Therefore, it is urgent to explore camera-based spatiotemporal trajectory travel time estimation methods. Summary of the invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a method for estimating the travel time of a spatiotemporal trajectory based on spatiotemporal aggregation of a context topology graph, which effectively improves the accuracy and robustness of vehicle travel time estimation.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions.

[0005] The present invention provides a method for estimating the travel time of a spatiotemporal trajectory based on spatiotemporal aggregation of a context topology graph, comprising the following steps:

[0006] Step 1: construct a camera topology map based on time series information, calculate the time interval histogram of the same vehicle passing through every two cameras, then use the Parzen window to smooth the time interval histogram to make the discrete probability value continuous, and then represent the cameras as a graph structure according to the path probability between cameras to obtain the camera topology map;

[0007] Step 2: construct a spatiotemporal trajectory travel time estimation network that takes into account the context topology map. By aggregating spatiotemporal attribute information (such as distance, vehicle speed, etc.), the camera topology map output in step 1 is subjected to graph convolution and temporal convolution learning. Then, local trajectory time prediction and global trajectory time prediction are jointly learned to obtain the spatiotemporal trajectory travel time estimation network parameters that take into account the context topology map.

[0008] Step 3: Use the spatiotemporal trajectory travel time estimation network trained in step 2 taking into account the context topology graph, input the starting camera ID of the vehicle's driving trajectory, and predict the time required for the vehicle to pass through the trajectory.

[0009] In the above technical solution, the step 1 specifically includes:

[0010] (1-1) Camera path time histogram calculation: The camera path time distribution is calculated based on the captured vehicle ID, camera ID and shooting timestamp;

[0011] (1-2) Histogram smoothing: using the Parzen window to smooth the path time-frequency graph between different cameras obtained in step (1-1), and converting the discrete frequency data into continuous probability;

[0012] (1-3) Find neighboring camera nodes. For any camera, calculate the probability set of other cameras arriving within a certain time interval and the probability set of all other cameras arriving within a certain time interval based on the smoothed path time probability. Select the camera node with a probability greater than 90% in the two probability sets as the neighbor node of the camera.

[0013] (1-4) Add edges to the camera nodes. For a path starting from a camera, add a directed edge from the node to the end of the path. Conversely, add a directed edge in the opposite direction. The weight of the edge between two nodes is set to the arrival probability of the two camera nodes in time. These weights constitute the adjacency matrix between camera nodes, and the camera topology map is obtained.

[0014] In the above technical solution, the step 2 specifically includes:

[0015] (2-1) Construct a spatiotemporal attribute aggregation module to aggregate the distance between two cameras, vehicle speed and other attributes as the edge features between the two camera nodes, and aggregate all edge features connecting camera nodes as the camera node features;

[0016] (2-1-1) Aggregate the features of the edges between nodes. First, we need to obtain the geographical location of the camera. Cameras with GPS locations directly use the GPS location. Cameras without GPS use the map matching method to obtain the actual longitude and latitude. The average speed and distance between two nodes at time t are mapped in a high-dimensional manner as the edge features.

[0017] (2-1-2) Aggregate node features, obtain the geocode of the camera location by dividing the grid, calculate the coverage area of ​​all cameras, determine the side length of the grid by making each camera fall into a unique grid, and use the row and column number of the grid where the camera is located as the geocode of the camera. Aggregate all edge features and geocodes connected to the node, and then perform high-dimensional mapping to obtain the camera node features;

[0018] (2-2) Construct a spatiotemporal graph convolution module and a multi-scale aggregation module to calculate the single-scale spatiotemporal correlation and multi-scale spatiotemporal correlation of the camera topology subgraph;

[0019] (2-2-1) Construct a spatiotemporal graph convolution module, which consists of two parts: graph convolution GCN and temporal convolution TCN. After inputting the camera topology subgraph, the GCN layer is responsible for aggregating spatial features. The features aggregated by GCN are input into the TCN part to learn temporal correlation.

[0020] (2-2-2) Construct a multi-scale aggregation module, use the gated recurrent unit GRU to aggregate multi-scale information, and obtain global and local spatiotemporal correlations at the same time;

[0021] (2-3) Joint task learning: For a given trajectory, the topological subgraph corresponding to all cameras on the trajectory is input, and the GRU unit is used for feature transfer to jointly optimize the local trajectory travel time and the entire trajectory travel time;

[0022] (2-3-1) Local trajectory travel time prediction: add the potential features of the two cameras connected to the trajectory and input them into the local prediction layer to obtain the travel time of the trajectory. The local prediction layer consists of two fully connected layers;

[0023] (2-3-2) Predict the travel time of the entire trajectory. The potential features of all cameras on the trajectory are summed and pooled to aggregate the features of the entire trajectory, and then input into the global prediction layer to obtain the travel time of the entire trajectory. The global prediction layer consists of two fully connected layers.

[0024] (2-3-3) Model training, using MAPE loss function to simultaneously optimize the local trajectory travel time and the entire trajectory travel time.

[0025] In the above technical solution, the step 3 specifically includes:

[0026] (3-1) Input the starting camera ID of the vehicle's path trajectory into the spatiotemporal trajectory travel time estimation network that takes into account the contextual topology graph trained in step 2 to obtain the entire travel time and segmented local travel time of the vehicle's path along the trajectory;

[0027] (3-2) The segmented local travel time is summed up to obtain the cumulative travel time, and the average of the cumulative travel time and the entire travel time is calculated to obtain the travel time of the vehicle passing through the trajectory.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. A camera topology map was constructed to constrain the space-time trajectory, which can improve the accuracy of the space-time trajectory travel time estimation.

[0030] 2. Multi-scale spatiotemporal aggregation and joint task learning are designed to capture the spatiotemporal correlation of independent camera topological subgraphs and multi-scale spatiotemporal correlation, which can be better adapted to various traffic conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a network architecture diagram for estimating the spatiotemporal trajectory travel time based on spatiotemporal aggregation of contextual topology graphs according to the present invention.

[0032] Figure 2 The present invention is implemented in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is further described below through embodiments and in conjunction with the accompanying drawings.

[0034] Figure 1 The network architecture diagram of the spatiotemporal trajectory travel time estimation based on spatiotemporal aggregation of context topology map proposed by the present invention includes: camera topology map construction based on timing information; spatiotemporal trajectory travel time estimation network taking into account the context topology map.

[0035] like Figure 2 As shown, a method for estimating the travel time of a spatiotemporal trajectory based on spatiotemporal aggregation of a context topology graph provided by an embodiment of the present invention mainly comprises the following steps:

[0036] Step 1: Construct a camera topology map based on time series information, calculate the time interval histogram of the same vehicle passing through every two cameras, and then use the Parzen window to smooth the time interval histogram to make the discrete probability values ​​continuous. Then, represent the cameras as a graph structure according to the path probability between cameras to obtain the camera topology map.

[0037] Step 2: construct a spatiotemporal trajectory travel time estimation network that takes into account the contextual topological map. By aggregating spatiotemporal attribute information (such as distance, vehicle speed, etc.), perform graph convolution and temporal convolution learning on the camera topological map output in step 1, and then jointly learn local trajectory time prediction and global trajectory time prediction to obtain the spatiotemporal trajectory travel time estimation network parameters that take into account the contextual topological map.

[0038] Step 3: Use the spatiotemporal trajectory travel time estimation network trained in step 2 taking into account the context topology graph, input the starting camera ID of the vehicle's driving trajectory, and predict the time required for the vehicle to pass through the trajectory.

[0039] Next, this method is described in detail.

[0040] The step 1 specifically comprises the following steps:

[0041] (1-1) Camera path time histogram calculation: The camera path time distribution is calculated based on the captured vehicle ID, camera ID, and shooting timestamp. The camera path time histogram calculation method is as follows:

[0042]

[0043]

[0044] Where I, c, and f represent the vehicle ID, camera ID, and shooting timestamp, respectively. Δ represents a fixed time interval. represents the number of vehicles in the kth histogram from camera i to camera j, Represents the sum of the number of all histograms. Corresponding to the path time distribution of camera i, the center point of the time interval is 0. The time interval on the left is less than 0, indicating that the vehicle is driving from camera i to the corresponding camera, and the time interval on the right is greater than 0, indicating that the vehicle is driving from other cameras to camera i.

[0045] (1-2) Histogram smoothing: Use the Parzen window to smooth the path time frequency graph between different cameras obtained in step (1-1) to convert the discrete frequency data into continuous probability. The smoothing process is calculated as follows:

[0046] P c (I i =I j |k,c i ,c j )=P c (I i =I j |k,c i ,c j )K(k)

[0047]

[0048] Among them, K is the Gaussian kernel, σ is a constant, and after smoothing, the path time probability will no longer have an obvious peak.

[0049] (1-3) Find neighboring camera nodes. For any camera, calculate the probability set of other cameras arriving within a certain time interval and the probability set of all other cameras arriving within a certain time interval based on the smoothed path time probability. Select the camera node with a probability greater than 90% in the two probability sets as the neighbor node of the camera.

[0050] (1-4) Add edges to camera nodes. For a path starting from a camera, add a directed edge from the node to the end of the path. Conversely, add a directed edge in the opposite direction. The weight of the edge between two nodes is set to the arrival probability of the two camera nodes in time. These weights constitute the adjacency matrix between camera nodes. For any camera c i , whose camera topology subgraph is represented as G i =(V i ,E i), where V i represents a set of nodes, E i represents a set of directed edges. Specifically, Represents a node in a topological subgraph, Represents the slave node To Node A i It is the adjacency matrix of the topological subgraph, with a size of n×n, where n represents the total number of camera nodes.

[0051] The step 2 specifically includes the following steps:

[0052] (2-1) Construct a spatiotemporal attribute aggregation module to aggregate the distance between two cameras, vehicle speed and other attributes as the features of the edge between the two camera nodes, and aggregate all edge features connecting the camera nodes as the features of the camera node.

[0053] (2-1-1) Aggregate the features of the edges between nodes. First, we need to obtain the geographical location of the camera. Cameras with GPS locations use the GPS location directly. Cameras without GPS use the map matching method to obtain the actual longitude and latitude. The average speed and distance between two nodes at time t are mapped in a high-dimensional manner as the edge features. The feature calculation method for each edge at time t is as follows:

[0054] e(t)=tanh(W a ·[s e (t),d e ])

[0055] Among them, s e (t) represents the average speed of the road section at time t, d e Indicates the distance of the road section, W a represents the learnable weights and [·] represents the dimension-wise concatenation operation.

[0056] (2-1-2) Aggregate node features and obtain the geocode of the camera location by dividing the grid. Calculate the coverage area of ​​all cameras, determine the side length of the grid by making each camera fall into a unique grid, and use the row and column number of the grid where the camera is located as the geocode of the camera. Aggregate all edge features and geocodes connected to the node, and then perform high-dimensional mapping to obtain the camera node features. The calculation method of camera node features is as follows:

[0057]

[0058] Among them, W b Represents the learnable weight. Through the above operations, each camera node achieves the aggregation of spatiotemporal attributes.

[0059] (2-2) Construct a spatiotemporal graph convolution module and a multi-scale aggregation module to calculate the single-scale spatiotemporal correlation and multi-scale spatiotemporal correlation of the camera topology subgraph.

[0060] (2-2-1) Construct a spatiotemporal graph convolution module, which consists of two parts: graph convolution GCN and temporal convolution TCN. After inputting the camera topology subgraph, the GCN layer is responsible for aggregating spatial features. The features aggregated by GCN are input to the TCN part to learn temporal correlation. The specific calculation method of the GCN layer aggregating spatial features is as follows:

[0061] Z (l+1) =ReLU(L·Z (l) ·W (l) )

[0062] in, A is a positive definite adjacency matrix, I is the identity matrix, Z (l) Represents the features learned in the lth layer. TCN uses a gating mechanism, and the specific calculation process is as follows:

[0063]

[0064] Among them, θ1, θ2 are learnable parameters of temporal convolution, and ⊙ represents element-by-element multiplication. i represents the potential features of camera node i, and TCN is applied to each node in the camera topology subgraph.

[0065] (2-2-2) Construct a multi-scale aggregation module, use the gated recurrent unit GRU to aggregate multi-scale information, and obtain global and local spatiotemporal correlations at the same time. The specific calculation process is as follows:

[0066] c (0) =G(r (0) ,c (-1) ),

[0067] r (1) =T(S(r (0) )),

[0068] r (2) =T(S([r (0) ,r (1) ])),

[0069] r (l+1) =T(S([G(r (l-1) ,c (l-2) ),r (l) ])),

[0070] l=2,3,4,...

[0071] Among them, T represents TCN aggregation operation, S represents GCN aggregation operation, r(l) represents the latent features learned at layer l. (l) represents the hidden state of the topological subgraph after processing, c (0) The initial value is 0.

[0072] G stands for GRU operation. Specifically, GRU is defined as follows:

[0073] u=tanh(W u1 r(t)+U u1 c(t-1)+b u1 ),

[0074] x = tanh(W x1 r(t)+U x1 c(t-1)+b x1 ),

[0075] c′(t)=tanh(W h1 r(t)+U h1 (x⊙c(t-1))+b h1 ),

[0076] c(t)=u⊙r(t-1)+(1-u)⊙c′(t)

[0077] Among them, W u1 ,U u1 ,W x1 ,U x1 ,W h1 ,U h1 represents the learnable parameter, b u1 ,b x1 ,b h1 Indicates the offset.

[0078] (2-3) Joint task learning: For a given trajectory, the topological subgraph corresponding to all cameras on the trajectory is input, and the GRU unit is used for feature transfer to jointly optimize the local trajectory travel time and the entire trajectory travel time.

[0079] (2-3-1) Local trajectory travel time prediction: add the potential features of the two cameras connected to the trajectory and input them into the local prediction layer to obtain the travel time of the trajectory. The local prediction layer consists of two fully connected layers.

[0080] (2-3-2) Predict the travel time of the entire trajectory. The potential features of all cameras on the trajectory are summed and pooled to aggregate the features of the entire trajectory, and input into the global prediction layer to obtain the travel time of the entire trajectory. The global prediction layer consists of two fully connected layers.

[0081] (2-3-3) Model training uses the MAPE loss function to optimize the local trajectory travel time and the entire trajectory travel time at the same time. The specific loss function is as follows:

[0082]

[0083] Among them, P represents the number of cameras included in the entire trajectory, t whole , t ij , They represent the entire trajectory travel time, the entire trajectory travel estimated time, the local trajectory travel time and the local trajectory travel estimated time respectively.

[0084] The step 3 specifically comprises the following steps:

[0085] (3-1) Input the starting camera ID of the vehicle's path trajectory into the spatiotemporal trajectory travel time estimation network trained in step (2) taking into account the contextual topology graph to obtain the entire travel time and segmented local travel time of the vehicle's path.

[0086] (3-2) The segmented local travel time is summed up to obtain the cumulative travel time, and the average of the cumulative travel time and the entire travel time is calculated to obtain the travel time of the vehicle passing through the trajectory.

[0087] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

[0088] Although the present invention is described in detail and embodiments are given above, the present invention and applicable embodiments are not limited thereto, and those skilled in the art can make various modifications based on the principles of the present invention, and can also apply some methods in the present invention to other systems. Therefore, all modifications made according to the principles of the present invention should be understood to fall within the protection scope of the present invention.

Claims

1. A method for estimating travel time of spatiotemporal trajectories based on spatiotemporal aggregation of contextual topological graphs, characterized in that The method comprises the following steps: (1) Construct a camera topology map based on time series information, calculate the time interval histogram of the same vehicle passing through every two cameras, then use the Parzen window to smooth the time interval histogram to make the discrete probability value continuous, and then represent the cameras as a graph structure according to the path probability between cameras to obtain the camera topology map; specifically, it includes the following sub-steps: (1-1) Camera path time histogram calculation: The camera path time distribution is calculated based on the captured vehicle ID, camera ID and shooting timestamp; (1-2) Histogram smoothing: using the Parzen window to smooth the path time-frequency graph between different cameras obtained in step (1-1), and converting the discrete frequency data into continuous probability; (1-3) Find neighboring camera nodes. For any camera, calculate the probability set of other cameras arriving within a certain time interval and the probability set of all other cameras arriving within a certain time interval based on the smoothed path time probability. Select the camera node with a probability greater than 90% in the two probability sets as the neighbor node of the camera. (1-4) Add edges to camera nodes. For a path starting from a camera, add a directed edge from the node to the end of the path. Conversely, add a directed edge in the opposite direction. The weight of the edge between two nodes is set to the arrival probability of the two camera nodes within a certain time. These weights constitute the adjacency matrix between camera nodes, and the camera topology map is obtained. (2) constructing a spatiotemporal trajectory travel time estimation network that takes into account the context topology map, by aggregating spatiotemporal attribute information, performing graph convolution and temporal convolution learning on the camera topology map output in step (1), and then jointly learning local trajectory time prediction and global trajectory time prediction to obtain the spatiotemporal trajectory travel time estimation network parameters that take into account the context topology map; specifically, the following sub-steps are included: (2-1) Construct a spatiotemporal attribute aggregation module to aggregate the spatiotemporal attributes between two cameras, including distance and vehicle speed, as the features of the edge between the two camera nodes, and aggregate all edge features connecting the camera nodes as the features of the camera node; (2-1-1) Aggregate the features of the edges between nodes. First, we need to obtain the geographical location of the camera. Cameras with GPS locations directly use the GPS location. Cameras without GPS use the map matching method to obtain the actual longitude and latitude. The average speed and distance between two nodes at time t are mapped in a high-dimensional manner as the edge features. (2-1-2) Aggregate node features, obtain the geographic code of the camera location by dividing the grid, calculate the coverage area of ​​all cameras, determine the side length of the divided grid by making each camera fall into a unique grid, use the row and column number of the grid where the camera is located as the geocode of the camera, aggregate all edge features and geocodes connecting the node, and then perform high-dimensional mapping to obtain the camera node features; (2-2) Construct a spatiotemporal graph convolution module and a multi-scale aggregation module to calculate the single-scale spatiotemporal correlation and multi-scale spatiotemporal correlation of the camera topology subgraph; (2-2-1) Construct a spatiotemporal graph convolution module, which consists of two parts: graph convolution GCN and temporal convolution TCN. After inputting the camera topology subgraph, the GCN layer is responsible for aggregating spatial features. The features aggregated by GCN are input into the TCN part to learn temporal correlation. (2-2-2) Construct a multi-scale aggregation module, use the gated recurrent unit GRU to aggregate multi-scale information, and obtain global and local spatiotemporal correlations at the same time; (2-3) Joint task learning: For a given trajectory, the topological subgraph corresponding to all cameras on the trajectory is input, and the GRU unit is used for feature transfer to jointly optimize the local trajectory travel time and the entire trajectory travel time; (2-3-1) Local trajectory travel time prediction: add the potential features of the two cameras connected to the trajectory and input them into the local prediction layer to obtain the travel time of the trajectory. The local prediction layer consists of two fully connected layers; (2-3-2) Predict the travel time of the entire trajectory. The potential features of all cameras on the trajectory are summed and pooled to aggregate the features of the entire trajectory, and then input into the global prediction layer to obtain the travel time of the entire trajectory. The global prediction layer consists of two fully connected layers. (2-3-3) Model training, using MAPE loss function to optimize the local trajectory travel time and the entire trajectory travel time simultaneously; (3) Using the spatiotemporal trajectory travel time estimation network that takes into account the contextual topology graph trained in step (2), the starting camera ID of the vehicle's driving trajectory is input to predict the time required for the vehicle to pass through the trajectory.

2. The method for estimating the travel time of a spatiotemporal trajectory based on spatiotemporal aggregation of a context topology graph according to claim 1, characterized in that: The "using the spatiotemporal trajectory travel time estimation network taking into account the context topology graph trained in step (2), inputting the starting camera ID of the vehicle's driving trajectory, and predicting the time required for the vehicle to pass through the trajectory" includes the following sub-steps: (3-1) Inputting the starting camera ID of the vehicle's path trajectory into the spatiotemporal trajectory travel time estimation network trained in step (2) taking into account the contextual topology graph to obtain the entire travel time and segmented local travel time of the vehicle's path along the trajectory; (3-2) The segmented local travel time is summed up to obtain the cumulative travel time, and the average of the cumulative travel time and the entire travel time is calculated to obtain the travel time of the vehicle passing through the trajectory.

Images