Low-latency Real-time Traffic Prediction Cloud Platform for Coupling Internet of Things and Edge Computing Devices

Abstract

The present invention discloses a low-latency real-time traffic prediction cloud platform coupled with the Internet of Things and edge computing devices. The present invention uses image capture devices on the road to capture vehicle flow images, and then performs license plate positioning and recognition on them through edge computing devices at the local end, so as to obtain vehicle flow information passing through different positions, thereby greatly reducing the amount of data that needs to be uploaded to the cloud platform, improving processing efficiency and reducing the load on the cloud. Moreover, since the acquisition of the vehicle information is only carried out at the local end, a faster response speed can be provided, and the latency for the cloud platform to obtain real-time vehicle flow information can be reduced. The present invention can aggregate all vehicle flow trajectory information within the entire prediction area on the cloud platform, and predict the traffic flow at future moments through the traffic flow prediction model carried on the cloud platform.

Description

[0001] This application is a divisional application of the invention patent with application number 202111623527.6, application date December 28, 2021, and invention name "A traffic flow prediction system and cloud platform based on the Internet of Things and edge computing". Technical Field

[0002] The present invention relates to the field of traffic flow prediction, and in particular to a traffic flow prediction system and a cloud platform based on the Internet of Things and edge computing. Background Art

[0003] With the continuous acceleration of urbanization, traffic congestion in cities is becoming more and more serious. Therefore, traffic flow prediction is an important part of smart transportation. However, in existing technologies, traffic flow prediction often cannot use real-time data, so its timeliness and application feasibility are flawed.

[0004] For example, in the invention patent with application number CN201810603991.0, a method for predicting short-term urban traffic flow based on the spatiotemporal similarity of traffic flow is disclosed. The method includes the following steps: S1. Based on the spatiotemporal similarity of traffic flow, define the time state vector and spatiotemporal state vector of traffic flow; S2. Construct the "current spatiotemporal state vector" of traffic flow in the current period; S3. Construct the "historical spatiotemporal state vector" of traffic flow in the same period on different dates in history; S4. Use the distance metric function to calculate the "spatiotemporal similarity distance" between the current and each historical spatiotemporal state vector. ; S5. Select the date where k historical state vectors with the smallest spatiotemporal similarity distance are located, and find out the traffic flow of the predicted period corresponding to these k historical dates; S6. Based on the traffic flow of the predicted period corresponding to these k historical dates, use the prediction function to calculate the traffic flow of the target road section in the next period; S7. According to the predicted results and actual results of the traffic flow, the prediction error of the target road section is evaluated and analyzed. The historical data in this scheme comes from the floating car data of taxis, and its data samples cannot represent all vehicles, and such data often have data quality problems caused by signal reasons.

[0005] Therefore, how to improve the practical applicability of traffic flow prediction systems is a technical problem that needs to be solved urgently. Summary of the invention

[0006] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a traffic flow prediction system and cloud platform based on the Internet of Things and edge computing, which can effectively solve the above-mentioned problems.

[0007] The technical solutions specifically adopted in the present invention are as follows:

[0008] In a first aspect, the present invention provides a traffic flow prediction system based on the Internet of Things and edge computing, which includes a cloud platform and image capture devices and edge computing devices installed at different locations on the road;

[0009] The image capture device is used to capture images of passing vehicles in real time, and the driving direction of the vehicles captured by the same image capture device is fixed;

[0010] The edge computing device is paired with the image capture device one by one to obtain the traffic flow image captured by the image capture device at the same location, and locate each license plate area in the traffic flow image through the built-in license plate positioning model, and then recognize the license plate number in each license plate area through the license plate recognition model;

[0011] The cloud platform is connected to edge computing devices at different locations through the Internet of Things, and is internally provided with a data receiving module, a data fusion module, and a traffic flow prediction module;

[0012] The data receiving module is used to receive in real time the license plate number recognition results reported by edge computing devices at different locations, as well as the corresponding timestamp and vehicle driving direction;

[0013] The data fusion module is used to associate and fuse each license plate number reported by each edge computing device with the corresponding timestamp, the vehicle driving direction and the coordinates of the edge computing device to form a trajectory point. All continuous trajectory points of each license plate number are restored to the driving trajectory of the corresponding vehicle by calling the path planning algorithm, and the driving trajectories of all vehicles are stored as traffic data in the historical traffic database;

[0014] The traffic flow prediction module is used to read the stored traffic flow data from the historical traffic flow database, and predict the traffic flow at a future moment based on the trained traffic flow prediction model.

[0015] Preferably, the image capture device and the edge computing device are installed in pairs at the intersection of the road.

[0016] Preferably, the license plate positioning model is a YOLO model.

[0017] Preferably, the license plate recognition model is a CNN convolutional neural network model.

[0018] Preferably, the traffic flow prediction module is provided with a designated module for inputting a prediction area and a prediction time.

[0019] Preferably, the traffic flow prediction model is a multi-direction traffic flow prediction model, which includes a first fully connected neural network, a second fully connected neural network, a three-dimensional residual convolutional network, and a recalibration layer. The input of the multi-direction traffic flow prediction model is a three-dimensional traffic flow matrix, a time signal vector, and a point of interest signal. The first fully connected neural network outputs a time signal matrix according to the time signal vector, the second fully connected neural network outputs a point of interest signal matrix according to the point of interest signal, and the three-dimensional residual convolutional network outputs a result matrix according to the fusion features of the three-dimensional traffic flow matrix, the point of interest signal matrix, and the time signal matrix. Finally, after the result matrix undergoes a weighted compression operation in the recalibration layer, a multi-direction traffic flow prediction result is obtained.

[0020] Preferably, both the license plate positioning model and the license plate recognition model are pre-trained and then downloaded to the edge computing device.

[0021] Preferably, the traffic flow prediction model in the cloud platform is continuously trained using the incremental learning method.

[0022] Preferably, the image capture device is a camera installed above the intersection, and each camera only shoots in the direction of the traffic flow.

[0023] Preferably, the generation methods of the three-dimensional traffic flow matrix, the time signal vector, and the point of interest signal are as follows:

[0024] S1. Obtain the historical traffic flow data before the prediction time in the area to be predicted. The historical traffic flow data includes the positions of different vehicles at different times in the area to be predicted and the driving directions of the vehicles; extract a number of traffic flow data time slices from the historical traffic flow data at a preset time slice interval;

[0025] S2. Perform grid division on the area to be predicted to divide it into a series of grids. The vehicles in each traffic flow data time slice are mapped to the corresponding grids in the area to be predicted according to their coordinates, and the driving direction of the vehicle is defined as the movement state of the vehicle in the grid. The movement states include up, down, left, and right; count the total number of vehicles in each movement state contained in each grid in each time slice and use it as the grid value, and map the grid value to a matrix element, so as to construct a two-dimensional traffic flow matrix for each movement state in each time slice respectively. The two-dimensional traffic flow matrices of each movement state in all time slices are stacked according to the time dimension to form a three-dimensional traffic flow matrix;

[0026] S3. Extract the hour field and the minute field from the prediction time, and splice them to form a binary time signal vector;

[0027] S4. Obtain the spatial geographical locations of all points of interest (POIs), map the POIs of different functional categories into the grids of the area to be predicted, count the total number of POIs of each group of functional categories in each grid and use it as the grid value, map the grid value into a matrix element, so as to construct a POI slice in the form of a two-dimensional matrix for each group of functional categories of POIs respectively, and the POI slices of all functional categories are superimposed to form a POI signal in the form of a three-dimensional tensor.

[0028] In a second aspect, the present invention provides a cloud platform, which is used to cooperate with image capture devices and edge computing devices installed at different positions on the road to realize traffic flow prediction;

[0029] The image capture device is used to capture the passing vehicle flow images in real time, and the driving direction of the vehicles captured by the same image capture device is fixed;

[0030] The edge computing device is paired with the image capture device one by one, and is used to obtain the vehicle flow images captured by the image capture device at the same position, locate each license plate area in the vehicle flow image through the built-in license plate positioning model, and then identify the license plate number in each license plate area through the license plate recognition model;

[0031] The cloud platform is communicatively connected to edge computing devices at different positions through the Internet of Things, and is internally provided with a data receiving module, a data fusion module and a traffic flow prediction module;

[0032] The data receiving module is used to receive in real time the license plate number recognition results reported by edge computing devices at different positions, as well as the corresponding timestamps and vehicle driving directions;

[0033] The data fusion module is used to associate and fuse each license plate number reported by each edge computing device with the corresponding timestamp, vehicle driving direction and the coordinates of the edge computing device to form a trajectory point. All consecutive trajectory points of each license plate number are used to restore the driving trajectory of the corresponding vehicle by calling the path planning algorithm, and the driving trajectories of all vehicles are stored in the historical vehicle flow database as vehicle flow data;

[0034] The traffic flow prediction module is used to read the stored vehicle flow data from the historical vehicle flow database and predict the traffic flow at a future moment based on the trained traffic flow prediction model.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention uses image capture devices on roads to capture vehicle flow images, and then performs license plate positioning and recognition on them through edge computing devices at the local end, so as to obtain vehicle flow information passing through different positions, thereby greatly reducing the amount of data that needs to be uploaded to the cloud platform, improving processing efficiency and reducing the load on the cloud. Moreover, since the acquisition of the vehicle information is only carried out at the local end, a faster response speed can be provided, and the delay of the cloud platform in obtaining real-time vehicle flow information can be reduced. The present invention can aggregate all vehicle flow trajectory information within the entire prediction area on the cloud platform, and predict the traffic flow at future moments through the traffic flow prediction model carried on the cloud platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of a traffic flow prediction system based on the Internet of Things and edge computing;

[0038] Figure 2 It is a diagram of the module composition in the cloud platform. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0040] On the contrary, the present invention covers any alternatives, modifications, equivalent methods and solutions made within the spirit and scope of the present invention defined by the claims. Further, in order to enable the public to better understand the present invention, some specific details are described in detail in the following detailed description of the present invention. Those skilled in the art can fully understand the present invention without the description of these details.

[0041] As a preferred implementation form of the present invention, a traffic flow prediction system based on the Internet of Things and edge computing is provided, which includes a cloud platform, as well as image capture devices and edge computing devices installed at different positions on the road.

[0042] Among them, the edge computing devices are paired with the image capture devices one by one. Each pair of image capture devices and edge computing devices are connected by signal lines and installed at a position on the road. For the convenience of image capture, the edge computing devices and image capture devices are preferably installed at road intersections.

[0043] The function of the image capture device is to capture the passing vehicle flow images in real time, and the driving direction of the vehicles captured by the same image capture device is fixed. In addition, the function of the edge computing device is to obtain the vehicle flow images captured by the image capture device at the same position, and locate each license plate area in the vehicle flow image through the built-in license plate positioning model, and then identify the license plate number in each license plate area through the license plate recognition model.

[0044] In practical applications, the image capture device can directly use the cameras installed above the intersection, and each camera is only oriented towards the direction of vehicle flow for shooting. Since the driving direction of the vehicles captured by each image capture device is fixed, the driving direction of the vehicles recognized from the images captured by the image capture device is also fixed. In the present invention, the license plate positioning model and the license plate recognition model can be implemented using any network model capable of achieving license plate positioning and license plate recognition. For example, the license plate positioning model can use the YOLO model, preferably the YOLO V3 model, and the license plate recognition model can use the CNN convolutional neural network model. Both the license plate positioning model and the license plate recognition model need to be pre-trained and then downloaded to the edge computing device.

[0045] Since the edge computing device has processed the image data locally, it only needs to send the license plate number data to the cloud platform, which greatly reduces the amount of network uplink data, and thus can improve the real-time performance of the cloud platform for data acquisition.

[0046] In addition, the cloud platform is communicatively connected to edge computing devices at different locations through the Internet of Things, and is internally provided with a data receiving module, a data fusion module, and a traffic flow prediction module.

[0047] Among them, the data receiving module is used to receive in real time the license plate number recognition results reported by edge computing devices at different locations, as well as the corresponding timestamps and vehicle driving directions.

[0048] In practical applications, the vehicle driving direction can be determined according to the interface or ID of the image capture device and the edge computing device from which it comes, and each edge computing device can pre-store the corresponding vehicle driving direction in the cloud platform.

[0049] Among them, the data fusion module is used to associate and fuse each license plate number reported by each edge computing device with the corresponding timestamp, vehicle driving direction, and the coordinates of the edge computing device to form a trajectory point. All consecutive trajectory points of each license plate number are used to restore the driving trajectory of the corresponding vehicle by calling the path planning algorithm, and the driving trajectories of all vehicles are stored as traffic flow data in the historical traffic flow database.

[0050] In the present invention, the path planning algorithm can be any algorithm capable of generating the path trajectory of a vehicle based on all consecutive trajectory points of the vehicle. Preferably, the Dijkstra algorithm is used. Of course, it is also possible to directly call map APIs such as Baidu Map or Gaode Map, and use each trajectory point as a waypoint to generate the path trajectory. And during the process of generating this path trajectory, its time information also needs to be carried, that is, the remaining trajectory points between any two trajectory points on a path can generate the corresponding time information through interpolation.

[0051] Among them, the traffic flow prediction module is used to read the stored traffic flow data from the historical traffic flow database and predict the traffic flow at future moments based on the trained traffic flow prediction model.

[0052] The traffic flow prediction model in the cloud platform needs to be pre-trained before use. In order to ensure the accuracy of the model, the data continuously stored in the cloud platform can also be used as samples, and the incremental learning method can be adopted for continuous training.

[0053] In order to consider different traffic flow prediction requirements, a specified module for inputting the prediction area and prediction time can be provided in the traffic flow prediction module, so as to input different prediction areas and different prediction times as needed.

[0054] In the present invention, the adopted traffic flow prediction model can be any network model capable of realizing traffic flow prediction, such as spatio-temporal graph neural network, etc.

[0055] As a preferred embodiment of the present invention, the above traffic flow prediction model can adopt a multi-directional traffic flow prediction model. This model can distinguish directions after rasterizing vehicle trajectories and distinguish moving states based on the directions when the trajectories pass through the grids, so as to realize multi-directional traffic flow prediction. The multi-directional traffic flow prediction model includes a first fully connected neural network, a second fully connected neural network, a three-dimensional residual convolution network, and a recalibration layer. The input of the multi-directional traffic flow prediction model is a traffic flow three-dimensional matrix, a time signal vector, and a point of interest signal. The first fully connected neural network outputs a time signal matrix according to the time signal vector, the second fully connected neural network outputs a point of interest signal matrix according to the point of interest signal, the three-dimensional residual convolution network outputs a result matrix according to the fusion features of the traffic flow three-dimensional matrix, the point of interest signal matrix, and the time signal matrix, and finally, after passing through a weighted compression operation in the recalibration layer for the result matrix, a multi-directional traffic flow prediction result is obtained.

[0056] In the above cloud platform, the method for traffic flow prediction using the multi-directional traffic flow prediction model includes the following steps:

[0057] S1. Obtain the historical traffic flow data before the prediction time in the area to be predicted. The historical traffic flow data includes the positions and driving directions of different vehicles in the area to be predicted at different times; extract a number of traffic flow data time slices from the historical traffic flow data at preset time slice intervals.

[0058] In this embodiment, the area to be predicted in S1 is a rectangular area, and the time span of the historical traffic flow data is [t, t + (m - 1)*τ], and m traffic flow data time slices are extracted at τ-minute time slice intervals.

[0059] S2. Perform rasterization on the area to be predicted to divide it into a series of grids. For each vehicle in each traffic flow data time slice, map it to the corresponding grid in the area to be predicted according to its coordinates, and define the vehicle driving direction as the moving state of the vehicle in the grid. The moving states include four types: upward, downward, leftward, and rightward. Count the total number of vehicles in each moving state contained in each grid within each time slice and use it as the grid value. Map the grid value to a matrix element, so as to construct a two-dimensional traffic flow matrix for different moving states in each time slice respectively. The two-dimensional traffic flow matrices of each moving state in all time slices are stacked in the time dimension to form a three-dimensional traffic flow matrix. Among them, the moving state of the vehicle in the grid can be determined according to the vehicle driving direction when the vehicle driving trajectory passes through the grid.

[0060] In this embodiment, the specific implementation steps of S2 are as follows:

[0061] S21. Perform rasterization on the area to be predicted, and divide it into I*J grids in total. The grid in the i-th row and j-th column is P ij ;

[0062] S22. For each vehicle in each traffic flow data time slice, map it to the corresponding grid in the area to be predicted according to its coordinates, and define the vehicle driving direction as the moving state of the vehicle in the grid. The moving states include four types: upward, downward, leftward, and rightward.

[0063] Since the vehicle driving direction is actually a 360° direction space, it is necessary to divide this 360° direction space at intervals of 90°. Establish an XY coordinate system on the map plane with the vehicle's location as the origin. Use the two lines y = x and y = -x to divide the entire 360° direction space into four sub-spaces with upward openings, downward openings, leftward openings, and rightward openings. The opening direction of the corresponding sub-space where the vehicle driving direction with the vehicle's location as the origin is located is used as the moving state of the vehicle in the grid.

[0064] S23. Count the total number of vehicles in each moving state contained in each grid within each time slice. Denote the traffic flow of vehicles with the moving state d in grid P ij within time slice t as Construct the corresponding to all I*J grids into a two-dimensional traffic flow matrix of the moving state d in the entire area to be predicted within time slice t Two-dimensional traffic flow matrix The element value in the i-th row and j-th column of the

[0065] S24. The two-dimensional traffic flow matrices of all m traffic flow data time slices Concatenate according to the time dimension to form a three-dimensional traffic flow matrix

[0066] S3. Extract the hour field and minute field from the moment to be predicted, and concatenate them to form a binary time signal vector.

[0067] In this embodiment, the specific implementation steps of S3 are as follows:

[0068] Convert the moment t to be predicted pred into a time signal vector h containing two elements, namely the hour field t pred_hour and the minute field t pred_minute h = [t t , t pred_hour , t pred_minute .

[0069] S4. Obtain the spatial geographical locations of all points of interest (POIs), map the POIs of different functional categories to the grids in the area to be predicted, count the total number of POIs of each group of functional categories in each grid as the grid value, and map the grid value to a matrix element, so as to construct a two-dimensional matrix form of POI slice for each group of functional categories of POIs respectively, and the POI slices of all functional categories are superimposed to form a three-dimensional tensor form of POI signal.

[0070] In the present invention, the point of interest (POI) is a geographical entity that realizes urban functions, reflecting the influence of different origins and destinations on the change of traffic volume. For example, the catering POI affects the traffic volume in the surrounding area during lunch and dinner times, while the tourist attraction POI mainly affects the traffic volume on weekends and holidays. In this embodiment, the POIs can be grouped into 9 categories according to food and beverage, shopping service, daily life service, medical service, accommodation service, tourist attraction, education service, transportation service, and others. Of course, other classification forms can also be adopted.

[0071] In this embodiment, the specific implementation steps of S4 are as follows:

[0072] S41. Obtain the spatial geographical locations of all POIs, and map all POIs to each grid P ij in the area to be predicted according to their locations;

[0073] S42. Divide all POIs into n groups according to different functional categories, count the number of POIs in each POI group g in the grid P ij as the grid value of the grid P ij ; Construct the grid values of all grids corresponding to each POI group g into a POI slice γ g corresponding to the POI group g, γ gThe size is I*J;

[0074] S43. Concatenate the sliced points of interest corresponding to all n groups of points of interest to obtain the point-of-interest signal Ψ = [γ1, γ2, …, γ n , whose size is n*I*J.

[0075] S5. Use the three-dimensional traffic flow matrix, the time signal vector, and the point-of-interest signal as the inputs of a trained multi-directional traffic flow prediction model. As mentioned before, this multi-directional traffic flow prediction model includes a first fully connected neural network, a second fully connected neural network, a three-dimensional residual convolutional network, and a recalibration layer.

[0076] In this embodiment, the specific implementation steps of S5 are as follows:

[0077] S51: Input the three-dimensional traffic flow matrix the time signal vector h t and the point-of-interest signal Ψ into the trained multi-directional traffic flow prediction model, which includes a first fully connected neural network, a second fully connected neural network, a three-dimensional residual convolutional network, and a recalibration layer;

[0078] S52. The time signal h t is input into the first fully connected neural network consisting of L ts layers of cascaded fully connected layers. The input of the first layer of the fully connected layer is the time signal h t , the input of the next layer of the fully connected layer is the output of the previous layer of the fully connected layer, and the output of the last layer of the fully connected layer is Map the vector element by element into a matrix of size (I*J) to obtain a time signal matrix H of size (I, J) t ;

[0079] After the point-of-interest signal Ψ = [γ1, γ2, …, γ n is input, first obtain the average self-weight z g for each sliced point of interest γ g in it:

[0080]

[0081] Obtain the average self-weight matrix Z = {z1, z2, …, z n} of the point-of-interest signal Ψ, where n represents the number of point-of-interest groups;

[0082] Then input the average self-weight matrix Z into the L psThe second fully-connected neural network cascaded by layer-wise fully-connected layers is calculated layer by layer. The input of the first fully-connected layer is the average self-weight matrix Z, the input of the next fully-connected layer is the output of the previous fully-connected layer, and the output of the last fully-connected layer is

[0083] Subsequently, the output is mapped to a variable between 0 and 1 using a gate mechanism, and its calculation process is as follows: The calculation process is as follows:

[0084]

[0085] where f si is the ReLU activation function;

[0086] Finally, the point-of-interest signal matrix is obtained through calculation

[0087]

[0088] where ⊙ represents matrix multiplication;

[0089] S54. Feature fusion is performed on the three-dimensional traffic flow matrix the point-of-interest signal matrix and the time signal matrix H t The fusion feature of the k-th traffic flow data time slice in is calculated as follows:

[0090]

[0091] where and are both trainable parameters, and m is the number of traffic flow data time slices in Finally, the fused traffic flow matrix

[0092] S55. The fused traffic flow matrix X Γ is input into a three-dimensional residual convolutional network cascaded by L c layers of three-dimensional residual convolutional layers and calculated layer by layer. The input traffic flow matrix X Γ of the first three-dimensional residual convolutional layer, and the result obtained after the three-dimensional residual convolutional operation of each three-dimensional residual convolutional layer is used as the input of the next three-dimensional residual convolutional layer. The outputs of all three-dimensional residual convolutional layers in the three-dimensional residual convolutional network are concatenated to form the final result matrix X ST ; for any l-th three-dimensional residual convolutional layer, the three-dimensional residual convolutional operation performed is as follows:

[0093] First, a three-dimensional convolution operation is performed on the input of the current three-dimensional residual convolutional layer to obtain the convolution result:

[0094]

[0095] Among them, Cov3D represents a three-dimensional convolution operation, represents the output of the (l-1)-th layer three-dimensional residual convolution layer, where and are the trainable parameters of the l-th layer three-dimensional convolution layer (i.e., the aforementioned three-dimensional convolution operation Cov3D), and f c is the ReLU activation function;

[0096] Then, a batch normalization operation is performed on each element in the convolution result output by the three-dimensional convolution layer to obtain the batch normalization result The formula is as follows:

[0097]

[0098] Among them, E[x] represents the mean of each-dimensional matrix, Var[x] is the variance of each-dimensional matrix, ∈ is a constant set to prevent the variance from being 0, and γ and β are learnable parameters;

[0099] Finally, the batch normalization result is added to the output matrix of the previous layer to obtain the output matrix of the l-th layer three-dimensional residual convolution layer. The formula is:

[0100]

[0101] S56. In the recalibration layer, a weighted compression operation is performed on all dimensions of the finally output result matrix X ST to obtain the prediction result The calculation formula is as follows:

[0102]

[0103] Among them is a learnable parameter matrix.

[0104] It should be noted that in the above S5, the multi-direction traffic flow prediction model is pre-trained with training data, and during the training process, the loss value between the prediction result and the real result Φ is continuously iterated through the loss function. When the iteration termination condition is reached, the multi-direction traffic flow prediction model is output for actual prediction.

[0105] As an implementation form of the loss function Loss of a multi-direction traffic flow prediction model, the formula is:

[0106]

[0107] wherein is a matrix is the value of each element in is the value of each element in matrix Φ, and M is the number of training samples.

[0108] The above multi-direction traffic flow prediction model can effectively predict traffic flows in different directions, and its prediction accuracy is significantly better than that of traditional mathematical methods and machine learning-related methods.

[0109] In another embodiment of the present invention, a cloud platform is further provided, which is used to cooperate with image capture devices and edge computing devices installed at different positions on the road to achieve traffic flow prediction;

[0110] The image capture device is used to capture real-time images of passing vehicle flows, and the driving direction of the vehicles captured by the same image capture device is fixed;

[0111] The edge computing device is paired with the image capture device one by one, and is used to obtain the vehicle flow images captured by the image capture device at the same position, locate each license plate area in the vehicle flow image through the built-in license plate positioning model, and then identify the license plate number in each license plate area through the license plate recognition model;

[0112] The cloud platform is communicatively connected to edge computing devices at different positions through the Internet of Things, and is internally provided with a data receiving module, a data fusion module, and a traffic flow prediction module;

[0113] The data receiving module is used to receive in real time the license plate number recognition results reported by edge computing devices at different positions, as well as the corresponding timestamps and vehicle driving directions;

[0114] The data fusion module is used to associate and fuse each license plate number reported by each edge computing device with the corresponding timestamp, vehicle driving direction, and the coordinates of the edge computing device to form a trajectory point. All consecutive trajectory points of each license plate number are used to restore the driving trajectory of the corresponding vehicle by calling a path planning algorithm, and the driving trajectories of all vehicles are stored in the historical vehicle flow database as vehicle flow data;

[0115] The traffic flow prediction module is used to read the stored vehicle flow data from the historical vehicle flow database, and predict the traffic flow at a future moment based on the trained traffic flow prediction model.

[0116] It should be noted that in the above cloud platform, the specific implementation methods in each module can also adopt the practices in the aforementioned traffic flow prediction system based on the Internet of Things and edge computing, and will not be repeated here.

[0117] The above-described embodiments are only a preferred solution of the present invention, but they are not intended to limit the present invention. Those of ordinary skill in the relevant technical field can still make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all technical solutions obtained by means of equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A low-latency real-time traffic prediction cloud platform for coupling the Internet of Things and edge computing devices, characterized in that, It is used to cooperate with image capture devices and edge computing devices installed at different positions on the road to achieve traffic flow prediction; The image capture device is used to capture real-time images of passing vehicle flows, and the driving direction of the vehicles captured by the same image capture device is fixed; The edge computing device is paired with the image capture device one by one, and is used to obtain the vehicle flow images captured by the image capture device at the same position, locate each license plate area in the vehicle flow image through the built-in license plate location model, and then identify the license plate numbers in each license plate area through the license plate recognition model; The cloud platform is communicatively connected to edge computing devices at different positions through the Internet of Things, and is internally provided with a data receiving module, a data fusion module and a traffic flow prediction module; The data receiving module is used to receive in real time the license plate number recognition results reported by edge computing devices at different positions, as well as the corresponding timestamps and vehicle driving directions; The data fusion module is used to associate and fuse each license plate number reported by each edge computing device with the corresponding timestamp, vehicle driving direction and the coordinates where the edge computing device is located to form a trajectory point. All consecutive trajectory points of each license plate number are used to restore the driving trajectory of the corresponding vehicle by calling the path planning algorithm, and the driving trajectories of all vehicles are stored as traffic flow data in the historical traffic flow database; The traffic flow prediction module is used to read the stored traffic flow data from the historical traffic flow database, and predict the traffic flow at a future moment based on a trained traffic flow prediction model; The traffic flow prediction model is a multi-directional traffic flow prediction model, which includes a first fully connected neural network, a second fully connected neural network, a three-dimensional residual convolutional network and a recalibration layer. The input of the multi-directional traffic flow prediction model is a traffic flow three-dimensional matrix, a time signal vector and a point of interest signal. The first fully connected neural network outputs a time signal matrix according to the time signal vector, the second fully connected neural network outputs a point of interest signal matrix according to the point of interest signal, the three-dimensional residual convolutional network outputs a result matrix according to the fusion features of the traffic flow three-dimensional matrix, the point of interest signal matrix and the time signal matrix, and finally, after a weighted compression operation on the result matrix in the recalibration layer, a multi-directional traffic flow prediction result is obtained.

2. The cloud platform according to claim 1, characterized in that, The image capture device and the edge computing device are installed in pairs at the intersection positions of the road.

3. The cloud platform according to claim 1, characterized in that, The license plate location model is a YOLO model.

4. The cloud platform according to claim 1, wherein The license plate recognition model is a CNN convolutional neural network model.

5. The cloud platform according to claim 1, characterized in that, The traffic flow prediction module is provided with a specified module for inputting a prediction area and a prediction moment.

6. The cloud platform according to claim 1, wherein, Both the license plate location model and the license plate recognition model are pre-trained and then downloaded to the edge computing device.

7. The cloud platform according to claim 1, characterized in that The traffic flow prediction model in the cloud platform is continuously trained by using an incremental learning method.

8. The cloud platform according to claim 1, characterized in that, The image capture device is a camera installed above the intersection, and each camera only shoots in the driving direction of the vehicle flow.

Images