A fine-grained urban traffic flow inference method based on spatial heterogeneity

Abstract

The application provides a fine-grained urban traffic flow inference method based on spatial heterogeneity, first constructs a traffic flow fine-grained inference model, infers fine-grained urban traffic flow, then performs masking, and finally constructs an encoder-double decoder model, sequentially passes coarse-grained urban traffic flow and the like through the trained traffic flow fine-grained inference model and the encoder-double decoder model, and completes fine-grained urban traffic flow inference and positioning and compensation of fine-grained difficult areas. The fine-grained urban traffic flow inference method based on spatial heterogeneity provided by the application fully considers spatial heterogeneity under the urban geographical background, can more effectively perform fine-grained urban traffic flow inference, and guides the model to pay more attention to urban areas with complex flow patterns in the optimization process, and thus better restores the traffic flow conditions of the areas.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spatial data mining, and in particular to a fine-grained urban traffic flow inference method based on spatial heterogeneity. BACKGROUND

[0002] In the field of intelligent transportation systems (ITS), urban traffic flow data is of great importance, as it contains rich spatio-temporal information about human movement patterns in cities. Therefore, obtaining fine-grained urban traffic flow data is very helpful for numerous smart city applications, such as urban planning and traffic management. Generally, fine-grained urban traffic flow monitoring requires a large amount of infrastructure, including a wide distribution of sensors and related support equipment, such as surveillance cameras and power supply systems. However, the construction and maintenance of a dense sensor network covering an entire city requires a huge economic cost, especially for metropolises. Therefore, there is an urgent need for ITS applications to develop an algorithmic technique that can automatically infer fine-grained urban traffic flow data from coarse-grained urban traffic flow data.

[0003] Fine-grained urban flow inference (FUFI) aims to reconstruct fine-grained urban traffic flow data from coarse-grained observation data. This means that only a small number of monitoring sensors are needed, and then the fine-grained urban traffic flow can be restored through an algorithm. However, the fine-grained urban flow inference problem is challenging. First, there is a complex spatial correlation between coarse-grained and fine-grained urban traffic flow. Considering urban traffic flow, the traffic flow in one region is largely influenced by neighboring regions. However, the traffic flow distribution of adjacent fine-grained partitions within the same coarse-grained region can also differ significantly. Second, urban traffic flow between different regions exhibits spatial heterogeneity, which means that traffic flow in different regions exhibits different distribution patterns. Some regions have traffic flow that exhibits general flow pattern rules, while other regions may have traffic flow that exhibits complex and irregular pattern characteristics.

[0004] Currently, fine-grained urban traffic flow inference models mainly focus on using deep models to capture the complex spatial correlation between coarse-grained and fine-grained urban traffic flow. Therefore, compared with traditional statistical interpolation methods, deep models are more accurate for fine-grained urban traffic flow inference. However, these models usually ignore the spatial heterogeneity of urban flow, which can lead to the existing methods being more inclined to focus on easy-to-reconstruct regions during optimization, while being less effective for difficult regions with complex flow distribution characteristics. SUMMARY

[0005] In order to solve the problems of the prior art, the present application provides a fine-grained urban traffic flow inference method based on spatial heterogeneity, which fully considers the spatial heterogeneity in the urban geographical background and can more effectively infer the fine-grained urban traffic flow, so as to guide the model to pay more attention to the urban areas with complex flow patterns in the optimization process, and thus better restore the traffic flow conditions of this type of area.

[0006] The technical scheme adopted by the present application to solve the technical problems is to provide a fine-grained urban traffic flow inference method based on spatial heterogeneity, comprising the following steps:

[0007] S1, constructing a traffic flow fine-grained inference model, inputting the training input data constructed by integrating the coarse-grained urban traffic flow data and the extracted features of external factors into the traffic flow fine-grained inference model, and inferring the fine-grained urban traffic flow;

[0008] S2, generating error mapping and performing binary classification, adaptively generating a mask, and using the mask to mask the areas with error lower than the threshold value to obtain the trained traffic flow fine-grained inference model;

[0009] S3, constructing an encoder-double decoder model, sequentially inputting the training input data into the trained traffic flow fine-grained inference model and the encoder-double decoder model, and completing the fine-grained urban traffic flow inference and the positioning and compensation of fine-grained difficult areas.

[0010] The traffic flow fine-grained inference model of step S1 comprises, from the input end to the output end, a first convolutional layer, a deep residual module, an up-sampling module, a second convolutional layer and a rectified linear unit activation layer.

[0011] The deep residual module comprises a group of residual block modules, a convolutional layer and a batch normalization layer, wherein the residual block module adopts an original residual structure composed of two convolutional layers, two batch normalization layers and a rectified linear unit activation layer.

[0012] The up-sampling module comprises a pixel shuffle (PixelShuflle) layer.

[0013] The mask processing of step S2 is implemented by the following process:

[0014] Masked FG=FG⊙Mask

[0015] Masked Input=input⊙AvePool(Mask,S)

[0016] Wherein, Masked FG represents the masked fine-grained urban traffic flow, FG represents the real fine-grained urban traffic flow, Mask represents the generated mask, represents Hadamard, Masked Input represents the masked training input data, input represents the training input data, AvePool() represents the average pooling operation, and S represents the up-sampling factor.

[0017] The encoder-decoder model in step S3 specifically includes the following processing procedures:

[0018] S3.1, a multi-layer down-sampling process is used to extract multi-scale features of the urban traffic flow data, and the multi-layer down-sampling process adopts the following calculation:

[0019] F1=Extraction(I-FG)

[0020] F i+1 =Extraction(MaxPool(F i )),i=1,2,…,n

[0021] Wherein, Extraction() represents a feature extraction module operation, MaxPool() represents a maximum pooling operation, I-FG represents the fine-grained urban traffic flow generated by the traffic flow fine-grained model, F1 represents a bottom layer scale feature, and F i represents an i-th layer scale feature, n represents the number of down-sampling layers except the bottom layer, and the last layer scale feature is a multi-scale feature after multi-layer down-sampling;

[0022] S3.2, the multi-scale features flow to a double-branch, including an attention branch and a residual feature branch, the multi-scale features input into the attention branch and the residual feature branch are up-sampled through a bilinear interpolation process, and then are combined with the n-th scale feature to the first scale feature in turn to obtain attention branch features and residual features;

[0023] S3.3, the attention branch features flow to a convolution layer, and an attention map is generated by a Sigmoid activation function;

[0024] S3.4, the attention map and the residual features are combined in the form of Hadamard product to calculate residual errors.

[0025] The application has the beneficial effects of the technical scheme in that:

[0026] The application provides a fine-grained urban traffic flow inference method based on spatial heterogeneity, which fully considers the spatial heterogeneity in the urban geographical background, can more effectively perform fine-grained urban traffic flow inference, and guide the model to pay more attention to urban areas with complex flow patterns in the optimization process, thereby better restoring the traffic flow conditions of the areas. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a flow chart of the fine-grained urban traffic flow inference method based on spatial heterogeneity.

[0028] Figure 2 is a detail view of the traffic flow fine-grained inference model.

[0029] Figure 3 is a flow chart of the adaptive shielding strategy.

[0030] Figure 4 is a framework diagram of the encoder-decoder model.

[0031] Figure 5 is a visualization diagram of inference errors between different methods. DETAILED DESCRIPTION

[0032] The application will be further described below in combination with the drawings and examples.

[0033] The application provides a fine-grained urban traffic flow inference method based on spatial heterogeneity, which fully considers the spatial heterogeneity in the urban geographical background, can more effectively perform fine-grained urban traffic flow inference, and guide the model to pay more attention to urban areas with complex flow patterns in the optimization process, thereby better restoring the traffic flow conditions of the areas. Figure 1 , comprising the following steps:

[0034] S1, constructing a traffic flow fine-grained inference model, inputting the training input data constructed by integrating the coarse-grained urban traffic flow data and the extracted features of external factors into the traffic flow fine-grained inference model, and inferring the fine-grained urban traffic flow data.

[0035] Referring to Figure 2 , the traffic flow fine-grained inference model comprises a first convolutional layer, a deep residual module, an up-sampling module, a second convolutional layer and a rectified linear unit activation layer in sequence from the input end to the output end. In this embodiment, the deep residual module comprises 16 residual block modules (Resblock), a convolutional layer and a batch normalization layer, wherein the Resblock adopts a original residual structure composed of two convolutional layers, two batch normalization layers and a rectified linear unit activation layer. The up-sampling module comprises a pixel recombination layer (PixelShuffle).

[0036] During the inference process, the first step is to use the convolution layer (Conv) to extract the bottom features of the input. Then a deep residual module is designed to further extract deep urban traffic flow features. In this module, 16 Resblocks are used to achieve a balance between efficiency and effectiveness, and each ResBlock adopts the original residual structure, which includes two convolution layers, two batch normalization layers (BN) and a rectified linear unit activation layer (ReLU). Finally, the up-sampled features are used to restore the global fine-grained urban traffic flow data through the convolution layer and the rectified linear unit activation layer. Except for the final output convolution layer (channel number 64, convolution kernel size 3), all convolution layers in the framework of the embodiment are set to the same configuration (channel number 1, convolution kernel size 3).

[0037] S2, referring to Figure 3 Adaptive masking strategy is performed. Error map is generated and binary classification is performed, and mask is adaptively generated. The mask is used to mask the area where the error is lower than the threshold, and the trained traffic flow fine-grained inference model is obtained. The following process is used to achieve it:

[0038] First, the error map of the training data at the current iteration is generated by the traffic flow fine-grained inference model. Then, through the binary classification process, the mask is adaptively generated according to the error map. Since the focus is on the difficult area, the binary classification process is designed to mark the area with the highest error of ρ% as 1, and other areas as 0. The mask of the i-th row and j-th column area in the urban traffic flow map is as follows:

[0039]

[0040] Where I-FG is the fine-grained urban traffic flow inferred in step S1, E(FG, I-FG) is the list of inference errors of all areas, e i，j is the inference error of the i-th row and j-th column area in the inference error list, top ρ% is a hyperparameter, and the default value of ρ is 20.

[0041] Then, the obtained mask is used to mask the area where the input coarse-grained and fine-grained urban traffic flow error is small. The process is as follows:

[0042] Masked FG=FG⊙Mask

[0043] Masked Input=input⊙AvePool(Mask,S)

[0044] In this context, Masked FG represents the fine-grained urban traffic flow after masking, FG represents the actual fine-grained urban traffic flow, Mask represents the generated mask, which is the general term for the masks of each area, ⊙ represents the Hadamard product, Masked Input represents the training input data after masking, input represents the training input data, AvePool() represents the average pooling operation, and S represents the upsampling factor.

[0045] S3. Construct an encoder-dual decoder model for finer-grained localization and compensation of difficult regions. The encoder-dual decoder model framework is as follows: Figure 4 As shown, the coarse-grained urban traffic flow to be analyzed is sequentially passed through a trained fine-grained traffic flow inference model and an encoder-dual-decoder model to complete fine-grained urban traffic flow inference and the location and compensation of fine-grained difficult areas.

[0046] The encoder-dual decoder model specifically includes the following processing steps:

[0047] S3.1. Multi-scale features of urban traffic flow data are extracted using a multi-level downsampling process. The multi-level downsampling process employs the following calculations:

[0048] F1 = Extraction(I-FG)

[0049] F i+1 =Extraction(MaxPool(F i )), i = 1, 2, 3, 4

[0050] Where Extraction() represents the feature extraction module operation, MaxPool() represents the max pooling operation, I-FG represents urban traffic data, and F1 represents the low-level scale features. i This represents the scale feature of the i-th layer, and the scale feature of the last layer is the multi-scale feature that has undergone multiple downsampling layers.

[0051] S3.2 Then, to enhance multi-scale feature fusion in the decoder, the features extracted from the encoder will flow to a dual branch, including an attention branch and a residual feature branch, used for fine-grained hard region localization and compensation, respectively. The attention branch is designed to localize fine-grained hard regions. In the attention branch, input features (such as...) Figure 4 (A1) is first upsampled through a bilinear interpolation process, and then compared with the corresponding scale features extracted by the encoder (such as... Figure 4F4) is combined. This residual fusion mechanism can avoid overfitting problem and improve the robustness of the model. Then, the combined features will flow to the Extraction block to generate features for the next layer (e.g. A2). In this way, this hierarchical decoder structure can make full use of the multi-scale features extracted by the encoder, which helps to locate the fine-grained difficult regions. This process can be described as:

[0052] A i+1 = Extraction [cat(Upscale i (A 5-i ), F i+1 )], i = 1, 2, 3, 4

[0053] R ×2 = Extraction [cat(Upscale i (A 5-i ), F ×2 )], i = 1, 2, 3, 4

[0054] where Upscale c→1 represents the bilinear interpolation process, and cat represents the concatenation operation. After processing, the attention branch features and residual features are obtained.

[0055] S3.3, at the end of the attention branch, the final feature map (e.g. A5) will flow to the convolution layer, and then the attention map representing the difficulty of each unit region will be generated by the Sigmoid activation function. Figure 4

[0056] S3.4, after locating the fine-grained difficult regions, the residual features will also be extracted by a similar attention branch (without the final Sigmoid function), and finally the attention map and the residual features will be combined in the form of Hadamard product to calculate the residual error, which is described as:

[0057] Attention = sigmoid(Conv c→1 (A5))

[0058] Residual feature = Conv c→1

[0059] Improved I-FG = I-FG + Attention ⊙ Residual feature

[0060] where Conv t ​​is a convolutional layer that converts the feature map (e.g., A5 or R5) from c channels (the default value of c is 64) to 1 channel, sigmoid is an activation layer, and Improved I-FG is the final result after the entire two stages. Therefore, through the second stage, the reconstruction result of the traffic flow fine-grained inference model can be further enhanced.

[0061] A spatial heterogeneity-based fine-grained urban traffic flow inference method provided by the present application is used to process a Beijing taxi flow data set.

[0062] I. Data collection

[0063] The taxi flow data set TaxiBJ used by the present application is a benchmark data set used in many traffic flow related works. TaxiBJ is composed of four time period (P1, P2, P3, P4) data, as shown in Table 1. The specific composition of the data set is shown in Table 1.

[0064]

[0065] Table 1: Data set description

[0066] In this embodiment, the data of each time period is divided into training data, validation data and test data, and the ratio is 2:1:1. For example, in P1 (7 / 1 / 2013-10 / 31 / 2013), the data of the first two months is used as the training set, and the data of the next month and the last month is used as the validation set and the test set respectively, to test the performance of the method implemented at different times.

[0067] II. Method implementation and comparison

[0068] In order to evaluate the advantages of the present application, it is compared with several methods from classic to state-of-the-art,

[0069] Mean (average split): Mean assumes that the traffic of a coarse-grained unit is uniformly distributed in the corresponding fine-grained unit. For example, assuming that the traffic in a coarse-grained unit is 4, and the magnification factor is 2, then the traffic of each fine-grained unit will be 1.

[0070] HA (historical average level): Unlike Mean, HA calculates the proportion of each fine-grained unit in the corresponding coarse-grained unit according to the training data (i.e. historical data), and then uses this proportion as a weight to distribute the traffic of each fine-grained unit.

[0071] SRCNN: SRCNN is the first image super-resolution method that introduces convolutional neural network (CNN). SRCNN has a high-efficiency and light-weight network structure, which uses a bicubic interpolation method to improve the low-resolution image to high-resolution.

[0072] SRResNet: SRResNet is a variant of SRGAN, which is the first image super-resolution method that can recover natural images from down-sampled 4 times images. Since SRGAN aims to improve the visual quality of natural images, SRResNet is chosen as the baseline method to obtain better flow reconstruction accuracy.

[0073] UrbanFM: UrbanFM is different from SRCNN and SRResNet, it is a deep learning method specially designed for fine-grained urban flow inference (FUFI). UrbanFM takes into account the influence of external factors on traffic and the complex spatial correlation between coarse-grained flow and fine-grained flow. It is generally considered to be the most advanced model in the field of FUFI.

[0074] In addition, in order to understand the impact of each stage of the invention on the results, we also compared the invention (FocusNet) with its different variants:

[0075] FocusNet-ns: a fine-grained traffic flow inference model that does not consider spatial heterogeneity modeling, that is, without stage one and two.

[0076] FocusNet-I: only the first stage is implemented, without the second stage part.

[0077] The invention uses an optimization algorithm called Adam Optimizer to train the model, divides the data into small batches of size 16 for training, and sets the learning rate of backpropagation to 0.001. The training process is divided into two stages,

[0078] In the first stage, a method similar to "curriculum strategy" is adopted. First, let the model learn normal data flow in the first 10 training cycles without any special processing. Then, between the 11th and 50th cycles, an adaptive masking strategy is introduced to guide the model to focus on processing those difficult areas, helping the model better understand and process the difficult parts of the traffic data. In the second stage, the model is again trained based on normal data flow until the model converges, that is, the performance of the model no longer improves. The mask is generated according to the error of the model in each iteration, and those areas with error values lower than the top 20% are selected, so the mask will include about 80% of the data. In addition to this method, related models such as SRCNN, SRResNet and UrbanFM are also implemented for comparison, and they are all trained and experimented according to their optimal configurations.

[0079] III. Results and evaluation

[0080] The present invention uses three common evaluation metrics to comprehensively evaluate the performance of different methods, which are Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Details are as follows:

[0081]

[0082]

[0083]

[0084] where I-FG t is the inferred value at time t, FG t is the corresponding true value.

[0085] The present invention is compared with other methods, and the results of each model under different evaluation metrics are listed as shown in Table 2. The best results are marked in bold.

[0086]

[0087]

[0088] Table 2 Comparison of results based on different time span datasets (P1-P4)

[0089] From the table, the following conclusions can be drawn: FocusNet shows the best performance under all time spans, which indicates that the present invention performs well in different scenarios and is robust. In addition, in almost all cases (except one case in P2), the performance of FocusNet is better than that of FocusNet-I, and the performance of FocusNet-I is better than that of FocusNet-ns, which confirms the effectiveness of the two stages of the present invention.

[0090] The Mean method performs the worst because it does not consider any prior information. On the contrary, the HA method performs much better because it uses historical traffic information. Although SRCNN is trained on historical traffic data, its performance is not satisfactory. This may be because the network structure of SRCNN is relatively shallow (only three convolutional layers), which cannot effectively capture the complex spatial correlation between coarse-grained flow and fine-grained flow. In contrast, because the network of SRResNet is deeper, better results are obtained.

[0091] UrbanFM-64 performs better than SRResNet in terms of MAE and MAPE, but worse in terms of RMSE. However, UrbanFM-128 is superior to SRResNet in all cases, which indicates that increasing more filters may improve the robustness of UrbanFM.

[0092] Considering the performance of both SRResNet and UrbanFM-128, we choose them as the baseline models for the subsequent experiments. These results help us understand which method performs best in terms of traffic data reconstruction.

[0093] V. Error visualization

[0094] Figure 5 The inferred error values |FGof different methods are shown in Figure 6. t -I-FG t The results show that the traditional statistical method (HA) can achieve relatively small errors in most areas, except for some difficult areas, i.e., the areas with brighter pixels. The deep learning methods (such as SRResNet and UrbanFM) can outperform HA because they achieve better results in these difficult areas. However, these deep learning methods may be limited by spatial heterogeneity when dealing with areas with different traffic patterns, while the proposed fine-grained urban traffic flow inference method based on spatial heterogeneity can effectively identify this problem.

[0095] VI. Analysis of masking strategy

[0096] Different mask ratios are used for the Focus-I method to study the effectiveness of the proposed masking strategy. The results, as shown in Table 3, show that most mask ratios can improve the performance of Focus-I, indicating that the proposed masking strategy is effective. Different mask ratios result in different evaluation index trends. For example, in P1, Focus-I achieves the best MAE at mask ratios of 50% and 80%, and the best MAPE at a mask ratio of 80%. However, from a global perspective, mask ratios of 50% and 80% can achieve relatively good results in the four time spans, indicating that appropriate mask ratios can guide the model to focus on a limited number of difficult areas, thereby helping to improve the accuracy of the model.

[0097]

[0098] Table 3 Focus-I results with different mask ratios

[0099] In addition, the shielding strategy is also applied to other methods, and the results are shown in Table 4. The results show that SRResNet-I and UrbanFM-I(128) obtain better results in most cases after using the shielding strategy. In addition, in some cases, it can even be seen that SRResNet-I is superior to UrbanFM-I in some indicators (for example, RMSE and MAE in P1). It needs to be emphasized that the shielding strategy does not increase the complexity of the model, so these results show that the shielding strategy can be effectively used in combination with existing methods to improve their performance.

[0100]

[0101] Table 4 Baseline results with shielding strategy

[0102] Seven, research on encoder-decoder model

[0103] In this section, the baseline model containing stages one and two of the present application is implemented, and the results are listed in Table 5. The results show that UrbanFM-II(128) (as well as SRResNet-II) performs better than UrbanFM-I (as well as SRResNet-I) and their original versions in all cases. In addition, the improvements compared with UrbanFM-I and SRResNet-I are also listed to demonstrate the individual effect of stage two. The results show that the two stages proposed by the present application can be well used in combination with existing methods to improve their performance.

[0104]

[0105] Table 5 Baseline model results with stages one and two

[0106] In summary, the present application provides a fine-grained urban traffic flow inference method based on spatial heterogeneity, which can more effectively perform fine-grained urban traffic flow inference by considering the spatial heterogeneity of different regions in the city. This method can pay more attention to urban areas with complex flow patterns, and thus better restore the traffic flow situation of the entire region.

Claims

1. A method for fine-grained urban traffic flow inference based on spatial heterogeneity, characterized in that The method comprises the following steps: S1, constructing a traffic flow fine-grained inference model, inputting training input data constructed by integrating coarse-grained urban traffic flow data and external factor extracted features into the traffic flow fine-grained inference model to infer fine-grained urban traffic flow; the traffic flow fine-grained inference model comprises a first convolutional layer, a deep residual module, an up-sampling module, a second convolutional layer and a rectified linear unit activation layer from input end to output end; the deep residual module comprises a group of residual block modules, a convolutional layer and a batch normalization layer, wherein the residual block module adopts an original residual structure composed of two convolutional layers, two batch normalization layers and a rectified linear unit activation layer; the up-sampling module comprises a pixel recombination layer; S2, generating error mapping and performing binary classification, adaptively generating a mask, and using the mask to mask the area with error lower than the threshold value to obtain the trained traffic flow fine-grained inference model; the mask processing specifically adopts the following process: ， ， wherein, Masked FG denotes the masked fine-grained urban traffic flow, FG denotes the real fine-grained urban traffic flow, Mask denotes the generated mask, denotes the Hadamard product, Masked Input denotes the masked training input data, input denotes the training input data, AvePool () denotes an average pooling operation, S denotes an up-sampling factor; S3, constructing an encoder-double decoder model, sequentially inputting the training input data into the trained traffic flow fine-grained inference model and the encoder-double decoder model to complete fine-grained urban traffic flow inference and positioning and compensation of fine-grained difficult areas; the encoder-double decoder model specifically comprises the following processing process: S3.1, extracting multi-scale features of urban traffic flow data by using a multi-layer down-sampling process, and the multi-layer down-sampling process adopts the following calculation: ， ， wherein Extraction () represents a feature extraction module operation, MaxPool () represents a max-pooling operation, I - FG representing a fine-grained urban traffic flow generated by the traffic flow fine-grained model, F 1 represents a bottom layer scale feature, F i representing the first i layer scale feature, n representing a number of down-sampling layers except the bottom layer, the last layer scale feature being a multi-scale feature after multi-layer down-sampling; S3.2, flowing the multi-scale features to a double branch, including an attention branch and a residual feature branch; the multi-scale features input into the attention branch and the residual feature branch are up-sampled by a bilinear interpolation process, and then combined with the n-th scale feature to the first scale feature in sequence to obtain attention branch features and residual features; S3.3, flowing the attention branch features to a convolutional layer to generate an attention map by a Sigmoid activation function; S3.4, combining the attention map and the residual features in the form of Hadamard product to calculate residual error.

Images