Urban traffic speed estimation method based on multi-source crowd-sensing data

Abstract

The application discloses a kind of urban traffic speed estimation method based on multi-source crowd-sensing data, which is based on the road pedestrian data and on-road navigation data collected by smart phone, and the final estimated speed is obtained through the steps of missing data filling, self-view speed aggregation and multi-view speed summation.The fine-grained large-scale urban traffic speed estimation method can estimate the speed of all types of roads, including suburban sections and small roads, rather than focusing on the main roads in the city center.The application is based on data-driven, without the need to install additional devices on the road, with low cost, strong universality, stronger practicality, theoretical and applicability compared with prior art, and has important significance for improving traffic management and planning.

Description

Technical Field

[0001] This invention relates to a traffic speed estimation method, and more particularly to an urban traffic speed estimation method based on multi-source crowdsensing data. Background Technology

[0002] Fine-grained, large-scale urban traffic speed estimation is of great significance for urban traffic management and improvement. Traditional coarse-grained traffic speed estimation only calculates speeds on small road segments based on a limited number of traffic sensors. Nowadays, mobile phones are increasingly used for navigation purposes. When mobile phone users use map or ride-hailing apps, service providers record GPS coordinates. This mobile navigation data has become an important data source for traffic monitoring and sensing, and is widely used in traffic state estimation. However, mobile navigation data has uneven spatial coverage; more data is typically collected in hotspot areas, while little or no data is collected in suburban areas, making fine-grained traffic speed estimation impossible. Besides mobile navigation data obtained for navigation purposes, location-based services are involved when users use mobile applications such as Weibo and Meituan, and many pedestrians use their phones while walking. Simultaneously, when some pedestrians use location-based mobile applications, they randomly scan for nearby vehicle Wi-Fi signals and report their current location. We can filter vehicle Wi-Fi signals using the list of Wi-Fi signals reported by pedestrians and approximate vehicle locations based on pedestrian positions. The obtained data can cover more sidewalk areas without deploying any additional equipment. Therefore, it is possible to integrate roadside pedestrian data with on-road navigation data to obtain fine-grained large-scale urban traffic speed estimates in a low-cost and accurate manner. Summary of the Invention

[0003] The purpose of this invention is to improve and standardize existing research and technologies, proposing a method for estimating urban traffic speed based on multi-source crowdsourced sensing data. This method provides an overall data processing flow for traffic speed estimation, which can promote urban traffic planning and management and has practical value.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] A method for estimating urban traffic speed based on multi-source crowd sensing data includes the following steps:

[0006] Step 1, Dataset Preprocessing: Clean the raw dataset collected by smartphones to obtain roadside pedestrian data and on-road navigation data;

[0007] Step 2, Average speed calculation: Calculate the speed X of each road segment in different time periods using the dataset from Step 1;

[0008] Step 3, Missing Data Imputation: Using the current speeds X of each road segment in different time periods obtained in Step 2, combined with the historical speeds H of each road segment in different time periods, the missing speed data in X is imputed to obtain the imputed speeds.

[0009] Step 4, Self-view speed aggregation: Utilize the filled speed calculated in Step 3. Based on historical speed H, the spatial dependence between different road segments is quantified, and useful neighbor information is aggregated to obtain the aggregated roadside pedestrian speed data V. d and on-road navigation speed data V w ;

[0010] Step 5, Multi-view speed fusion: Based on the aggregated roadside pedestrian speed data V d and on-road navigation speed data V w By using a multilayer perceptron (MLP), multi-source velocity data is fused based on timestamps and whether the current velocity data is filled data, to obtain the fused velocity. Finally, the fusion speed is corrected based on road type characteristics to obtain a fine-grained, large-scale urban traffic speed estimate.

[0011] Further, step 1 specifically includes: obtaining road navigation data by filtering the APP usage list in the original data, that is, the personal location data reported by users when using programs such as Gaode Map navigation and Didi Chuxing; obtaining roadside pedestrian data by filtering the scanned WIFI signal list in the original data and determining whether there is a vehicle WIFI signal in the list, that is, when a user accidentally scans the WIFI signal of a passing vehicle while using their mobile phone and reports their personal location, the location of the roadside pedestrian is approximated as the location of the passing vehicle.

[0012] Furthermore, step 2 specifically includes: using the data obtained by cleaning and filtering in step 1, projecting the trajectory data into the road network using the Hidden Markov Network Matching Algorithm, thereby obtaining the speed X of each road segment in different time periods. The Hidden Markov Network Matching Algorithm is also known as the Hidden Markov Model Map Matching Algorithm and is a well-known technology.

[0013] Furthermore, step 3 specifically includes: firstly, introducing a mask matrix M to represent the missing units of velocity X:

[0014]

[0015] Secondly, a historical velocity matrix H is constructed using historical data to provide additional information to help fill in the missing data. Another mask matrix N is introduced to represent the missing units of H:

[0016]

[0017] Define a weighted matrix W to measure the importance of each item in the historical velocity matrix H, and then perform matrix decomposition using H, E, M, N, and X:

[0018]

[0019]

[0020]

[0021] Then based on the updated U + and V + Constructing the loss through matrix factorization:

[0022]

[0023] Then update the weight W:

[0024]

[0025] Where α and β are the learning rate parameters;

[0026] After iterative updates, we obtain a fully learned weighted matrix W. We can then use both X and H to estimate the missing data.

[0027]

[0028]

[0029] Where λ represents the penalty term parameter, and U and V represent the two submatrices decomposed from the original matrix.

[0030] Further, step 4 specifically includes: using self-view speed aggregation to capture the spatial correlation between adjacent roads, aggregating neighboring road segment information that is highly correlated with the central road segment, and firstly, calculating the spatial correlation e between road segment i and road segment j based on the historical speed matrix. i，j Furthermore, it retains highly relevant information while ignoring irrelevant information:

[0031]

[0032]

[0033] Then, based on spatial correlation e i，j The fusion coefficient 'a' between road segments was calculated. i，j This leads to the roadside pedestrian speed data V obtained after self-view aggregation. dand on-road navigation speed data V w :

[0034]

[0035]

[0036] Where ε represents a very small constant to prevent overflow, k represents a constant scaling value, and N i V represents the adjacency set of road segment i. i A unified representation of pedestrian speed data V on the roadside d and on-road navigation speed data V w The i-th row in the diagram represents the feature representation of road segment i, which will not be described separately.

[0037] Further, step 5 specifically includes: effectively fusing multi-source data using multi-view speed fusion, and determining F based on the timestamp feature and whether the current speed data is filled data. d and F w Then, the features are passed through an embedding layer and concatenated, based on the aggregated roadside pedestrian speed data V obtained in step 4. d and on-road navigation speed data V w The fusion speed is obtained through a multilayer perceptron (MLP).

[0038] Z = Embedding(Concat(F) d ,F w ))

[0039]

[0040] Finally, the estimated speed is corrected based on external factors (i.e., road type). Based on historical data, the speed distribution for each road type can be easily obtained. Treating this speed distribution as a normal distribution, speeds falling at the tail end of the distribution are corrected to arrive at the final estimated speed.

[0041] Compared with the prior art, the present invention has the following innovative advantages and significant effects:

[0042] 1) This invention integrates data from multiple sources, combines roadside pedestrian data and on-road navigation data, achieves 100% coverage of speed estimation across the entire road network, fills in more than 70% of missing data, and greatly reduces the cost of speed estimation without the need for additional equipment such as loop detectors or cameras, which cost tens of thousands of yuan per unit.

[0043] 2) This invention proposes a standardized data processing flow, and the specific implementation of each step can vary, which has strong flexibility and scalability. Attached Figure Description

[0044] Figure 1 This is a flowchart of the fine-grained large-scale urban traffic speed estimation method of the present invention;

[0045] Figure 2 This is a schematic diagram of the correlation between neighboring road segments in step 3 of the present invention. In the figure, (1) represents the navigation data on the road and (2) represents the pedestrian data on the roadside.

[0046] Figure 3 This is a schematic diagram of the improvement of self-view speed aggregation accuracy in step 3 of the embodiment of the present invention. In the figure, (1) is the navigation data on the road and (2) is the pedestrian data on the roadside. Detailed Implementation

[0047] The specific implementation method and working principle of the present invention will be described in detail below with reference to the accompanying drawings:

[0048] Example

[0049] This embodiment processes user data collected from a certain location between March 21, 2020 and March 28, 2020. The data collection process was anonymized. The specific variables included in the dataset are shown in Table 1.

[0050] Table 1. Crowd Sensing Dataset

[0051]

[0052]

[0053] In this embodiment, the implementation dataset for fine-grained large-scale urban traffic speed estimation is the aforementioned user data from a certain location. The detailed implementation steps are as follows:

[0054] Step 1 involves cleaning the raw dataset in Table 1, removing duplicate records, etc. On-road navigation data is obtained by filtering the list of apps used in the raw data, selecting navigation applications such as Didi and Gaode Maps. Pedestrian data is obtained by filtering the scanned Wi-Fi list, selecting data containing vehicle Wi-Fi signals.

[0055] Step 2: Using the data cleaned and filtered in Step 1, the trajectory data is projected onto the road network using a Hidden Markov Model (HMM) road network matching algorithm to obtain the speed X of each road segment at different time periods. Then, the road coverage of different datasets from 8:00 AM to 8:30 AM is calculated, as shown in Table 2. Each item in the table represents the coverage rate of different types of road segments under different datasets within the example time period. A higher coverage rate indicates less data loss.

[0056] Table 2 compares road coverage data from 8:00 AM to 8:30 AM for the same dataset.

[0057] Road type Road navigation data Road test pedestrian data All road sections 80.14％ 81.52％ Main road 92.26％ 91.44％ secondary roads 87.65％ 88.24％ Class III road 79.76％ 83.49％ Other roads 70.00％ 70.79％

[0058] The comparison revealed that different data coverage levels performed differently, dominating in different road segments rather than consistently performing better in all situations. On-road mobile navigation data was mainly concentrated on main roads, while mobile roadside data was more evenly distributed.

[0059] Step 3: Fill in the speed X of each road segment in different time periods obtained in Step 2. For example, calculate the missing speed data for road segments with records from March 28, 2020. To provide additional speed pattern information, use on-road navigation data and roadside pedestrian data from March 21 to March 27, 2020 as historical data. Calculations show that the data for that day represents 74.5% and 76.8% of the on-road navigation data and roadside pedestrian data, respectively. Specifically, this embodiment uses a learning rate α of 1e-4, β of 1e-4, and β1 and β2 of 0.999 in the Adam optimizer. The example first performs 10K steps of pre-training on the normal matrix factorization to obtain a good initialization, and then runs the meta-learning process of the weighted matrix for 30K steps. Finally, the matrix factorization process based on the weighted matrix was trained for 10K steps. The proposed method was compared with ordinary matrix factorization, tensor factorization, linear interpolation, GAIN, and KNN imputation methods. The experimental results are shown in Table 3.

[0060] Table 3 Comparison of data filling effects of different methods

[0061]

[0062] In the table, MAE represents mean absolute error, RMSE represents root mean square error, and MAPE represents mean absolute percentage error. The lower the error value, the better the method. As can be seen from the table, the method proposed in this embodiment has the lowest mean error value under various evaluation error standards compared with other models, which is significantly better than other methods and has a good missing data filling effect.

[0063] Step 4: Self-view speed aggregation is used to capture the spatial correlation between adjacent roads, aggregating information on neighboring road segments highly correlated with the central road segment. First, this embodiment calculates the spatial correlation between road segments based on the historical speed matrix, retaining highly correlated portions and ignoring irrelevant information.

[0064]

[0065]

[0066] The results are as follows Figure 2 As shown, the closer the geographical distance between road segments, the greater the similarity in speed; at the same time, the farther the distance, the greater the difference.

[0067] Subsequently, this embodiment calculates the fusion coefficient 'a' between road segments based on spatial correlation. i,j The velocity V after self-view aggregation was obtained:

[0068]

[0069]

[0070] The results are as follows Figure 3 As shown, the aggregation method proposed in this embodiment can effectively combine neighboring road segment information to improve the speed estimation accuracy of intermediate road segments.

[0071] Step 5: Based on the data aggregated from the self-view in Step 4, multi-view speed fusion is used to effectively integrate multi-source data, and the estimated speed is corrected according to external factors, namely road type. The results are shown in Table 4.

[0072] Table 4 Comparison of data fusion effects of different methods

[0073] Model MAE RMSE MAPE Weighted average 7.896 10.459 25.706 Gradient iteration 7.635 10.197 25.340 Linear Regression 7.467 10.094 25.901 Enhanced decision tree 7.345 9.740 25.779 Instance methods 7.320 9.837 23.877

[0074] As can be seen from the table, the error values ​​of the method proposed in this invention are smaller than those of other models under the three evaluations, and the effect is significantly better than other methods, with good data fusion effect.

[0075] The above description is merely an embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall still fall within the protection scope of the technical solutions of the present invention.

Claims

1.A method for urban traffic speed estimation based on multi-source crowd-sensing data, characterized in that, The method comprises the following steps: Step 1, data set preprocessing: cleaning the original data set collected by the smart phone, and obtaining road pedestrian data and road navigation data respectively; Step 2, average speed calculation: calculating the speed X of each road section in different time periods by using the data set in step 1; Step 3, missing data filling: using the speed X of the current each road segment in different time periods obtained in step 2, in combination with the historical speed H of each road segment in different time periods, the missing speed data in X is filled to obtain the filled speed Specifically includes: First, introduce a mask matrix M to represent the missing units of the speed X: Second, use historical data to establish a historical speed matrix H to provide additional information to help fill in the missing data, and introduce another mask matrix N to represent the missing units of the historical speed matrix H: Define a weight matrix W to measure the importance of each item in the historical speed matrix H, then use H, W, M, N and X to perform matrix decomposition: After updating the U + and V + Construct the loss by matrix factorization: Then update the weight W: Where, alpha and beta are learning rate parameters; After iterative updating, a learned weighting matrix W is obtained, which can be used to estimate the missing data simultaneously using X and H Where, lambda represents a penalty term parameter, and U and V represent two sub-matrices obtained by decomposing the original matrix; Step 4, aggregate from view speed: use the filled speed calculated in step 3 According to the historical speed H, the spatial dependence between different road segments is quantified, the useful neighbor information is collected, and the aggregated road pedestrian speed data V is obtained d And the moving navigation speed data V on the road w ; Step 5, multi-view speed fusion: according to the aggregated lane-based pedestrian speed data V d and on-road moving navigation speed data V w , the multi-source speed data is fused by using a multi-layer perceptron according to the time stamp and whether the current speed data is padding data, to obtain fused speed Finally, the fused speed is corrected according to the road type characteristics, to obtain a fine-grained large-scale urban traffic estimated speed 2.The urban traffic speed estimation method based on multi-source crowd-sensing data according to claim 1, wherein, The step 1 specifically comprises: obtaining road navigation data by filtering the APP usage list in the original data; obtaining road pedestrian data by filtering the scanned WIFI signal list in the original data according to whether there is a vehicle-mounted WIFI signal in the list. 3.The urban traffic speed estimation method based on multi-source crowd-sensing data according to claim 1, wherein, The step 2 specifically comprises: using the data filtered by step 1, using a hidden Markov road network matching algorithm to project the trajectory data into the road network, thereby obtaining the speed X of each road section in different time periods. 4.The urban traffic speed estimation method based on multi-source crowd-sensing data according to claim 1, wherein, The step 4 specifically includes: using self-view speed aggregation to capture the spatial correlation between adjacent roads, aggregating the neighbor road segment information highly correlated with the center road segment, first, calculating the spatial correlation e between road segment i and road segment j according to the historical speed matrix i,j , and retaining the highly correlated part and ignoring the irrelevant information: Afterwards, according to the spatial correlation e i,j The fusion coefficient a between road segments is calculated i,j , and the aggregated road pedestrian speed data V d and the road moving navigation speed data V w are obtained. where ε represents a small constant to prevent overflow, k represents a constant scaling value, N i represents the set of adjacent links of link i, V i , uniformly represents the link pedestrian speed data V d and the on-road moving navigation speed data V w of the i-th row in V 5. The method of claim 1, wherein, The step 5 specifically comprises: adopting multi-view speed fusion multi-source speed data, according to whether the time stamp feature is a filling data feature F d , whether the current speed data is a filling data feature F w , then passing the feature through an embedding layer, and splicing the features according to the aggregated road pedestrian speed data V d and the on-road moving navigation speed data V w obtained in step 4, to obtain a fused speed through a multi-layer perceptron Z = Embedding(Concat(F d ,F w )) Finally, the estimated speed is corrected according to the road type feature. According to historical data, the speed distribution of each type of road is obtained. The speed distribution is regarded as a normal distribution. The speed falling at the tail end of the distribution is corrected to obtain the final estimated speed

Images