A public transport travel automatic identification method and system based on multi-source data fusion

Abstract

The application relates to the computer technical field and discloses a public transport travel automatic identification method and system based on multi-source data fusion, aiming to solve the technical problems of low identification accuracy of a single data source and unsatisfactory multi-source data fusion effect. The method comprises five steps of multi-source heterogeneous data synchronous acquisition, multi-dimensional space-time data intelligent alignment, multi-source credibility dynamic evaluation, differentiated travel feature deep extraction and multi-level fusion decision. The system comprises a multi-modal data acquisition module, a space-time intelligent alignment module, a multi-source credibility evaluation module, a travel feature extraction module and a multi-level fusion decision module. Through the technical scheme, the application can realize collaborative analysis of multi-source data in a unified space-time framework, dynamically adjust the data source weight to solve the information conflict problem, and fully utilize the feature differences of different travel modes, thereby significantly improving the identification accuracy of the public transport travel mode.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to an automatic identification method and system for public transportation travel based on multi-source data fusion. Background Technology

[0002] With the acceleration of urbanization and the increasing sophistication of public transportation systems, accurately identifying residents' public transportation travel behavior is of significant value for optimizing traffic planning, improving operational management, and formulating reasonable fare policies. Currently, public transportation travel identification mainly relies on single data sources or simple combinations thereof, with mobile phone location data, bus IC card swipe data, GPS location data, and video surveillance data being common data sources. Specifically, mobile phone location data can reflect the approximate location changes of users, bus IC card data can record passenger boarding and alighting information, GPS data can provide real-time vehicle trajectories, and video surveillance data can capture passenger flow at stations.

[0003] However, existing travel identification methods based on a single data source have significant limitations. For example, mobile phone location data typically has an accuracy range of several hundred meters to several kilometers, making it difficult to accurately distinguish whether a user is using the subway, bus, or other modes of transportation, resulting in low accuracy in travel mode identification. Similarly, public transport IC card data only records passenger card swiping behavior at fixed stops, failing to capture the passenger's travel status during the journey and lacking sufficient ability to identify transfers. While GPS location data has advantages in vehicle trajectory tracking, it cannot directly determine whether a passenger actually boarded the corresponding vehicle, resulting in a "vehicle-passenger separation" problem.

[0004] To overcome the shortcomings of a single data source, existing technologies attempt to combine multiple data sources, but the fusion effect remains unsatisfactory. This is because different data sources exhibit significant differences in data collection frequency, spatial precision, and temporal granularity, making precise alignment of data across the spatiotemporal dimensions difficult. Furthermore, during the fusion process, there is a lack of effective coordination mechanisms to address information conflicts between different data sources. When travel judgments from different data sources contradict each other, existing fusion strategies cannot intelligently determine the credibility of each data source, leading to a decrease in the reliability of the recognition results. In addition, traditional data fusion methods do not adequately consider the complexity of public transportation travel scenarios and fail to effectively utilize the differences in speed, route, and dwell time characteristics among different travel modes. This results in limited improvement in recognition accuracy after multi-source data fusion, making it difficult to meet the high accuracy requirements of practical applications. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic identification method and system for public transportation travel based on multi-source data fusion, which can effectively solve the technical problems of low identification accuracy of single data source and unsatisfactory multi-source data fusion effect in the prior art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: an automatic identification method and system for public transportation travel based on multi-source data fusion, the method including a multi-source heterogeneous data synchronous acquisition step, a multi-dimensional spatiotemporal data intelligent alignment step, a multi-source credibility dynamic evaluation step, a differentiated travel feature deep extraction step, and a multi-level fusion decision step.

[0007] In the multi-source heterogeneous data synchronous acquisition step, the system synchronously acquires heterogeneous data streams including mobile phone positioning data, public transport IC card swiping data, GPS positioning data, and video surveillance data through deployed multimodal sensing devices. Specifically, this step obtains the user's coarse-grained location trajectory through the mobile communication base station positioning module, obtains the passenger's boarding and alighting station information through the public transport IC card reader, obtains the vehicle's real-time running trajectory through the vehicle-mounted GPS terminal, and obtains passenger flow capture images through the station video surveillance equipment. All sensing devices are synchronously calibrated through a unified timestamp service to construct a raw multi-source data set reflecting the user's travel behavior, providing a multi-dimensional information foundation for subsequent data fusion analysis.

[0008] In the multi-dimensional spatiotemporal data intelligent alignment step, this invention addresses the significant differences in data acquisition frequency, spatial precision, and temporal granularity among different data sources. It employs adaptive time window slicing technology and spatial grid discretization method to map multi-source data to a unified spatiotemporal reference frame. The specific process includes: dynamically adjusting the time window width according to the acquisition frequency of each data source; using an interpolation algorithm based on Kalman filtering to fill in data gaps in the time series; and unifying spatial coordinates of different precision levels to a grid system of the same resolution through grid coding. This alignment step eliminates the heterogeneity barrier of multi-source data in the spatiotemporal dimension, creating a computable unified data foundation for subsequent credibility assessment and fusion decision-making.

[0009] In the multi-source credibility dynamic evaluation step, this invention establishes a credibility quantification model based on a combination of data quality features and contextual relevance. This step comprehensively considers the signal quality, historical accuracy, and consistency with other data sources of each data source under specific spatiotemporal conditions. In a specific implementation, the system calculates a multi-dimensional credibility score for each data record. The scoring dimensions include signal strength index, location drift index, spatiotemporal continuity measurement, and cross-validation consistency. The credibility score enables dynamic adaptive adjustment of the weights of different data sources. This evaluation step effectively solves the problem of intelligent judgment when the travel judgment results given by different data sources contradict each other.

[0010] In the differentiated travel feature deep extraction step, this invention utilizes the significant differences in speed features, path features, and dwell time features among different public transportation modes to construct a multi-dimensional travel feature vector library. This step extracts feature parameters such as mean speed, speed variance, acceleration distribution, path curvature, station dwell time, and transfer interval from the spatiotemporally aligned data. Principal component analysis is then used to reduce feature dimensionality and eliminate redundancy. The extraction step fully considers the essential differences in the operating modes of different travel modes such as subways, regular buses, BRT, and taxis, forming a feature expression with class separability, providing high-dimensional feature space support for subsequent fusion decision-making.

[0011] In the multi-level fusion decision-making step, this invention adopts a two-layer fusion architecture that combines feature-level fusion and decision-level fusion. At the feature level, the feature vectors of multi-source data are serially concatenated and normalized, and then input into a deep neural network classifier for preliminary travel mode determination. At the decision level, the output results of the multi-classifier are weighted and fused by voting. The fusion weight is dynamically determined by the aforementioned credibility assessment results. Specifically, the system integrates speed feature judgment from mobile phone positioning data, station sequence matching from public transport IC card data, vehicle trajectory association from GPS data, and passenger body tracking from video surveillance data. Through the two-layer fusion architecture, collaborative decision-making from multiple information sources is achieved. This step significantly improves the recognition accuracy and reliability of results in complex travel scenarios.

[0012] Preferably, in the multi-source heterogeneous data synchronous acquisition step, mobile phone positioning data is acquired through a hybrid positioning method of Beidou navigation positioning system and mobile communication base station, with positioning accuracy controlled within a coarse-grained trajectory within a range of 500 meters; bus IC card data is uploaded in real time through an embedded vehicle terminal, with card swiping time accuracy down to the second level; GPS positioning data is acquired through Beidou navigation positioning terminal, with a sampling frequency of not less than once every 10 seconds; and video surveillance data is acquired through a high-definition network camera, with a frame rate of not less than 25 frames per second.

[0013] Preferably, in the multidimensional spatiotemporal data intelligent alignment step, the time window width is adaptively adjusted according to the characteristics of the data source. The dynamic time window width of mobile phone positioning data is 30 to 120 seconds, the window width of bus IC card data is synchronized with the bus interval, the window width of GPS data is 10 to 30 seconds, and the spatial grid discretization adopts a standard grid unit of 100 meters by 100 meters. The key area can be subdivided into 50 meters by 50 meters.

[0014] Preferably, in the multi-source credibility dynamic evaluation step, the credibility score adopts a quantitative system from 0 to 100, and the score weight of each dimension is optimized by offline training based on the historical fusion accuracy. When the travel judgment results of different data sources conflict, the system prioritizes the output of the data source with the higher credibility score, and determines whether to trigger the secondary verification process through the confidence threshold mechanism.

[0015] Preferably, in the differentiated travel feature deep extraction step, the speed feature extraction includes instantaneous speed, average speed, maximum speed, peak acceleration and average jerk; the path feature extraction includes total path length, straight-line distance, path curvature index and number of stations passed through; and the dwell feature extraction includes station dwell time, transfer waiting time and station entry recognition time. Each feature vector has a dimension of not less than 30.

[0016] Preferably, in the multi-level fusion decision-making step, feature-level fusion adopts a deep residual network classifier with no less than 18 network layers, decision-level fusion adopts a weighted Bayesian voting mechanism, the fusion weights are dynamically updated according to the real-time confidence score, and the output results include the travel mode category and the corresponding confidence value. When the confidence value is lower than the preset threshold, the system marks it as a suspected anomaly and triggers the manual review process.

[0017] The present invention also provides an automatic identification system for public transportation travel based on multi-source data fusion. The system includes a multimodal data acquisition module, a spatiotemporal intelligent alignment module, a multi-source credibility assessment module, a travel feature extraction module, and a multi-level fusion decision module.

[0018] The multimodal data acquisition module is configured to build a perception network covering the entire urban public transportation scenario. The module includes a mobile phone positioning receiving unit, a bus IC card reading unit, an on-board GPS terminal unit, and a station video monitoring unit. All acquisition units transmit data in real time to the back-end data processing center through a 5G communication network and use a unified time synchronization protocol to ensure the time consistency of the data. This module not only provides raw travel data, but also provides status identification information that reflects the data quality, providing raw information support for the upper-layer fusion logic.

[0019] The spatiotemporal intelligent alignment module is connected to the multimodal data acquisition module. It integrates an adaptive time slicing engine and a spatial grid mapping engine. This module is responsible for receiving multi-source heterogeneous data streams and using an embedded Kalman filter algorithm to perform time series interpolation and spatial coordinate standardization. This module can dynamically adjust the processing parameters according to the acquisition characteristics of each data source to achieve accurate alignment of multi-source data under a unified spatiotemporal framework, providing standardized data input for subsequent credibility assessment and feature extraction.

[0020] The multi-source credibility assessment module is connected to the spatiotemporal intelligent alignment module and is configured to perform multi-dimensional data quality measurement and credibility quantification. This module integrates a signal quality analyzer, a spatiotemporal continuity detector, and a cross-validation engine to perform real-time credibility scoring on each data record. The scoring results are output to the fusion decision module in the form of a weight matrix. This module can automatically identify data anomalies and signal interference and automatically reduce the contribution of low-quality data during the fusion process to ensure that the fusion result is dominated by high-credibility data.

[0021] The travel feature extraction module integrates a feature engineering pipeline and a deep feature learning network. This module automatically extracts multi-dimensional travel features such as speed features, path features, and stop features from standardized data after spatiotemporal alignment. It also performs feature dimensionality reduction and redundancy elimination through principal component analysis to form feature vectors with high class separability. This module also has online learning capabilities, which can continuously optimize the feature extraction strategy based on new travel data to ensure the adaptability of feature representation to actual travel scenarios.

[0022] The multi-level fusion decision module is connected to the travel feature extraction module and the multi-source credibility assessment module. It adopts a two-layer fusion architecture that combines feature-level fusion and decision-level fusion. The module contains a deep neural network classifier and a Bayesian voting fusion unit, which are responsible for the preliminary judgment at the feature level and the result optimization at the decision level, respectively. The module comprehensively considers the credibility score and feature matching degree of each data source and outputs the final travel mode identification result and confidence value. When the confidence of the identification result is lower than the preset threshold, the abnormal marking and manual review process is automatically triggered.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. The present invention discloses an automatic identification method and system for public transportation travel based on multi-source data fusion. Through the synchronous acquisition of multi-source heterogeneous data, a multi-modal perception network is constructed. It integrates four mainstream data sources: mobile phone positioning data, public transport IC card data, GPS positioning data, and video surveillance data, forming complementary travel information coverage. This technical feature breaks through the technical bottleneck of limited identification capability of a single data source, provides a rich information source foundation for subsequent multi-level fusion, and effectively solves the problem of misjudgment of travel mode caused by insufficient positioning accuracy or missing information from a single data source in the background technology.

[0025] 2. The present invention provides an automatic identification method and system for public transportation travel based on multi-source data fusion. By using a multi-dimensional spatiotemporal data intelligent alignment step, it solves the technical problem of difficulty in accurately aligning different data sources in the spatiotemporal dimension. It adopts an adaptive time window slicing technique and a spatial grid discretization method to uniformly map heterogeneous data to a standard spatiotemporal reference frame. This technical feature eliminates the fusion barriers caused by differences in data acquisition frequency, spatial accuracy, and temporal granularity, enabling multi-source data to be collaboratively analyzed under a unified computing framework, creating a computable unified data foundation for credibility assessment and feature extraction.

[0026] 3. The present invention provides an automatic public transportation travel identification method and system based on multi-source data fusion. Through a multi-source credibility dynamic evaluation step, a credibility quantification model combining data quality and contextual relevance is established. This enables intelligent judgment when travel judgment results from different data sources contradict each other. This technical feature can dynamically adjust the contribution weight of each data source in the fusion process based on multi-dimensional indicators such as signal quality, historical accuracy, and cross-validation consistency. This effectively solves the technical problem in the background technology where the lack of an effective coordination mechanism for information conflicts leads to a decrease in the reliability of identification results.

[0027] 4. The automatic identification method and system for public transportation travel based on multi-source data fusion described in this invention fully utilizes the essential differences in speed features, route features, and stop features of different public transportation modes through a differentiated travel feature deep extraction step, constructing a feature vector space with high class separability. This technical feature enables the fusion decision module to accurately distinguish between travel modes such as subway, regular bus, and BRT, significantly improving the identification accuracy after multi-source data fusion compared to traditional methods, and meeting the demand for high accuracy in practical applications. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall technical solution architecture of the public transportation travel automatic identification method based on multi-source data fusion proposed in this invention;

[0029] Figure 2 This is a schematic diagram of an automatic identification system for public transportation travel based on multi-source data fusion. Detailed Implementation

[0030] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0031] Example 1

[0032] This embodiment is applied to an automatic identification scenario for urban public transportation travel, which deploys an automatic public transportation travel identification system based on multi-source data fusion. The system includes a data acquisition module group, a data preprocessing module, a feature extraction module, a two-layer fusion decision module, and a result output module. These modules establish data interaction connections through a communication bus to jointly achieve accurate identification of users' public transportation travel behavior.

[0033] The data acquisition module group is set up at various key nodes of the urban public transportation network, responsible for collecting multi-source heterogeneous public transportation travel data. The mobile phone positioning data acquisition unit is set up at the mobile communication base station side, obtaining coarse location information of user terminals through base station positioning technology, with a collection frequency set to once per second and positioning accuracy controlled within the range of 500 to 2000 meters. The bus IC card swiping data acquisition unit is set up at the bus card readers and subway turnstiles, recording passengers' card swiping time, card swiping station number, and line number in real time, with data transmission latency controlled within 100 milliseconds. The GPS positioning data acquisition unit is set up on buses and subway trains, obtaining the vehicle's real-time latitude and longitude coordinates through onboard GPS devices, with a collection frequency set to five times per second and positioning accuracy controlled within 10 meters. The video surveillance data acquisition unit is set up at bus stops and subway entrances / exits, collecting passenger flow video streams through high-definition cameras, with a resolution of no less than 1920×1080 pixels and a frame rate of 25fps.

[0034] The data preprocessing module is located in the system's backend server cluster and includes a data cleaning unit, a data spatiotemporal alignment unit, and a data standardization unit. The data cleaning unit performs quality checks on the collected raw data, removing obviously abnormal data records, including data with location drift exceeding 5000 meters, data where the card swipe time is earlier than the vehicle's arrival time, and data with GPS coordinate jumps exceeding 1000 meters. The data spatiotemporal alignment unit performs precise spatiotemporal alignment of data from different data sources based on a unified time base and spatial reference system. It uses a linear interpolation algorithm to unify data from different collection frequencies to a 1-second sampling interval and a Gaussian projection algorithm to unify spatial data from different coordinate systems to the same coordinate system. The data standardization unit converts the cleaned and aligned data into a unified data format, forming standardized travel data records. Each record includes fields such as user identifier, timestamp, longitude, latitude, and travel status tag.

[0035] The feature extraction module is located behind the data preprocessing module and includes a spatial feature extraction unit, a temporal feature extraction unit, and a behavioral feature extraction unit. The spatial feature extraction unit extracts spatial attribute features of users from standardized data, including instantaneous speed, trajectory curvature, path offset, and station dwell time. Instantaneous speed is obtained by dividing the Euclidean distance between two adjacent locations by the time difference; trajectory curvature is calculated using a three-point circular arc fitting method; path offset is obtained by comparing the deviation distance between the actual trajectory and the theoretical route; and station dwell time is obtained by detecting the time the user spends in the station area. The temporal feature extraction unit extracts temporal attribute features of users from standardized data, including travel time period, transfer interval, and waiting time. Travel time period is obtained by dividing 24 hours into several time slices and marking the time slice to which the user's travel belongs; transfer interval is obtained by detecting the time difference between two consecutive card swipe records; and waiting time is obtained by detecting the time the user spends in the platform area. The behavioral feature extraction unit extracts behavioral attribute features of users from standardized data, including station entry frequency, travel frequency, and frequently used routes. Station entry frequency is obtained by counting the number of times a user enters a station within a unit of time. Trip frequency is obtained by counting the number of trips a user takes each day. Frequently used routes are obtained by analyzing a user's historical travel data.

[0036] The dual-layer fusion decision module is the core processing unit of this invention, comprising a feature-level fusion unit, a decision-level fusion unit, and a credibility assessment unit. The feature-level fusion unit fuses multi-source features from the feature extraction module, merging features extracted from different data sources using a weighted average method. The weights are dynamically adjusted based on the credibility of each data source in its corresponding feature dimension. The credibility assessment unit calculates the credibility weight of each data source based on data quality indicators, feature consistency indicators, and time continuity indicators. Data quality indicators include data completeness, data accuracy, and data timeliness; feature consistency indicators include the feature similarity of the same sample across different data sources; and time continuity indicators include the continuity of the data collection time series. The decision-level fusion unit determines the user's travel mode based on the feature-level fusion results and preset travel mode identification rules. Travel modes include walking, cycling, taking public transportation, taking the subway, taxis, and ride-hailing services. The decision-level fusion unit internally maintains a classification model trained using a machine learning algorithm. This model takes the fused feature vector as input and outputs the probability distribution of each travel mode.

[0037] The results output module is located at the system front end and includes a travel mode identification unit and a results storage unit. The travel mode identification unit determines the final travel mode judgment based on the output of the decision-level fusion unit and generates a results report containing travel mode, travel time, travel route, and transfer information. The results storage unit persistently stores the identification results in a database to support subsequent data analysis and business applications.

[0038] When users travel by public transportation, the mobile phone positioning data acquisition unit continuously acquires the user's terminal location information through the mobile communication network and sends the location data to the data preprocessing module. The bus IC card swiping data acquisition unit immediately captures the card swiping record when the user swipes their card, including information such as card number, swiping time, station number, and route number, and sends the record to the data preprocessing module. The GPS positioning data acquisition unit acquires the real-time location information of buses and subway trains and sends the vehicle location data to the data preprocessing module. The video surveillance data acquisition unit continuously collects video streams from platforms and subway entrances / exits, and uses video analysis algorithms to detect passenger flow, sending the passenger flow data to the data preprocessing module.

[0039] After receiving raw data from the data acquisition module group, the data preprocessing module first performs quality checks and processing by the data cleaning unit, removing abnormal and invalid data. The cleaned data then enters the data spatiotemporal alignment unit. This unit aligns data from different data sources along the timeline based on a unified timestamp and uses spatial interpolation methods to spatially align data from different acquisition frequencies, ensuring consistency in the spatiotemporal dimension for data from different data sources. The standardization unit converts the aligned data into a standardized data format, generating data records containing complete spatiotemporal information.

[0040] Standardized data records are fed into the feature extraction module for feature extraction processing. The spatial feature extraction unit extracts spatial attribute features from each data record, including calculating instantaneous velocity based on consecutive location points, calculating trajectory curvature based on trajectory point sequences, calculating path offset based on a comparison of the actual trajectory and the theoretical route, and calculating station dwell time based on detection results within the station area. The temporal feature extraction unit extracts temporal attribute features from the data records, including determining the travel period based on card swipe time, calculating transfer intervals based on consecutive card swipe records, and calculating waiting time based on platform area detection results. The behavioral feature extraction unit extracts behavioral attribute features from historical data, including calculating station entry frequency, travel frequency, and frequently used routes based on historical travel records.

[0041] The extracted multi-source features are sent to the two-layer fusion decision module for processing. In the feature-level fusion stage, the feature-level fusion unit receives feature vectors from the feature extraction module, while the credibility assessment unit simultaneously calculates the credibility weights of each data source. The credibility assessment unit first calculates the data quality indicators for each data source, including data completeness (number of valid data fields divided by the total number of fields), data accuracy (the ratio of consistency with reference data), and data timeliness (the difference between the data collection time and the current time). Then, it calculates the feature consistency indicator, which is obtained by comparing the cosine similarity of feature vectors extracted from different data sources by the same user within the same time window. Finally, it calculates the time continuity indicator, which is obtained by analyzing the interval distribution of the data collection time series. Based on these three indicators, the credibility assessment unit uses a weighted summation method to calculate the comprehensive credibility weight of each data source and passes it to the feature-level fusion unit. The feature-level fusion unit performs weighted fusion of features from different data sources according to the credibility weights, forming a fused feature vector.

[0042] In the decision-level fusion stage, the decision-level fusion unit receives the fused feature vector as input and feeds it into a pre-trained classification model to determine the mode of travel. The classification model is trained using either a support vector machine (SVM) or random forest algorithm, with the fused feature vector as input and the probability distribution of each mode of travel as output. The decision-level fusion unit selects the mode of travel with the highest probability as the initial determination. However, when the probability differences among multiple modes of travel are small or there are significant conflicts, the decision-level fusion unit triggers a credibility reassessment process. This reassessment process first checks the data quality changes of each data source within the current time window. If a significant decline in the data quality of a particular data source is found, its credibility weight is reduced. Then, it checks whether the feature values ​​of each dimension in the fused feature vector are within the typical feature range of the corresponding mode of travel. If there is a significant deviation, that feature dimension is marked as an anomaly. Finally, based on the reassessment results, the decision boundary of the classification model is adjusted, the probability distribution of each mode of travel is recalculated, and the final mode of travel determination is obtained.

[0043] The results output module receives the final judgment result from the decision-level fusion unit, and the travel mode identification unit generates a structured travel report, including information such as user identifier, travel mode, travel start time, travel end time, departure station, arrival station, number of transfers, and transfer routes. The results storage unit writes the travel report to the database and also supports real-time push to user terminals or traffic management department systems.

[0044] In this embodiment, the system acquires multi-source heterogeneous data through a data acquisition module group. After cleaning and alignment by the data preprocessing module, multi-dimensional features are extracted using the feature extraction module. Then, through two stages of feature-level fusion and decision-level fusion by the two-layer fusion decision module, the system ultimately achieves accurate identification of the user's public transportation travel mode. This system and method effectively solve the technical problems of low identification accuracy from a single data source and poor fusion effect of multi-source data, thus improving the accuracy and reliability of public transportation travel identification.

[0045] Example 2

[0046] Building upon Example 1, this example optimizes public transportation travel identification in high-density passenger flow scenarios. This scenario deploys an enhanced multi-source data fusion identification system, primarily targeting dense passenger flow at bus stops and subway stations during morning and evening rush hours.

[0047] The enhanced system adds a high-density passenger flow data acquisition unit to the data acquisition module group. This unit includes a multi-camera collaborative tracking subunit and an infrared people counting subunit. The multi-camera collaborative tracking subunit is deployed at key locations on the platform, using video streams acquired from multi-angle cameras for human detection and tracking, enabling accurate counting and trajectory tracking of passenger flow entering and exiting the station. The infrared people counting subunit is installed above the turnstile channels, using infrared beam sensors to count the number of people passing through, serving as a supplementary verification of passenger flow data.

[0048] In the data preprocessing module, this embodiment adds a passenger flow data fusion unit. This unit is responsible for fusing video surveillance data and infrared counting data to generate high-precision passenger flow data. The passenger flow data fusion unit uses a multi-sensor fusion algorithm based on Kalman filtering to model and compensate for errors in video counting and infrared counting, thereby improving the accuracy of passenger flow counting.

[0049] In the feature extraction module, this embodiment adds a passenger flow density feature extraction unit. This unit extracts features such as platform passenger flow density, queue length, and congestion level from the fused passenger flow data. Platform passenger flow density is calculated by the number of passengers per unit area, queue length is calculated by converting the pixel length of the waiting queue detected by video analysis, and congestion level is calculated by the ratio of the carriage's passenger capacity to the actual passenger volume.

[0050] In the two-layer fusion decision module, this embodiment improves the feature-level fusion unit by adding a weighted adjustment mechanism based on passenger flow characteristics. When the passenger flow density at the platform exceeds a preset threshold, the feature-level fusion unit automatically increases the fusion weight of bus IC card data and video surveillance data while decreasing the fusion weight of mobile phone location data, because mobile phone location data is more prone to location drift in high-density scenarios. Simultaneously, the decision-level fusion unit adds travel mode auxiliary determination rules based on passenger flow characteristics. When a high degree of match is detected between a user's card swiping record and platform passenger flow data, the confidence level of the public transport travel mode determination can be improved.

[0051] The workflow of this embodiment adds the following processing steps to the workflow of Embodiment 1: A multi-camera collaborative tracking subunit collects real-time video streams from the platform, performs human detection and target tracking, and generates passenger flow trajectory data; an infrared people counting subunit counts the number of passengers passing through the passage and generates infrared counting data; a passenger flow data fusion unit performs Kalman filtering on the video counts and infrared counts to generate high-precision passenger flow data; a passenger flow density feature extraction unit extracts density features from the passenger flow data and adds them to the feature vector; a feature-level fusion unit automatically adjusts the fusion weights of each data source according to the passenger flow density; and a decision-level fusion unit determines the travel mode based on the passenger flow features.

[0052] Through the improvements in this embodiment, the system can more accurately identify public transportation travel behavior in high-density passenger flow scenarios, effectively cope with data noise and fusion challenges in complex environments, and further improve the applicability and recognition accuracy of the system.

[0053] Example 3

[0054] Building upon Example 1, this example optimizes scenarios for intercity travel and combined travel modes. This scenario covers the identification needs of various travel modes, including intercity rail, highway passenger transport, and urban public transport.

[0055] The cross-scenario system has added railway ticketing data acquisition units and highway passenger transport data acquisition units to its data acquisition module group. The railway ticketing data acquisition unit interfaces with the railway ticketing system to obtain users' ticket purchase and ticket inspection records, including train information, departure station, arrival station, and travel time. The highway passenger transport data acquisition unit interfaces with the passenger transport company's management system to obtain passengers' ticket purchase and ticket inspection information.

[0056] In the data preprocessing module, this embodiment adds a multi-modal data association unit. This unit is responsible for performing correlation analysis on data from different modes of transportation to identify the user's complete travel chain. The multi-modal data association unit uses a time- and space-based sequence alignment algorithm to link the user's railway travel records, highway passenger transport records, and urban public transport card swipe records to reconstruct the user's complete travel trajectory.

[0057] In the feature extraction module, this embodiment adds a cross-modal feature extraction unit, which extracts transfer features, connection features, and long-distance travel features from multi-modal travel data. Transfer features include connection time between different modes of transportation, transfer spatial distance, and transfer convenience evaluation. Connection features include the degree of matching between two adjacent modes of transportation in time and space. Long-distance travel features include the total mileage of a single trip, travel time, and average speed.

[0058] In the two-layer fusion decision module, this embodiment extends the decision-level fusion unit by adding a multi-modal travel chain identification model. This model employs a sequence-to-sequence learning framework based on recurrent neural networks, taking the user's travel event sequence as input and outputting a prediction of the complete travel chain. The travel event sequence includes elements such as travel mode, departure time, arrival time, and origin and destination locations. By learning from a large amount of historical travel data, the model has mastered typical patterns of different travel chains and can identify whether a user uses "high-speed rail + subway," "bus + public transport," or other combinations to complete intercity travel.

[0059] The workflow of this embodiment adds the following processing steps to the workflow of embodiment one: the railway ticketing data collection unit acquires the user's railway travel records, and the highway passenger transport data collection unit acquires the user's highway passenger transport records; the multi-mode transportation data association unit performs spatiotemporal alignment and association of data from different modes of transportation to generate a complete travel event sequence; the cross-mode feature extraction unit extracts transfer features, connection features, and long-distance travel features from the travel event sequence; the multi-mode travel chain identification model analyzes the travel event sequence and predicts the user's complete travel chain; the result output module outputs a comprehensive travel identification result containing multiple modes of transportation.

[0060] Through the improvements in this embodiment, the system can identify public transportation travel behavior in intercity travel and multi-modal travel scenarios, expanding the application scope of the system and meeting broader business needs such as intercity travel analysis and transportation hub passenger flow analysis.

Claims

1. A method for automatic identification of public transportation trips based on multi-source data fusion, characterized in that, Includes the following steps: S1. Simultaneously acquire mobile phone positioning data, public transport IC card swiping data, GPS positioning data and video surveillance data through multimodal sensing devices, and perform synchronous calibration of each sensing device through a unified timestamp service; S2. Adaptive time window slicing technology and spatial grid discretization method are used to map multi-source data to a unified spatiotemporal reference frame, and Kalman filter interpolation algorithm is used to fill in data gaps; S3. Establish a quantitative model based on data quality characteristics and contextual credibility, calculate a multidimensional credibility score for each data record, and dynamically adjust the weights of each data source according to the score; S4. Extract speed features, path features, and dwell features from the spatiotemporally aligned data, and perform feature dimensionality reduction through principal component analysis to form feature vectors with class separability. S5. A two-layer fusion architecture combining feature-level fusion and decision-level fusion is adopted. At the feature level, multi-source feature vectors are serially concatenated and input into a deep neural network classifier for preliminary judgment. At the decision level, the output results of multiple classifiers are weighted and fused by voting to output the travel mode recognition result.

2. The method according to claim 1, characterized in that, The mobile phone location data mentioned in S1 is obtained through a hybrid positioning method combining navigation positioning system and mobile communication base station.

3. The method according to claim 1, characterized in that, The time window width described in S2 is dynamically adjusted according to the data source acquisition frequency, and the spatial grid adopts standardized grid cells.

4. The method according to claim 1, characterized in that, The credibility scoring dimensions described in S3 include signal strength index, position drift index, spatiotemporal continuity measure, and cross-validation consistency.

5. The method according to claim 1, characterized in that, The speed characteristics described in S4 include instantaneous speed, mean speed, and acceleration distribution, while the path characteristics include path curvature and station dwell time.

6. The method according to claim 1, characterized in that, The feature-level fusion described in S5 uses a deep residual network classifier, while the decision-level fusion uses a weighted Bayesian voting mechanism.

7. The method according to claim 1, characterized in that, Also includes: When the confidence level of the recognition result is lower than the preset threshold, the anomaly marking process is triggered.

8. An automatic public transportation travel identification system based on multi-source data fusion, characterized in that, include: The system includes a multimodal data acquisition module, which synchronously acquires heterogeneous data streams through a mobile phone positioning receiving unit, a public transport IC card reading unit, an on-board GPS terminal unit, and a station video monitoring unit. Each acquisition unit transmits data in real time to the back-end data processing center via a communication network. A spatiotemporal intelligent alignment module, connected to the multimodal data acquisition module, uses an adaptive time slicing engine and a spatial grid mapping engine, employing a Kalman filter algorithm for time series interpolation and spatial coordinate standardization. A multi-source credibility assessment module, connected to the spatiotemporal intelligent alignment module, performs real-time credibility scoring on data records using a signal quality analyzer, a spatiotemporal continuity detector, and a cross-validation engine. A travel feature extraction module extracts speed, path, and dwell features from the standardized spatiotemporally aligned data, performing feature dimensionality reduction through principal component analysis. A multi-level fusion decision module, connected to both the travel feature extraction module and the multi-source credibility assessment module, outputs travel mode identification results using a two-layer fusion architecture combining feature-level fusion and decision-level fusion.

9. The system according to claim 8, characterized in that, The multimodal data acquisition module uses a unified time synchronization protocol to ensure data time consistency.

10. The system according to claim 8, characterized in that, The multi-level fusion decision module contains a deep neural network classifier and a Bayesian voting fusion unit, which are used to make preliminary judgments at the feature level and optimize the results at the decision level.

Images