A bus passenger flow counting method based on deep learning

Abstract

This invention provides a deep learning-based method for bus passenger flow statistics, belonging to the field of intelligent transportation technology. The method includes: acquiring a bus structure diagram; deploying a detection device; collecting multi-source data based on the detection device; classifying the multi-source data to determine vehicle data and passenger data; detecting passenger boarding and alighting behavior based on the vehicle and passenger data; performing magnitude analysis on the behavior detection results to generate a structured magnitude report; correlating the structured magnitude report with facial recognition model parameters to obtain correlation results; configuring a facial recognition model based on the correlation results; fusing historical passenger flow, vehicle BeiDou trajectory, and weather data based on the facial recognition model to form a prediction model; predicting the number of passengers; and pushing the predicted number of passengers to the dispatch center to optimize vehicle allocation, achieving data-driven dynamic vehicle allocation, improving the accuracy of perception, realizing intelligent prediction of passenger flow, and optimizing resource allocation.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method for passenger flow statistics on buses based on deep learning. Background Technology

[0002] With the acceleration of urbanization, public transportation systems face three major challenges: large fluctuations in passenger flow, delayed scheduling, and waste of resources. Traditional scheduling relies on experience, with manual passenger flow statistics having an error rate as high as 20-30%, and cannot cope with sudden large passenger flows. Vehicle monitoring, passenger flow counting, and weather data are scattered and independent, lacking the ability to integrate and analyze multi-source information, resulting in serious response delays.

[0003] Therefore, this invention provides a method for passenger flow statistics on buses based on deep learning. Summary of the Invention

[0004] This invention provides a deep learning-based method for bus passenger flow statistics. It collects vehicle and passenger data in real time through multi-source sensors and combines behavior detection algorithms to identify boarding and alighting actions. Based on passenger flow density, it dynamically adjusts the parameters of the facial recognition model and integrates historical passenger flow, BeiDou trajectory, and weather data to construct a prediction model. Finally, the prediction results are pushed to the dispatch center to achieve data-driven dynamic vehicle allocation, improve the accuracy of perception, realize intelligent prediction of passenger flow, and optimize resource allocation.

[0005] This invention provides a deep learning-based method for bus passenger flow statistics, comprising:

[0006] Step 1: Obtain a bus structure diagram, deploy a detection device based on the bus structure diagram, and collect multi-source data based on the detection device;

[0007] Step 2: Classify the multi-source data to determine vehicle data and passenger data. Based on the vehicle data and passenger data, detect passenger boarding and alighting behavior to obtain the behavior detection results.

[0008] Step 3: Perform magnitude analysis on the behavior detection results to obtain a structured magnitude report, correlate the face recognition model parameters on the structured magnitude report to obtain the correlation results, and configure the face recognition model based on the correlation results;

[0009] Step 4: Based on the facial recognition model, historical passenger flow, vehicle Beidou trajectory and weather data are integrated to form a prediction model, obtain the predicted change in the number of passengers, and push the predicted change in the number of passengers to the dispatch center to optimize vehicle allocation.

[0010] This invention provides a deep learning-based method for bus passenger flow statistics, which involves obtaining a bus structure map, deploying a detection device based on the bus structure map, and collecting multi-source data based on the detection device, including:

[0011] Based on the bus structure diagram, a multi-source sensor array is determined, and the multi-source sensor array is used to perform digital twin verification on the bus. The actual layout of the multi-source sensor array is then determined based on the verification results.

[0012] Based on the actual deployment of the corresponding detection devices, multi-source data acquisition is carried out.

[0013] This invention provides a deep learning-based method for bus passenger flow statistics. The method involves determining a multi-source sensor array based on a bus structure diagram, using the multi-source sensor array to perform digital twin verification of the bus, and determining the actual arrangement of the multi-source sensor array based on the verification results. The method includes:

[0014] The bus structural diagram is analyzed to extract key dimension parameters, and the existing equipment inside the bus is marked with its location. Based on the key dimension parameters and location marks, available routes are determined.

[0015] Based on the available lines, the carriage is divided into multiple detection areas. Each detection area is configured with a main-auxiliary sensor pair to form a spatial dimension matrix. A sensor capability complementarity table is established based on the main-auxiliary sensor pair to form a functional dimension matrix. The spatial dimension matrix and the functional dimension matrix are combined to obtain a multi-source sensor array.

[0016] The bus is digitally twinned based on the multi-source sensor array, generating basic performance verification results, extreme scenario verification results, and multimodal fusion verification results.

[0017] The first layout is derived based on basic performance verification, the second layout is derived based on extreme scenario verification results, and the third layout is derived based on multimodal fusion verification results.

[0018] By combining the first, second, and third arrangements, Pareto optimality calculations are performed to obtain the actual arrangement of the multi-source sensor array.

[0019] This invention provides a deep learning-based method for bus passenger flow statistics, which classifies multi-source data to determine vehicle data and passenger data, detects passenger boarding and alighting behavior based on the vehicle and passenger data, and obtains behavior detection results, including:

[0020] Preprocess the multi-source data, determine the data type of the preprocessing results, and determine the corresponding feature extraction method and output dimension from the type-feature engineering table based on the data type;

[0021] Based on the aforementioned feature extraction method and output dimensions, vehicle data and passenger data are determined, thereby determining the vehicle data classification model, the passenger data classification model, and multimodal collaboration.

[0022] Based on the vehicle data classification model, passenger data classification model, and multimodal collaboration, passenger boarding and alighting behaviors are detected, and then the behavior detection results are obtained.

[0023] This invention provides a deep learning-based method for bus passenger flow statistics. It detects passenger boarding and alighting behaviors based on a vehicle data classification model, a passenger data classification model, and multimodal collaboration, thereby deriving behavior detection results. The method includes:

[0024] Based on the vehicle data classification model, vehicle features are distinguished; based on the passenger data analysis model, passenger features are identified; and the multimodal collaborative approach is used to formulate a feature fusion strategy corresponding to vehicle features and passenger features.

[0025] The vehicle features and passenger features are fused according to the feature fusion strategy to obtain the fusion result. The fusion result is then analyzed to detect passenger boarding and alighting behavior, and the behavior detection result is obtained.

[0026] This invention provides a deep learning-based method for bus passenger flow statistics, which involves performing magnitude analysis on behavior detection results to generate a structured magnitude report, correlating the structured magnitude report with facial recognition model parameters to obtain correlation results, and configuring a facial recognition model based on the correlation results. The method includes:

[0027] The behavior detection results are aggregated in the spatiotemporal dimensions, and the distribution of passenger boarding and alighting, behavior types and regional density within a unit time window are statistically analyzed. The magnitude is divided into levels based on preset thresholds, and a structured magnitude report is output.

[0028] Based on the structured volumetric report, volumetric features are extracted, and these features are associated with the key parameters of the face recognition model to obtain the association results. Then, a hierarchical configuration is performed based on the association results, and the face recognition model is obtained by combining all the hierarchical configurations.

[0029] This invention provides a deep learning-based method for bus passenger flow statistics. It integrates historical passenger flow data, vehicle BeiDou navigation data, and weather data using a facial recognition model to form a predictive model, derives a predicted change in passenger numbers, and pushes this predicted change in passenger numbers to the dispatch center to optimize vehicle allocation. The method includes:

[0030] Real-time passenger flow is obtained from the facial recognition system, and historical passenger flow is extracted from the historical passenger flow database. Vehicle Beidou trajectories are collected simultaneously, and weather data is obtained through the meteorological interface.

[0031] The time dimension is decomposed into periodic time features, while the topological features of the site are obtained.

[0032] Based on periodic time characteristics, ordinary time-series feature vectors and key time-series feature vectors are extracted from real-time passenger flow and historical passenger flow. Static features are extracted from weather data. The ordinary time-series feature vectors and the static features are first concatenated, and the key time-series feature vectors and the static features are second concatenated.

[0033] Based on the topological features, the first and second splices are connected to form a prediction model, which yields the predicted number of people to be affected. The predicted number of people to be affected is then pushed to the dispatch center to optimize vehicle allocation.

[0034] This invention provides a deep learning-based method for bus passenger flow statistics, which pushes predicted passenger flow changes to the dispatch center to optimize vehicle allocation, including:

[0035] Match key causal chains based on predicted changes in the number of people, and trigger resource allocation schemes based on the key causal chains.

[0036] Based on the resource allocation plan, a sand table simulation is performed, and the plan evaluation results are output. The dispatch center determines the dispatch instructions and optimizes vehicle allocation based on the plan evaluation results.

[0037] Compared with existing technologies, the beneficial effects of this application are as follows: real-time collection of vehicle and passenger data through multi-source sensors, combined with behavior detection algorithms to identify boarding and alighting actions; dynamic adjustment of facial recognition model parameters based on passenger flow density, and construction of a prediction model by integrating historical passenger flow, BeiDou trajectory and weather data; finally, pushing the prediction results to the dispatch center to realize data-driven dynamic vehicle allocation, improve the accuracy of perception, realize intelligent prediction of passenger flow, and optimize resource allocation.

[0038] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0039] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0040] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0041] Figure 1 This is a flowchart illustrating a deep learning-based method for counting bus passenger flow, as provided in an embodiment of the present invention. Detailed Implementation

[0042] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Example 1:

[0043] This invention provides a method for bus passenger flow statistics based on deep learning, such as... Figure 1 As shown, it includes:

[0044] Step 1: Obtain a bus structure diagram, deploy a detection device based on the bus structure diagram, and collect multi-source data based on the detection device;

[0045] Step 2: Classify the multi-source data to determine vehicle data and passenger data. Based on the vehicle data and passenger data, detect passenger boarding and alighting behavior to obtain the behavior detection results.

[0046] Step 3: Perform magnitude analysis on the behavior detection results to obtain a structured magnitude report, correlate the face recognition model parameters on the structured magnitude report to obtain the correlation results, and configure the face recognition model based on the correlation results;

[0047] Step 4: Based on the facial recognition model, historical passenger flow, vehicle Beidou trajectory and weather data are integrated to form a prediction model, obtain the predicted change in the number of passengers, and push the predicted change in the number of passengers to the dispatch center to optimize vehicle allocation.

[0048] In this embodiment, the bus construction drawing is a CAD design drawing or 3D model of the bus, which includes engineering data such as the carriage structure and equipment layout. For example, the door position and step height markings of a low-floor bus, and the stair position and upper seat distribution diagram of a double-decker bus.

[0049] In this embodiment, the detection device is a deployed physical sensor and supporting equipment, such as the Hikvision DS-2CD3 series vehicle camera.

[0050] In this embodiment, the initial layout of multi-source sensors (such as cameras, radar, and pressure sensors) is designed using a bus structural diagram. Real-world operating scenarios (such as passenger flow, obstruction, and extreme weather) are simulated in a digital twin environment to verify sensor performance. Based on the simulation results, the installation location and parameters are optimized to ultimately guide physical deployment, ensuring data collection coverage without blind spots, achieving precise deployment, and improving robustness.

[0051] In this embodiment, vehicle data refers to data related to the vehicle, such as the vehicle's speed, acceleration, and location, for example, the vehicle's GPS location data and speedometer data.

[0052] In this embodiment, passenger data refers to passenger-related data, such as passenger age, gender, behavior, etc., for example, passenger sensor data through the car door, passenger images captured by the camera.

[0053] In this embodiment, the behavior detection result refers to the passenger behavior judgment derived from the analysis of the fusion results. This is usually based on the output of a machine learning model or rule engine. For example, if the feature fusion result of passenger A shows that the vehicle is decelerating and the passenger is standing near the door, the behavior detection result is that passenger A is preparing to get off the vehicle. If the feature fusion result of passenger B shows that the vehicle is accelerating and the passenger has just passed through the door, the behavior detection result is that passenger B has just boarded the vehicle.

[0054] In this embodiment, multi-source heterogeneous data is processed through intelligent classification and feature engineering to construct classification models for vehicle and passenger data respectively. Combined with multimodal collaborative algorithms to fuse spatiotemporal features, high-precision detection of boarding and alighting behavior is achieved. The system automatically matches data types and feature extraction methods and dynamically optimizes the detection process.

[0055] In this embodiment, the structured quantitative report contains density level, behavioral distribution characteristics and anomaly markers. It is a standardized statistical result after being sorted out, usually in JSON or database records. For example, the report contains time period 09:00-10:00, area A1 exit, 200 passengers boarding, and high density level.

[0056] In this embodiment, the association result is the matching rule between features and parameters. For example, when "red density + secondary behavior warning" occurs at the same time, the face capture frame rate is increased from 15fps to 30fps, and the similarity threshold is increased from 0.85 to 0.92.

[0057] In this embodiment, the face recognition model is a dynamically adjusted complete recognition system. For example, in the railway station system during the Spring Festival travel rush: from 6:00 to 9:00, it automatically activates "peak mode": reducing the resolution but doubling the analysis frequency. When a fugitive is identified, it instantly switches to "precise mode": the highest resolution image quality + strict comparison.

[0058] In this embodiment, passenger flow behavior data is analyzed by spatiotemporal aggregation to generate a structured volumetric report. Based on passenger flow density, the parameters of the face recognition model (such as detection frame rate and ROI region) are dynamically adjusted to achieve closed-loop optimization of "passenger flow perception-model adaptation" and ensure optimal recognition performance in different scenarios.

[0059] In this embodiment, historical passenger flow refers to passenger flow records from the same time period in the past (such as the same period last year / the same day last week). For example, the total passenger flow of this station on the first day of National Day last year was 100,000.

[0060] In this embodiment, the vehicle's BeiDou trajectory is the real-time location and movement path of the bus / shuttle obtained through BeiDou satellite positioning. For example, the trajectory point sequence of bus A is: [(longitude X1, latitude Y1, time T1), (X2, Y2, T2)...].

[0061] In this embodiment, the weather data is real-time weather information (such as temperature, rainfall, and wind speed) obtained through an API. For example, the current weather data is: {"Temperature":28℃, "Rainfall":5mm, "Wind":Level 3}.

[0062] In this embodiment, the input to the prediction model is the combined features after splicing, and the output is the predicted passenger flow value for the future period. For example, the input is: ["morning peak", "transfer station", 500, "+50%", "heavy rain"], and the output is: predicted number of people entering the station in the next hour is 650 (30% more than usual).

[0063] In this embodiment, a hybrid prediction model with enhanced spatiotemporal features is constructed by integrating real-time facial recognition passenger flow, historical passenger flow patterns, vehicle BeiDou trajectories, and weather data. The model combines periodic passenger flow trends with station topology relationships to dynamically generate passenger flow prediction results, driving intelligent scheduling decisions and achieving data-driven precise capacity allocation. Then, by analyzing passenger flow prediction data and key causal chains (such as weather-passenger flow-resource demand), a resource allocation plan is dynamically generated. After verifying the effectiveness through digital twin simulation, the optimal scheduling command is output, achieving closed-loop optimization from prediction to execution.

[0064] The working principle and beneficial effects of the above technical solution are as follows: real-time data collection of vehicles and passengers is achieved through multi-source sensors, and behavior detection algorithms are used to identify boarding and alighting actions; the parameters of the facial recognition model are dynamically adjusted based on passenger flow density, and a prediction model is constructed by integrating historical passenger flow, Beidou trajectory and weather data; finally, the prediction results are pushed to the dispatch center to realize data-driven dynamic vehicle allocation, improve the accuracy of perception, realize intelligent prediction of passenger flow, and optimize resource allocation. Example 2:

[0065] This invention provides a deep learning-based method for bus passenger flow statistics, which involves obtaining a bus structure map, deploying a detection device based on the bus structure map, and collecting multi-source data based on the detection device. The method includes:

[0066] Based on the bus structure diagram, a multi-source sensor array is determined, and the multi-source sensor array is used to perform digital twin verification on the bus. The actual layout of the multi-source sensor array is then determined based on the verification results.

[0067] Based on the actual deployment of the corresponding detection devices, multi-source data acquisition is carried out.

[0068] In this embodiment, the multi-source sensor array is a heterogeneous sensor combination deployed inside the vehicle compartment to collect multi-dimensional data. For example, the visual sensor is a 1080P binocular camera on the top of the door (to detect passenger flow), the radar sensor is a 60GHz millimeter-wave radar on the side wall of the vehicle compartment (to detect through obstructions), and the environmental sensor is a composite probe of temperature, humidity and illuminance near the driver's cab.

[0069] In this embodiment, digital twin verification is a test process that simulates sensor performance in a virtual environment. For example, in Unity3D, the camera recognition rate is simulated under heavy rain, and in ANSYS, the impact of electromagnetic interference on radar point clouds is simulated.

[0070] In this embodiment, the verification results are quantitative performance indicators output by the digital twin test. For example, the camera's false negative rate in backlit scenes is ≤3%, and the radar's multi-target resolution accuracy in crowded conditions is ±5cm.

[0071] In this embodiment, the actual arrangement is an optimized physical installation scheme for the sensors. For example, the front door camera is installed at a height of 2.4 meters from the ground with a downward angle of 12°. The pressure sensor is arranged with three measuring points evenly distributed under the step plate.

[0072] The working principle and beneficial effects of the above technical solution are as follows: the initial layout of multi-source sensors (such as cameras, radar, and pressure sensors) is designed by using the bus structure diagram, and real operating scenarios (such as passenger flow, obstruction, and extreme weather) are simulated in the digital twin environment to verify the sensor performance; the installation position and parameters are optimized based on the simulation results, and finally the physical deployment is guided to ensure that the data collection coverage is without blind spots, achieve accurate deployment, and improve robustness. Example 3:

[0073] This invention provides a deep learning-based method for bus passenger flow statistics. The method involves determining a multi-source sensor array based on a bus structure diagram, using the multi-source sensor array to perform digital twin verification of the bus, and determining the actual arrangement of the multi-source sensor array based on the verification results. The method includes:

[0074] The bus structural diagram is analyzed to extract key dimension parameters, and the existing equipment inside the bus is marked with its location. Based on the key dimension parameters and location marks, available routes are determined.

[0075] Based on the available lines, the carriage is divided into multiple detection areas. Each detection area is configured with a main-auxiliary sensor pair to form a spatial dimension matrix. A sensor capability complementarity table is established based on the main-auxiliary sensor pair to form a functional dimension matrix. The spatial dimension matrix and the functional dimension matrix are combined to obtain a multi-source sensor array.

[0076] The bus is digitally twinned based on the multi-source sensor array, generating basic performance verification results, extreme scenario verification results, and multimodal fusion verification results.

[0077] The first layout is derived based on basic performance verification, the second layout is derived based on extreme scenario verification results, and the third layout is derived based on multimodal fusion verification results.

[0078] By combining the first, second, and third arrangements, Pareto optimality calculations are performed to obtain the actual arrangement of the multi-source sensor array.

[0079] In this embodiment, key dimensional parameters refer to dimensions that are crucial for design and analysis in the bus structural drawings, such as the length, width, and height of the passenger compartment, the dimensions of the seats, and the width of the aisles. These parameters are essential for determining the location of internal equipment and route layout of the bus; for example, the passenger compartment length (10 meters), width (2.5 meters), and height (3 meters).

[0080] In this embodiment, the analysis process refers to the action of extracting these key dimensional parameters from the bus construction drawings, such as using CAD software to measure and record the various dimensions of the carriage.

[0081] In this embodiment, available wiring refers to wiring inside the bus that can be used to install sensors or other equipment, determined based on the location and key dimensional parameters of the equipment, such as aisles on the floor of the bus or the space under the seats.

[0082] In this embodiment, location marking refers to marking the location of existing equipment (such as seats, handrails, emergency exits, etc.) inside the bus so that these locations can be taken into account in subsequent designs. For example, a point is marked at each seat location on the drawing.

[0083] In this embodiment, the process of dividing the detection area refers to dividing the carriage into multiple areas based on the available lines to facilitate the placement of sensors. For example, the carriage can be divided into three detection areas: the front, the middle, and the rear.

[0084] In this embodiment, the main-auxiliary sensor pair refers to the main sensor and auxiliary sensor configured in the detection area. The main sensor is usually responsible for the main detection task, while the auxiliary sensor provides supplementary information. For example, the main sensor is a temperature sensor and the auxiliary sensor is a humidity sensor.

[0085] In this embodiment, the spatial dimension matrix refers to a matrix that represents the spatial distribution of sensors, helping to determine the location and coverage of the sensors. For example, a 3x3 matrix represents the location of 9 sensors inside the carriage.

[0086] In this embodiment, the sensor capability complementarity table refers to a table that displays the complementary relationships between different sensor capabilities, which is used to optimize sensor configuration. For example, the table displays the detection range and accuracy of temperature and humidity sensors.

[0087] In this embodiment, the functional dimension matrix refers to a matrix that represents the functional characteristics of the sensors. It is combined with the spatial dimension matrix to evaluate the functionality of the entire sensor network. For example, a matrix can display the environmental parameters that each sensor can detect.

[0088] In this embodiment, a multi-source sensor array refers to an array composed of multiple different types of sensors that can work together to provide comprehensive monitoring data, such as an array composed of temperature, humidity, and carbon dioxide sensors.

[0089] In this embodiment, the basic performance verification result verifies the performance of the sensor under normal conditions, such as the accuracy of the temperature sensor at 25°C; the extreme scenario verification result verifies the performance of the sensor under extreme conditions, such as the accuracy of the temperature sensor at -10°C or 50°C; and the multimodal fusion verification result verifies the performance of the fused data from multiple sensors, such as combining temperature and humidity data to predict the comfort level inside the vehicle.

[0090] In this embodiment, the first arrangement is a sensor arrangement based on the basic performance verification results; the second arrangement is a sensor arrangement based on the extreme scenario verification results; and the third arrangement is a sensor arrangement based on the multimodal fusion verification results.

[0091] In this embodiment, Pareto optimality refers to the configuration where the performance of a particular sensor cannot be further improved without compromising the performance of other sensors. This is an optimization method used to find the optimal sensor arrangement, for example, minimizing the number of sensors while ensuring sufficient sensor coverage across all detection areas.

[0092] In this embodiment, the actual arrangement refers to the result of arranging the sensors according to the optimal sensor arrangement scheme calculated based on Pareto optimality. For example, 20 sensors, including temperature, humidity, and carbon dioxide sensors, are installed in the carriage to achieve the best monitoring effect. ,in, This indicates the final selected optimal sensor arrangement. This represents the weight of the i-th objective; This represents the value of the i-th objective function under sensor placement scheme x. This represents the maximum value of target i in the frontier; Let represent the minimum value of objective i in the frontier; γ represent the penalty coefficient of the implementation function; C(x) represent the total implementation function of sensor placement scheme x; x represents the candidate sensor placement scheme; P represents the Pareto front solution set; and i represents the index of the objective function.

[0093] The working principle and beneficial effects of the above technical solution are as follows: by analyzing the bus structure diagram to divide the detection area, designing a main-auxiliary sensor collaborative network, verifying basic performance, extreme scenarios and multimodal fusion effects in a digital twin, and finally, based on the Pareto optimal algorithm to integrate the three verification results, outputting the globally optimal sensor layout scheme, achieving accurate sensor coverage and improving the adaptability of bus information collection. Example 4:

[0094] This invention provides a deep learning-based method for bus passenger flow statistics, which classifies multi-source data to determine vehicle data and passenger data, detects passenger boarding and alighting behavior based on the vehicle and passenger data, and obtains behavior detection results, including:

[0095] Preprocess the multi-source data, determine the data type of the preprocessing results, and determine the corresponding feature extraction method and output dimension from the type-feature engineering table based on the data type;

[0096] Based on the aforementioned feature extraction method and output dimensions, vehicle data and passenger data are determined, thereby determining the vehicle data classification model, the passenger data classification model, and multimodal collaboration.

[0097] Based on the vehicle data classification model, passenger data classification model, and multimodal collaboration, passenger boarding and alighting behaviors are detected, and then the behavior detection results are obtained.

[0098] In this embodiment, the preprocessing result refers to the data after the original data has been cleaned, formatted, and standardized, such as unifying the timestamp format, filling in missing values, and removing outliers.

[0099] In this embodiment, the type-feature engineering table is a mapping table that maps different data types to corresponding feature extraction methods and output dimensions. It is constructed based on domain knowledge and data characteristics. For example, a table that lists data types (such as text, numerical values, time series, etc.), corresponding feature extraction methods (such as bag-of-words model, normalization, Fourier transform, etc.), and output dimensions (such as 100-dimensional vector, 20-dimensional vector, etc.).

[0100] In this embodiment, data type refers to the type of data, such as text, numerical values, images, time series, etc. Determining the data type is a prerequisite for selecting an appropriate feature extraction method; for example, numerical data and text data.

[0101] In this embodiment, the feature extraction method refers to the process of converting raw data into a format acceptable to the model. This involves extracting key information from the data and representing it in the form of vectors. For example, for text data, the bag-of-words model can be used to extract features; for numerical data, normalization or standardization can be used.

[0102] In this embodiment, the output dimension refers to the dimension of the feature vector obtained after feature extraction. This dimension determines the size of the model input, for example, a 100-dimensional vector.

[0103] In this embodiment, the input and output of the vehicle data classification model refer to the format of the vehicle data used by the model during training and prediction, and the results output by the model. For example, the input is the vehicle's acceleration and speed, and the output is whether the vehicle is braking.

[0104] In this embodiment, the input and output of the passenger data classification model refer to the format of the passenger data used by the model during training and prediction, and the results output by the model. For example, the input is an image of a passenger, and the output is whether the passenger is wearing a mask.

[0105] In this embodiment, multimodal collaborative classification refers to combining data from different modalities (such as text, images, and sound) for classification. This method can improve the accuracy and robustness of classification; for example, it can combine vehicle GPS location data and passenger image data to detect passenger boarding and alighting behavior.

[0106] In this embodiment, key features are extracted by vehicle and passenger classification models respectively, and multimodal collaborative algorithms (such as attention mechanisms) are used to dynamically fuse vehicle status and passenger behavior features. Combined with spatiotemporal correlation analysis, high-precision detection of getting on and off the vehicle is achieved, forming a closed-loop analysis chain of "data perception - feature fusion - behavior determination".

[0107] The working principle and beneficial effects of the above technical solution are as follows: By processing multi-source heterogeneous data through intelligent classification and feature engineering, classification models for vehicle and passenger data are constructed respectively. By combining multimodal collaborative algorithms to fuse spatiotemporal features, high-precision detection of getting on and off vehicles is achieved. The system automatically matches data types and feature extraction methods and dynamically optimizes the detection process. Example 5:

[0108] This invention provides a deep learning-based method for bus passenger flow statistics. The method detects passenger boarding and alighting behaviors based on a vehicle data classification model, a passenger data classification model, and multimodal collaboration, thereby obtaining behavior detection results. The method includes:

[0109] Based on the vehicle data classification model, vehicle features are distinguished; based on the passenger data analysis model, passenger features are identified; and the multimodal collaborative approach is used to formulate a feature fusion strategy corresponding to vehicle features and passenger features.

[0110] The vehicle features and passenger features are fused according to the feature fusion strategy to obtain the fusion result. The fusion result is then analyzed to detect passenger boarding and alighting behavior, and the behavior detection result is obtained.

[0111] In this embodiment, vehicle features refer to attributes extracted from vehicle data used to describe the vehicle's state or behavior. These features can be the vehicle's physical characteristics, operating status, or other relevant information, such as the vehicle's current speed, acceleration, braking status, and steering angle.

[0112] In this embodiment, passenger characteristics refer to information extracted from passenger data used to describe passenger attributes. These characteristics may be the passenger's physiological characteristics, behavioral habits, or other relevant information, such as the passenger's age, gender, height, whether they are carrying luggage, etc.

[0113] In this embodiment, the feature fusion strategy refers to a method for merging features from different data sources into a comprehensive feature representation. This typically involves selecting which features, handling the correlation between different features, and fusing these features. For example, combining vehicle acceleration and velocity features with passenger body posture and position features to predict whether passengers are ready to get on or off the vehicle.

[0114] In this embodiment, the fusion result refers to the result obtained by merging vehicle features and passenger features according to a feature fusion strategy. This result is typically a dataset containing more information, which can be used for more complex analysis and modeling, such as a comprehensive feature vector containing vehicle speed, acceleration, passenger posture, and position.

[0115] The working principle and beneficial effects of the above technical solution are as follows: key features are extracted by vehicle and passenger classification models respectively, and multimodal collaborative algorithms (such as attention mechanisms) are used to dynamically fuse vehicle status and passenger behavior features. Combined with spatiotemporal correlation analysis, high-precision detection of getting on and off the vehicle behavior is achieved, forming a closed-loop analysis chain of "data perception - feature fusion - behavior determination" to achieve accurate detection and dynamic adaptation. Example 6:

[0116] This invention provides a deep learning-based method for bus passenger flow statistics, which involves performing magnitude analysis on behavior detection results to generate a structured magnitude report, associating the structured magnitude report with facial recognition model parameters to obtain association results, and configuring a facial recognition model based on the association results. The method includes:

[0117] The behavior detection results are aggregated in the spatiotemporal dimensions, and the distribution of passenger boarding and alighting, behavior types and regional density within a unit time window are statistically analyzed. The magnitude is divided into levels based on preset thresholds, and a structured magnitude report is output.

[0118] Based on the structured volumetric report, volumetric features are extracted, and these features are associated with the key parameters of the face recognition model to obtain the association results. Then, a hierarchical configuration is performed based on the association results, and the face recognition model is obtained by combining all the hierarchical configurations.

[0119] In this embodiment, spatiotemporal aggregation involves dividing and statistically analyzing behavioral data by both time segments and spatial regions. For example, a subway station divides its monitoring area into three spatial units: "turnstiles," "platforms," ​​and "passages." Using a 15-minute time window, data from each area is statistically analyzed during the period from 8:00 to 8:15.

[0120] In this embodiment, the number of passengers getting on and off the bus is the number of passengers entering and exiting a specific area during a fixed period. For example, at the east gate of the bus station from 7:30 to 8:00: there are 85 card swipe records when boarding and 72 QR code scan records when getting off.

[0121] In this embodiment, the behavior type distribution is the proportion of each type of behavior in the total detection volume. For example, the monitoring and analysis of the waiting area shows that: standing while waiting for the bus is 68%, walking is 25%, luggage is lost is 3%, and physical conflict is 4%.

[0122] In this embodiment, the area density is the real-time number of people per unit area. For example, the infrared sensor in the middle of the carriage detects 2.3 people / m², which exceeds the warning threshold of 2.0 people / m².

[0123] In this embodiment, the preset threshold and level are pre-set grading thresholds. For example, density grading: below 1.5 people / ㎡ = green, 1.5-2.5 people / ㎡ = yellow, above 2.5 people / ㎡ = red; abnormal behavior grading: 3 falls / hour = Level 1 warning, 5 falls / hour = Level 2 emergency.

[0124] In this embodiment, the structured volumetric report is a statistical analysis summary in a standardized format, such as "Report 2023-12-01 08:15:00: Area: Central part of platform 2, Density level: Red (2.8 people / ㎡), Behavioral alarm: Fall behavior triggered (3 times within 5 minutes), Recommended measures: Activate the flow restriction plan".

[0125] In this embodiment, the magnitude feature extraction is to extract key decision indicators from the report. For example, the core feature is extracted as "continuous high density + frequent abnormal behavior", and the related factors are "morning rush hour + rain and snow weather".

[0126] In this embodiment, the key parameters of face recognition are the core adjustable parameters that affect the recognition effect. For example, image acquisition: resolution (1080P / 720P), analysis speed: detection frame rate (10fps / 25fps), comparison accuracy: similarity threshold (0.8 / 0.9).

[0127] In this embodiment, the hierarchical configuration is a multi-level parameter combination scheme. For example, the basic mode is 1080P@15fps for sunny and normal days, the enhanced mode is 720P@20fps for rainy and snowy weather, and the emergency mode is 480P@30fps for large crowds and events.

[0128] The working principle and beneficial effects of the above technical solution are as follows: by analyzing passenger flow behavior data through spatiotemporal aggregation, a structured volumetric report is generated, and the face recognition model parameters (such as detection frame rate and ROI region) are dynamically adjusted based on passenger flow density to achieve closed-loop optimization of "passenger flow perception-model adaptation" and ensure optimal recognition performance in different scenarios. Example 7:

[0129] This invention provides a deep learning-based method for bus passenger flow statistics. It integrates historical passenger flow data, vehicle BeiDou trajectories, and weather data using a facial recognition model to form a predictive model, derives a predicted change in passenger numbers, and pushes this predicted change in passenger numbers to the dispatch center to optimize vehicle allocation. The method includes:

[0130] Real-time passenger flow is obtained from the facial recognition system, and historical passenger flow is extracted from the historical passenger flow database. Vehicle Beidou trajectories are collected simultaneously, and weather data is obtained through the meteorological interface.

[0131] The time dimension is decomposed into periodic time features, while the topological features of the site are obtained.

[0132] Based on periodic time characteristics, ordinary time-series feature vectors and key time-series feature vectors are extracted from real-time passenger flow and historical passenger flow. Static features are extracted from weather data. The ordinary time-series feature vectors and the static features are first concatenated, and the key time-series feature vectors and the static features are second concatenated.

[0133] Based on the topological features, the first and second splices are connected to form a prediction model, which yields the predicted number of people to be affected. The predicted number of people to be affected is then pushed to the dispatch center to optimize vehicle allocation.

[0134] In this embodiment, the real-time passenger flow is the current passenger flow counted in real time by the facial recognition system. For example, the subway station gate camera detected 200 people entering the station and 150 people exiting the station in the current 5 minutes.

[0135] In this embodiment, the historical passenger flow database is a structured database that stores historical passenger flow data. For example, a station passenger flow table stored hourly in a MySQL database contains fields such as "date, time period, and number of people entering and exiting".

[0136] In this embodiment, the periodic time feature is a periodic pattern extracted from time (such as hour / week / season), for example, the morning rush hour (7:00-9:00), weekends, and holiday markers.

[0137] In this embodiment, topological features are features that describe the spatial relationships between stations (such as adjacent stations, hub level). For example, station B is a transfer station for 3 subway lines, and its adjacent stations are C and D.

[0138] In this embodiment, the ordinary time series feature vector is a regular time series statistical value (such as the mean, the number of people in the sliding window), for example, the average passenger flow in the past hour is 500 people, which is a 10% increase compared to the previous period.

[0139] In this embodiment, the key time-series feature vector is the key time-series indicator (such as peak value, abrupt change point) that affects passenger flow. For example, the passenger flow suddenly increased by 50% in the same period yesterday (due to a sudden event).

[0140] In this embodiment, static features are weather / environmental attributes that remain unchanged in the short term, such as the current weather label "heavy rain" and the temperature "30°C".

[0141] In this embodiment, the first concatenation and the second concatenation are combinations of different feature vectors by dimension. For example, the first concatenation is: ordinary time-series feature (mean 500 people) + static feature (rainstorm) → [500, rainstorm], and the second concatenation is: key time-series feature (mutation + 50%) + static feature (rainstorm) → [+50%, rainstorm].

[0142] In this embodiment, the predicted number of people is the amount of change in passenger flow predicted by the model (absolute number or percentage). For example, it is expected that the passenger flow from 15:00 to 16:00 will increase by 200 people compared to a normal day.

[0143] The working principle and beneficial effects of the above technical solution are as follows: By integrating real-time facial recognition passenger flow, historical passenger flow patterns, vehicle BeiDou trajectories, and weather data, a hybrid prediction model with enhanced spatiotemporal features is constructed. The model combines periodic passenger flow trends with station topology relationships to dynamically generate passenger flow prediction results, driving intelligent scheduling decisions and achieving data-driven precise capacity allocation. Example 8:

[0144] This invention provides a deep learning-based method for bus passenger flow statistics, which pushes predicted passenger volume changes to the dispatch center to optimize vehicle allocation, including:

[0145] Match key causal chains based on predicted changes in the number of people, and trigger resource allocation schemes based on the key causal chains.

[0146] Based on the resource allocation plan, a sand table simulation is performed, and the plan evaluation results are output. The dispatch center determines the dispatch instructions and optimizes vehicle allocation based on the plan evaluation results.

[0147] In this embodiment, the key causal chain is the core factors affecting passenger flow changes and their logical relationship chains. For example, heavy rain (cause) → increased taxi demand (effect) → a 20% decrease in the number of people entering the subway station (final effect), the end of a concert (cause) → 500 people instantaneously gather at the bus stop (effect) → 3 emergency vehicles need to be added (final effect).

[0148] In this embodiment, the resource allocation plan is a specific resource adjustment plan formulated for predicted changes. For example, when it is predicted that the passenger flow will increase by 200 people: add 2 spare buses. Extend the last bus by 30 minutes. Open the emergency channel for diversion.

[0149] In this embodiment, the sand table deduction is to simulate the implementation effect of the plan in a virtual environment. For example, input conditions: it is predicted that the passenger flow during the evening peak will increase by 300 people, and there are 5 available spare vehicles. Deduction process: after simulating the dispatch of 3 vehicles: the waiting time is reduced from 15 minutes to 8 minutes, and the platform congestion degree is reduced from 90% to 60%.

[0150] In this embodiment, the evaluation result of the plan is the quantitative score of each plan after deduction. For example, Plan A (dispatch 3 vehicles): cost: 5000 yuan, efficiency improvement: the waiting time is reduced by 47%, comprehensive score: 85 points, Plan B (dispatch 2 vehicles + diversion): cost: 3000 yuan, efficiency improvement: the waiting time is reduced by 35%, comprehensive score: 78 points.

[0151] In this embodiment, the dispatching center is the central control unit for decision-making on resource allocation. For example, the large screen of a certain city's traffic command center shows: real-time monitoring of the passenger flow at 12 hub stations, receiving the top 3 allocation plans recommended by the system, and finally manually confirming to execute Plan B.

[0152] In this embodiment, the mobilization instruction is an operable command issued to the executing unit. For example, the instruction content: "Dispatch vehicle Beijing B12345 to Guomao Station before 18:00", "Activate the No. 2 emergency channel at Xidan Station", execution feedback: the vehicle has arrived and is confirmed by GPS, and the channel monitoring shows an open state.

[0153] In this embodiment, vehicle allocation is the adjustment of transport capacity resources implemented according to the instruction. For example, in the regular scenario: the departure frequency is fixed to increase by 10% during the morning and evening peaks, and in the emergency scenario: 20 reserve buses are urgently called from the garage during emergencies.

[0154] The working principle and beneficial effects of the above technical solution are: by analyzing passenger flow prediction data and the key causal chain (such as weather - passenger flow - resource demand), dynamically generate a resource allocation plan, and after verifying the effect through digital twin sand table deduction, output the optimal dispatching instruction to achieve closed-loop optimization from prediction to execution.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for passenger flow statistics on buses based on deep learning, characterized in that, include: Step 1: Obtain a bus structure diagram, deploy a detection device based on the bus structure diagram, and collect multi-source data based on the detection device; Step 2: Classify the multi-source data to determine vehicle data and passenger data. Based on the vehicle data and passenger data, detect passenger boarding and alighting behavior to obtain the behavior detection results. Step 3: Perform magnitude analysis on the behavior detection results to obtain a structured magnitude report, correlate the face recognition model parameters on the structured magnitude report to obtain the correlation results, and configure the face recognition model based on the correlation results; Step 4: Based on the facial recognition model, historical passenger flow, vehicle Beidou trajectory and weather data are integrated to form a prediction model, obtain the predicted change in the number of passengers, and push the predicted change in the number of passengers to the dispatch center to optimize vehicle allocation; Step 1 includes: Based on the bus structure diagram, a multi-source sensor array is determined, and the multi-source sensor array is used to perform digital twin verification on the bus. The actual layout of the multi-source sensor array is then determined based on the verification results. Based on the actual deployment of the corresponding detection devices, multi-source data acquisition is performed. The process of determining a multi-source sensor array based on the bus construction diagram, using the multi-source sensor array to perform digital twin verification of the bus, and determining the actual arrangement of the multi-source sensor array based on the verification results includes: The bus structural diagram is analyzed to extract key dimension parameters, and the existing equipment inside the bus is marked with its location. Based on the key dimension parameters and location marks, available routes are determined. Based on the available lines, the carriage is divided into multiple detection areas. Each detection area is configured with a main-auxiliary sensor pair to form a spatial dimension matrix. A sensor capability complementarity table is established based on the main-auxiliary sensor pair to form a functional dimension matrix. The spatial dimension matrix and the functional dimension matrix are combined to obtain a multi-source sensor array. The bus is digitally twinned based on the multi-source sensor array, generating basic performance verification results, extreme scenario verification results, and multimodal fusion verification results. The first layout is derived based on basic performance verification, the second layout is derived based on extreme scenario verification results, and the third layout is derived based on multimodal fusion verification results. By combining the first, second, and third arrangements, Pareto optimality calculations are performed to obtain the actual arrangement of the multi-source sensor array.

2. The method for bus passenger flow statistics based on deep learning according to claim 1, characterized in that, The multi-source data is classified to identify vehicle data and passenger data. Passenger boarding and alighting behavior is detected based on the vehicle and passenger data to obtain behavior detection results, including: Preprocess the multi-source data, determine the data type of the preprocessing results, and determine the corresponding feature extraction method and output dimension from the type-feature engineering table based on the data type; Based on the aforementioned feature extraction method and output dimensions, vehicle data and passenger data are determined, thereby determining the vehicle data classification model, the passenger data classification model, and multimodal collaboration. Based on the vehicle data classification model, passenger data classification model, and multimodal collaboration, passenger boarding and alighting behaviors are detected, and then the behavior detection results are obtained.

3. The method for calculating bus passenger flow based on deep learning according to claim 2, characterized in that, Based on the vehicle data classification model, passenger data classification model, and multimodal collaboration, passenger boarding and alighting behaviors are detected, and the behavior detection results are obtained, including: Based on the vehicle data classification model, vehicle features are distinguished; based on the passenger data analysis model, passenger features are identified; and the multimodal collaborative approach is used to formulate a feature fusion strategy corresponding to vehicle features and passenger features. The vehicle features and passenger features are fused according to the feature fusion strategy to obtain the fusion result. The fusion result is then analyzed to detect passenger boarding and alighting behavior, and the behavior detection result is obtained.

4. The method for bus passenger flow statistics based on deep learning according to claim 1, characterized in that, Perform magnitude analysis on the behavior detection results to obtain a structured magnitude report. Correlate the structured magnitude report with face recognition model parameters to obtain the correlation results. Configure the face recognition model based on the correlation results, including: The behavior detection results are aggregated in the spatiotemporal dimensions, and the distribution of passenger boarding and alighting, behavior types and regional density within a unit time window are statistically analyzed. The magnitude is divided into levels based on preset thresholds, and a structured magnitude report is output. Based on the structured volumetric report, volumetric features are extracted, and these features are associated with the key parameters of the face recognition model to obtain the association results. Then, a hierarchical configuration is performed based on the association results, and the face recognition model is obtained by combining all the hierarchical configurations.

5. The method for bus passenger flow statistics based on deep learning according to claim 1, characterized in that, A predictive model is formed by fusing historical passenger flow, vehicle BeiDou trajectory, and weather data using a facial recognition model. This model predicts changes in passenger numbers and pushes these predictions to the dispatch center to optimize vehicle allocation. This includes: Real-time passenger flow is obtained from the facial recognition system, and historical passenger flow is extracted from the historical passenger flow database. Vehicle Beidou trajectories are collected simultaneously, and weather data is obtained through the meteorological interface. The time dimension is decomposed into periodic time features, while the topological features of the site are obtained. Based on periodic time characteristics, ordinary time-series feature vectors and key time-series feature vectors are extracted from real-time passenger flow and historical passenger flow. Static features are extracted from weather data. The ordinary time-series feature vectors and the static features are first concatenated, and the key time-series feature vectors and the static features are second concatenated. Based on the topological features, the first and second splices are connected to form a prediction model, which yields the predicted number of people to be affected. The predicted number of people to be affected is then pushed to the dispatch center to optimize vehicle allocation.

6. The method for bus passenger flow statistics based on deep learning according to claim 5, characterized in that, The predicted changes in passenger numbers are pushed to the dispatch center to optimize vehicle allocation, including: Match key causal chains based on predicted changes in the number of people, and trigger resource allocation schemes based on the key causal chains. Based on the resource allocation plan, a sand table simulation is performed, and the plan evaluation results are output. The dispatch center determines the dispatch instructions and optimizes vehicle allocation based on the plan evaluation results.

Images