A bus passenger flow identification and statistical optimization method based on deep learning

Abstract

This invention relates to the field of computer technology, and in particular to a method for identifying and optimizing public transport passenger flow based on deep learning. This invention constructs a waiting feature vector by identifying and extracting features from waiting situation data at the platform. This waiting feature vector is then used as the input and output of a destination station probability distribution generation model. Based on probability ranking, a set of highlighted stations is determined, and route accessibility information is generated. The accessible highlighted stations are prominently displayed on the platform display terminal, and the corresponding available routes, remaining stops, and estimated arrival time are simultaneously provided. This allows passengers to select their destination station and route without having to compare multiple route and station lists line by line, thereby reducing the frequency of repeated confirmations and on-site inquiries, improving platform guidance efficiency, and alleviating the consultation pressure on platform staff and drivers.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method for identifying and optimizing public transport passenger flow based on deep learning. Background Technology

[0002] Existing public transport information services typically involve configuring platform display terminals at bus stops. These terminals display the route number, estimated arrival time of the vehicle, and a list of stops the route will reach. In some scenarios, they also provide arrival reminders in conjunction with platform voice announcements or mobile app push notifications. On the side of the bus, there are in-vehicle broadcasting equipment and driver's seat reminder terminals that announce the stop name in a preset sequence when the vehicle approaches or arrives at the stop, and provide the driver with arrival reminder information when necessary.

[0003] The aforementioned existing technologies mostly employ fixed display and fixed broadcast methods. Although the platform display terminal can list route and station information, it lacks a mechanism to highlight the destination station that passengers are most concerned about based on the characteristics of the waiting crowd at the target bus stop. Passengers still need to compare multiple routes line by line and repeatedly confirm whether they can reach the target station, resulting in frequent inquiries and low efficiency. Onboard arrival prompts are usually broadcast with a fixed number of times and a fixed intensity, making it difficult to adjust the broadcast coverage according to the number of passengers boarding, passenger flow information inside the vehicle, and changes in the noise level of the carriage. In noisy or distracted scenarios, passengers may not receive the prompts and may inquire again. This situation increases the burden on drivers to manually announce passenger flow and increases the risk of driver distraction. At the same time, existing passenger flow statistics are mostly based on headcount or crowding level judgment, and are mostly used for post-event summarization or operational evaluation. They are usually processed in different links from platform information display and vehicle arrival announcements, lacking unified data association and consistency verification methods. It is difficult to reflect changes in passenger flow structure and travel patterns in a timely manner in information prompting strategies, which can easily lead to discrepancies between information prompts and actual travel needs. When external conditions change, such as holidays, school commuting, morning and evening rush hours, and detours, information services are prone to deviating from actual travel needs.

[0004] To address this, a deep learning-based method for bus passenger flow identification and statistical optimization is proposed. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a method for bus passenger flow identification and statistical optimization based on deep learning.

[0006] To solve the above technical problems, the present invention provides the following technical solution: a method for bus passenger flow identification and statistical optimization based on deep learning, comprising the following steps: S1, acquiring waiting status data of the target bus station within a preset time window, and performing waiting passenger flow identification and feature extraction to obtain waiting feature information; S2, performing data preprocessing on the waiting feature information to obtain waiting feature vectors; S3, generating a set of candidate destination stations based on the set of available routes of the target bus station; S4, inputting the waiting feature vectors into a destination distribution generation model, and outputting a probability distribution of destination stations corresponding to the set of candidate destination stations; S5, determining a set of highlighted stations based on the probability distribution of destination stations, and generating scene label information based on the set of highlighted stations; S6, generating prominent guidance information based on the set of highlighted stations and outputting it to the station display terminal, the prominent guidance information including matching... The platform display terminal highlights the matching station set, display attribute parameters, and route accessibility explanation information based on the prominent guidance information and simultaneously outputs the route accessibility explanation information; S7, obtain the passenger flow status information and number of passengers boarding on the vehicle side, generate alighting prediction information based on the number of passengers boarding, the probability distribution of destination stations, and the passenger flow status information on the vehicle side, and generate arrival prompt control information based on the alighting prediction information, the highlighted station set, and the scene label information to execute the arrival prompt; S8, obtain the actual passenger flow statistics results aligned with the preset time window, update the destination distribution generation model parameters and arrival prompt control information parameters based on the actual passenger flow statistics results, and send the updated parameters to the platform side and the vehicle side via the wireless communication network to update the destination station probability distribution, the highlighted station set, the prominent guidance information, and the arrival prompt control information.

[0007] As a preferred technical solution of the present invention, the waiting situation data includes any one or a combination of image data collected by the platform camera device, target count statistics output by the platform millimeter-wave radar, and platform operation configuration data. The platform operation configuration data includes a set of available routes, route arrival information, departure intervals, temporary suspension information, and detour information. The waiting passenger flow identification includes performing pedestrian detection on the platform waiting area and combining it with multi-target tracking to obtain the number of people waiting, entering, and leaving, and obtaining the proportion of people in each sub-region and the density near the gate based on the sub-region division of the waiting area. When the occlusion intensity of the image data exceeds a preset threshold, the target count statistics output by the millimeter-wave radar are used to correct the number of people waiting on the image side, and the occlusion intensity is used as part of the waiting feature information and input into the target distribution generation model.

[0008] As a preferred technical solution of the present invention, the data preprocessing includes field consistency processing, abnormal data processing, numerical scale unification processing and category feature encoding processing, and sets a data source availability flag to generate a waiting feature vector when image data or radar data is missing.

[0009] As a preferred technical solution of the present invention, the candidate destination station set is obtained by merging and deduplicating the subsequent station sequences corresponding to the available route set, and the candidate destination station set is eliminated or replaced according to temporary suspension information and detour information; the highlighted station set consists of a preset number of destination stations whose probability ranking is in the probability distribution of destination stations, and the probability distribution of destination stations is smoothed or an adjacent time window maintenance strategy is adopted to reduce the jump of the highlighted station set in continuous time windows; the route accessibility explanation information includes at least the accessible highlighted station identifier for each available route and the corresponding number of remaining stations and estimated arrival time, and the platform display terminal synchronously displays the route accessibility explanation information to support passengers to quickly select routes.

[0010] As a preferred embodiment of the present invention, the arrival prompt control information includes a broadcast station set, broadcast style parameters, broadcast intensity parameters, and driving prompt parameters. The broadcast style parameters are selected based on scene label information, the broadcast intensity parameters are selected or adjusted from a preset level table based on the carriage noise level and passenger congestion level, and the driving prompt parameters are triggered based on the door zone density and the expected number of alighting passengers. The actual passenger flow statistics are obtained by summarizing any one or a combination of the on-board side statistics and the transaction side statistics, and are archived in association with the station identifier and time window identifier. Based on the actual passenger flow statistics, a true destination distribution is constructed to update the destination distribution generation model parameters, and the arrival prompt control information parameters are updated based on the difference between the actual passenger flow statistics and the alighting prediction information.

[0011] Compared with the prior art, the beneficial effects that this invention can achieve are: 1. This invention constructs a waiting feature vector by identifying and extracting features from waiting passenger flow data on the platform side. The waiting feature vector is then used as the input and output of the destination station probability distribution in the destination distribution generation model. Based on the probability ranking, a set of highlighted stations is determined and route accessibility explanation information is generated. The accessible highlighted stations are prominently displayed on the platform display terminal, and the corresponding available routes, remaining stops, and estimated arrival time are simultaneously provided. This allows passengers to complete the selection of destination station and route without having to compare multiple route station lists line by line, thereby reducing the frequency of repeated confirmations and on-site inquiries, improving platform guidance efficiency, and alleviating the consultation pressure on platform staff and drivers.

[0012] 2. This invention maps the probability distribution of destination stations output from the platform side and the set of highlighted stations to the future station sequence of the vehicle's travel route. It then combines the number of passengers boarding and the passenger flow status information on the vehicle side to generate the alighting prediction information for the next arrival station. Furthermore, it generates arrival prompt control information from the alighting prediction information, scene label information, and the set of highlighted stations, and adaptively selects or adjusts the broadcast style parameters, broadcast intensity parameters, and driving prompt parameters. This allows the voice broadcast and in-vehicle prompts to match peak alighting times, crowded door areas, and noisy environments, thereby improving the perceptibility and effectiveness of arrival prompts, reducing missed or misheard information during peak hours, and improving the efficiency and safety of passenger boarding and alighting organization.

[0013] 3. This invention aligns and archives platform-side waiting status data, vehicle-side passenger flow status information, and actual passenger flow statistics according to station identifiers and time window identifiers. Based on the actual passenger flow statistics, it constructs a true destination distribution and generates a loss function from the predicted distribution output by the destination distribution model to update the model parameters. Simultaneously, it updates the arrival prompt control parameters based on the difference between the actual passenger flow statistics and the alighting prediction information, and distributes the updated parameters to both the platform and vehicle sides. This allows the model and prompt strategy to be continuously corrected and iterated as the route is adjusted, the time period changes, and the passenger flow structure changes, thereby improving the long-term stability and generalization ability of destination station prediction and reducing guidance inaccuracies and prompt mismatches caused by data offset. Attached Figure Description

[0014] Fig. 1 This is a schematic diagram of the architecture of the present invention. Fig. 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0015] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0016] Example: Figs. 1-2As shown, a deep learning-based method for bus passenger flow identification and statistical optimization is collaboratively implemented by the platform side, the vehicle side, and the cloud side. The platform side includes a platform display terminal, a platform camera device, and a platform millimeter-wave radar; the vehicle side includes a vehicle camera device, a vehicle sensor device, and a vehicle communication terminal; the cloud side includes a model training server and a parameter distribution server. The platform side and the vehicle side interact with the cloud side via a wireless communication network, and the interaction messages all carry station identifiers, route identifiers, vehicle identifiers, time window identifiers, and version numbers. To ensure data availability and controllable communication overhead, the platform side prioritizes uploading structured features and statistical results, while image data is only sampled and uploaded according to trigger conditions during model regression verification or fault diagnosis.

[0017] The time window serves as the processing unit for the entire process. The waiting status data on the platform side, the passenger flow status information on the vehicle side, and the actual passenger flow statistics are aligned with the station identifier and the time window identifier. The preset time window length is set to 180 seconds, and the sliding update interval is set to 60 seconds.

[0018] Specifically, the steps include the following, and the order of these steps can be adjusted or executed in parallel without violating the business logic: S1, in the Within a specific time window, data on the waiting situation at the target bus stop is acquired, and passenger flow identification and feature extraction are performed to obtain waiting characteristic information. .

[0019] Waiting situation data includes any one or a combination of the following data sources: image data collected by platform camera devices, with an image frame rate of 10 to 30 frames per second and a resolution of 720 to 1080 lines; target statistics output by platform millimeter-wave radar, including target count, target distance distribution, target speed distribution, or sector occupancy ratio within a time window; and platform operation configuration data, including the set of available lines at the platform, line departure intervals, estimated arrival times, temporary suspension information, detour information, and platform geographical location.

[0020] Based on image data, passenger flow recognition for waiting areas is performed. This embodiment uses a combination of pedestrian detection and tracking: the waiting area is cropped for each frame of image, and the waiting area is pre-labeled by the platform; pedestrian detection is performed within the waiting area to obtain a set of detection boxes, and multi-target tracking is performed on the detection boxes of adjacent frames to obtain a set of waiting target trajectories; based on the trajectory set, the number of people waiting, entering, and leaving the window is counted; the pedestrian detection model can use a lightweight convolutional neural network or a transformer-structured detection network, and the inference version is run on the platform to meet real-time requirements.

[0021] When occlusion causes recognition instability, millimeter-wave radar is introduced for scale correction. The occlusion intensity can be calculated as follows: within a time window, the average overlap rate of the detection boxes on the image side is greater than a preset threshold, or the tracking loss rate is greater than a preset threshold, or the foreground occupancy ratio of the waiting area is greater than a preset threshold. If the occlusion intensity exceeds the threshold, the radar target count statistics are used as the correction basis to perform proportional correction or offset correction on the number of people waiting on the image side. This embodiment provides a reproducible correction method: Let the number of people estimated on the image side be... The number of targets on the radar side is The corrected number of people is ,Pick: .

[0022] in, The fusion coefficient is taken as 0.2 to 0.8, and in this embodiment it is taken as 0.5.

[0023] When only radar data is available, the number of people waiting is taken as follows: It outputs the sector occupancy ratio as a congestion level feature; when only image data exists, the number of people waiting is taken as... It outputs the occlusion intensity and image quality score.

[0024] Waiting area characteristics information This may include the following fields: Passenger flow related fields: Number of people waiting for the train Fields related to passenger flow include: number of passengers entering, number of passengers leaving, and rate of change in passenger flow; spatial distribution fields include: percentage of passengers in each sub-area after the waiting area is divided into multiple sub-areas, angle of the main direction of the queue, queue length, and density near the gate; time-related fields include: hour code, weekday identifier, holiday identifier, departure interval, and estimated arrival time; environmental and quality-related fields include: occlusion intensity, light intensity, rain and snow identifier, image quality score, and radar effective target rate; and route supply-related fields include: number of available routes and statistics on the estimated arrival time series of each route.

[0025] S2, Information on waiting characteristics Perform data preprocessing to obtain the waiting feature vector, and use the waiting feature vector as input to the destination distribution generation model.

[0026] Waiting area characteristic information Perform data preprocessing to obtain waiting area feature vectors. The preprocessing rules in this embodiment are as follows: Field consistency processing: The field names, units, and value ranges of different data sources are unified. For example, time is uniformly mapped to time window identifiers, and the number of people is uniformly counted as integers; Abnormal data processing: Data with negative numbers, change rates exceeding preset upper limits, and radar effective target rates below preset lower limits are marked as abnormal and replaced with the median of the nearest neighbor window or with a missing marker; Numerical scale unification processing: Continuous features are normalized or standardized. The normalization parameters are obtained and distributed by statistics during the cloud training phase; Category feature encoding processing: Category features such as hour, workday identifier, holiday identifier, and weather identifier are one-hot encoded or embedded encoded. The embedding vector dimension is 4 to 16, and this embodiment uses 8; Missing value strategy: When a data source is missing, the corresponding feature of the data source is marked as missing, and an availability flag is added to the vector to indicate whether the data source is available.

[0027] S3. Generate a set of candidate destination stations based on the set of available routes to the target bus stop. The platform side pre-stores or synchronizes a route and station information table from the cloud. This table records the sequence of subsequent stations and their identifiers after the target platform for each route. The subsequent station sequences corresponding to the available route sets are merged and deduplicated to obtain a candidate destination station set. If the operational configuration data includes temporary suspension information or detour information, the candidate destination site set will be removed or replaced to make the candidate set consistent with the operation of the day. To ensure real-time performance, the update cycle of the candidate destination site set can be 5 to 60 minutes. In this embodiment, it is 10 minutes and an immediate update is triggered when the operational configuration changes.

[0028] S4. Transfer the waiting feature vector Input the destination distribution generation model, output the set of candidate destination sites. One-to-one probability distribution of destination sites .

[0029] This embodiment illustrates a reproducible target distribution generation model structure and inference process; the target distribution generation model is a neural network structure, including a feature encoding subnetwork and a distribution output subnetwork: Feature encoding subnetwork: for the input waiting feature vector Multi-layer fully connected encoding is performed, with 2 to 5 layers; this embodiment uses 3 layers. The hidden layer dimensions are 256, 128, and 64 respectively. The activation function is ReLU, and a random dropout rate of 0.1 to 0.3 is added after the second layer; in this embodiment, 0.2 is used. When there are embedded encoding category features, the embedding vector is concatenated with continuous features and used as the encoding input. The output is a hidden representation vector. The dimension is 64; the distributed output subnetwork: the hidden representation vector Mapped to a score vector through a fully connected layer. The dimension of the score vector is equal to the number of candidate destination sites. The probability distribution is obtained by normalizing the score vector. Normalization uses softmax, and the calculation method is as follows: .

[0030] in, and Indexing candidate destination sites, For the first Within the first time window The scores of each candidate destination site For the first Within the first time window The probability of each candidate destination site.

[0031] To reduce frequent jumps in the highlighted site set caused by fluctuations in the probability distribution of the target site within adjacent time windows, this embodiment performs smoothing processing on the probability distribution of the target site before determining the highlighted site set; let the... The probability distribution of the destination site obtained by reasoning within a time window is as follows: Construct a smooth probability distribution : .

[0032] in, Indexing candidate destination sites, For the first The original probability of the i-th candidate destination site within each time window. The smoothed probability. This is a smoothing coefficient, with a value range of 0.6 to 0.95.

[0033] Based on the smoothed probability distribution The candidate destination sites are sorted, and the top-ranked destination sites by probability are selected to form a set of highlighted sites. Furthermore, when the candidate site set is updated in the current time window, causing a change in the site index, the overlapping sites from the previous window will be included. Inheritance is used as the initial value, and the initial value of non-overlapping stations is set to 0 to ensure the continuity of the smoothing process.

[0034] During platform-side inference, to reduce transmission and computational burden, the platform only needs to save the inference model and normalization parameters, and output the probability distribution or the first few station identifiers and corresponding probability values ​​in the probability ranking for each time window.

[0035] S5, based on highlighted site set The corresponding site attribute statistics generate scene tag information. The site attributes include category tags such as commercial area, residential area, hospital, school, and transportation hub. The tags are provided by the site knowledge table pre-set on the platform or in the cloud. The statistical method can be: the category that appears most frequently in the highlighted site cluster is used as the main scene tag, and the category weight vector is output as auxiliary scene information for subsequent broadcast style and display strategy selection.

[0036] S6, based on highlighted site set Candidate destination site set Scene tag information Generate prominent guidance information The information is then output to the platform display terminal, highlighting the guidance information. This includes the set of matching stations, display attribute parameters, and explanations of route reachability.

[0037] Matching station set generation: Match the highlighted station set with the pre-stored line station information on the station display terminal to obtain the matching station set; the matching granularity can be station identifier matching or normalized text matching of station names.

[0038] Display attribute parameter generation: Display attribute parameters include font size multiplier, color number, flashing frequency, emphasis icon identifier and sorting rules; In this embodiment, the font size multiplier is set to 1.2 to 1.8 for the site entries corresponding to the matching site set, and 1.5 is used in this embodiment, and an emphasis icon identifier is added before the site entries; The original display attributes of non-matching sites remain unchanged, so as to achieve highlight guidance without disturbing the original line information structure.

[0039] Route accessibility explanation information generation: To address the inefficiency of passenger line-by-line comparison, this embodiment generates route accessibility explanation information. Specifically, for each route in the current set of available routes at the platform, its subsequent station sequence is traversed to find stations that overlap with the set of highlighted stations. The set of highlighted station identifiers that the route can reach is output, and the remaining number of stations and estimated arrival time are calculated. The remaining number of stations is obtained by the difference between the index of the current station in the route sequence and the index of the highlighted stations. The estimated arrival time is calculated by the average inter-station travel time of the route or the estimated arrival time in the operation configuration. The route accessibility explanation information is output in structured fields, such as route identifier, list of accessible stations, list of corresponding remaining stations, and list of estimated arrival times.

[0040] After receiving the highlighted guidance information, the platform display terminal simultaneously displays the corresponding reachable highlighted stations and station time summaries on the display interface, allowing passengers to complete the route selection without having to compare each line.

[0041] When the platform display terminal supports interactive input, a destination station confirmation interaction can be added: after a passenger clicks on a highlighted station, the interface will only display the routes that can reach that station and the number of stops and time, thereby further reducing the frequency of inquiries.

[0042] S7. Obtain passenger flow status information and number of passengers boarding on the vehicle side; the passenger flow status information on the vehicle side is generated by the vehicle-mounted camera device and the vehicle-mounted sensor device, including any one or a combination of door area density, aisle occupancy ratio, number of passengers in each compartment, trend of gathering towards the door, and noise level in the compartment; the number of passengers boarding can be estimated from the door infrared counter, video counting, or card swipe records.

[0043] Based on passenger numbers and destination station probability distribution The system generates a disembarkation prediction information for the next destination station by combining the passenger flow status information on the vehicle side. This embodiment illustrates a prediction method: the probability distribution of the destination station is mapped into a disembarkation probability sequence of several future stations according to the route sequence, and the sequence is corrected by combining the current passenger volume and door area density to obtain the expected number of disembarkators and the disembarkation congestion risk level at the next station.

[0044] Based on drop-off prediction information and highlighted station sets Arrival prompt control information generated with scene label information The arrival notification control information includes the broadcast station set, broadcast style parameters, broadcast intensity parameters, and driving prompt parameters: the broadcast station set can be a subset of stations related to the current route from the highlighted station set, and can include next station prompts and transfer prompts; the broadcast style parameters are selected based on scene label information, for example, increasing the frequency of prompts in transportation hub scenarios, and reducing the broadcast intensity and using shorter prompts in hospital and school scenarios; the broadcast intensity parameters are selected or adjusted from a preset level table based on the noise level of the carriage, and the noise level can be calculated and graded from the microphone signal energy; the driving prompt parameters are used to remind the driver to pay attention to the risk of peak disembarkation or congestion at the next station, for example, prompting early arrival and optimizing door opening rhythm when the expected number of alighting passengers exceeds the threshold or the density of the door area exceeds the threshold.

[0045] The on-board unit executes arrival notification control information to generate voice broadcast control commands or in-vehicle display control commands.

[0046] When the vehicle-side lacks noise level or door zone density fields, the broadcast intensity parameters and driving prompt parameters are executed according to the default configuration, and the missing data is marked in the uploaded data so that robust strategies for missing conditions can be learned during the cloud training phase.

[0047] S8. Obtain the actual passenger flow statistics aligned with the preset time window. The actual passenger flow statistics are obtained by summing the statistics from the vehicle-side data and the transaction-side data, and are archived in association with the station identifier and time window identifier. The actual passenger flow statistics include the number of passengers boarding, the number of passengers alighting, and statistics of selectable destination stations. For the construction of the true distribution of destination stations, this embodiment covers the following situations: When there are drop-off station records: When the transaction side or the vehicle-side can provide passenger drop-off station identifiers, the number of passengers alighting at each destination station is counted according to the time window and normalized to obtain the true distribution of destination stations. In cases where drop-off station records are missing: When only boarding records exist but drop-off stations are missing, the drop-off station distribution can be constructed using the vehicle arrival sequence and the drop-off count from the door counter. This distribution can then be combined with prior knowledge of historical route destination distributions to assign individual destination stations, thus obtaining the true destination station distribution. The historical route destination distribution prior is obtained from cloud-based statistics by route, time period, and station dimensions and is updated periodically; in cases where sampling surveys or third-party statistics exist: when there are route destination station proportions obtained from sampling surveys or third-party statistical results, these are used as priors and integrated with the onboard passenger drop-off count to generate the data. .

[0048] Based on the actual distribution of the destination site Probability distribution of the destination site Construct a loss function and update the target distribution to generate model parameters; the loss function can be either cross-entropy loss or KL divergence loss, and this embodiment uses cross-entropy loss: .

[0049] in, Index candidate destination sites.

[0050] The parameter updates of the target distribution generation model can be performed using mini-batch gradient descent or adaptive moment estimation methods, with a learning rate of 0.0001 to 0.01. In this embodiment, a learning rate of 0.001 is used, and an upper limit for gradient clipping is set to ensure stable updates. To meet online stability requirements, the platform-side inference model is usually trained offline in the cloud and periodically distributed. Online updates can use a lightweight correction parameter update method, such as updating only the output layer bias term or temperature coefficient, and limiting the magnitude of a single update to no more than a preset upper limit.

[0051] Meanwhile, the arrival prompt control information parameters are updated based on the difference between the actual passenger flow statistics and the alighting prediction information. For example, the level threshold of the broadcast intensity parameter and the congestion risk trigger threshold are adaptively adjusted. The adjusted target distribution generation model parameters and control parameters are sent to the platform side and the vehicle side through the wireless communication network, carrying the version number and effective timestamp. The platform side and the vehicle side complete the hot update or switch during off-peak hours after verifying the version number.

[0052] To verify the technical effectiveness, this embodiment can be designed as follows for a reproducible experiment: Select several stations within the same city as the experimental group, and deploy the prominent guidance and route accessibility explanation display of this embodiment; select stations with similar passenger flow and similar number of routes as the control group, and use only fixed display and fixed announcement; statistical indicators include the average time for passengers to complete route confirmation, the number of on-site inquiries, the number of times station staff intervened, and the number of interactive clicks on the station interface; the experimental period can be 7 to 30 days, and statistics are collected separately for weekdays and holidays; the above experiment can verify that: after introducing the probability distribution of destination stations and the explanation of route accessibility, the confirmation time and inquiry frequency caused by passengers comparing line by line can be reduced, and the guidance effect can still be maintained even in the case of severe obstruction or missing data.

[0053] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for identifying and optimizing public transport passenger flow based on deep learning, characterized in that, Includes the following steps: S1. Obtain waiting status data of the target bus station within a preset time window, and perform waiting passenger flow identification and feature extraction to obtain waiting feature information; S2. Perform data preprocessing on the waiting area feature information to obtain the waiting area feature vector; S3. Generate a set of candidate destination stations based on the set of available routes to the target bus station; S4. Input the waiting feature vector into the destination distribution generation model, and output the destination station probability distribution corresponding to the candidate destination station set; S5. Determine the set of highlighted sites based on the probability distribution of the target sites, and generate scene label information based on the set of highlighted sites; S6. Generate prominent guidance information based on the highlighted station set and output it to the station display terminal. The prominent guidance information includes the matching station set, display attribute parameters and route accessibility explanation information. The station display terminal highlights the matching station set according to the prominent guidance information and outputs the route accessibility explanation information simultaneously. S7. Obtain passenger flow status information and number of passengers boarding on the vehicle side. Generate alighting prediction information based on the number of passengers boarding, the probability distribution of destination stations, and passenger flow status information on the vehicle side. Generate arrival prompt control information from the alighting prediction information, the set of highlighted stations, and the scene label information to execute the arrival prompt. S8. Obtain the actual passenger flow statistics results aligned with the preset time window, update the destination distribution generation model parameters and arrival prompt control information parameters based on the actual passenger flow statistics results, and send the updated parameters to the platform side and the vehicle side via the wireless communication network to update the destination station probability distribution, highlighted station set, highlighted guidance information and arrival prompt control information.

2. The method for bus passenger flow identification and statistical optimization based on deep learning according to claim 1, characterized in that, The waiting situation data includes any one or a combination of image data collected by platform camera devices, target count statistics output by platform millimeter-wave radar, and platform operation configuration data. The platform operation configuration data includes a set of available routes, estimated arrival information of routes, departure intervals, temporary suspension information, and detour information.

3. The method for bus passenger flow identification and statistical optimization based on deep learning according to claim 2, characterized in that, The passenger flow identification includes performing pedestrian detection on the platform waiting area and combining it with multi-target tracking to obtain the number of people waiting, entering and leaving, and obtaining the proportion of people in each sub-area and the density near the gate based on the sub-area division of the waiting area.

4. The method for bus passenger flow identification and statistical optimization based on deep learning according to claim 3, characterized in that, When the occlusion intensity of the image data exceeds a preset threshold, the target number statistics output by the millimeter-wave radar are used to correct the number of people waiting on the image side, and the occlusion intensity is used as part of the waiting feature information and input into the target distribution generation model.

5. The method for bus passenger flow identification and statistical optimization based on deep learning according to claim 4, characterized in that, The data preprocessing includes field consistency processing, abnormal data processing, numerical scale unification processing, and category feature encoding processing. When image data or radar data is missing, a data source availability flag is set to generate a waiting feature vector.

6. The method for bus passenger flow identification and statistical optimization based on deep learning according to claim 5, characterized in that, The candidate destination station set is obtained by merging and deduplicating the subsequent station sequences corresponding to the available route set, and the candidate destination station set is eliminated or replaced according to temporary suspension information and detour information.

7. The method for bus passenger flow identification and statistical optimization based on deep learning according to claim 6, characterized in that, The set of highlighted sites consists of a predetermined number of target sites whose probability ranking is among the top in the probability distribution of target sites. The probability distribution of target sites is smoothed or an adjacent time window preservation strategy is adopted to reduce the jumps of the set of highlighted sites within consecutive time windows.

8. The method for bus passenger flow identification and statistical optimization based on deep learning according to claim 7, characterized in that, The route accessibility information includes at least the highlighted station identifiers for each available route, the corresponding number of remaining stops, and the estimated arrival time. The platform display terminal synchronously displays the route accessibility information to support passengers in quickly selecting routes.

9. The method for bus passenger flow identification and statistical optimization based on deep learning according to claim 8, characterized in that, The arrival notification control information includes a broadcast station set, broadcast style parameters, broadcast intensity parameters, and driving prompt parameters. The broadcast style parameters are selected based on scene label information, the broadcast intensity parameters are selected or adjusted from a preset level table based on the carriage noise level and passenger congestion level, and the driving prompt parameters are triggered based on the door zone density and the expected number of passengers disembarking.

10. The method for bus passenger flow identification and statistical optimization based on deep learning according to claim 9, characterized in that, The actual passenger flow statistics are obtained by summarizing any one or a combination of the on-board side statistics and the transaction side statistics, and are archived in association with the station identifier and the time window identifier. Based on the actual passenger flow statistics, a true destination distribution is constructed to update the destination distribution generation model parameters, and the arrival prompt control information parameters are updated based on the difference between the actual passenger flow statistics and the alighting prediction information.

Images