A point flow prediction method based on point position fragment passenger flow

Abstract

The application discloses a point position passenger flow prediction method based on point position section passenger flow, and specifically comprises the following steps: firstly, collecting offline data, including a large number of point position continuous long-period passenger flow samples, position information, site information, point position information, environment information, surrounding information, environment information, merchant brand information and the like; then, constructing section passenger flow samples, and then performing data processing, feature engineering and model training; then, collecting and constructing section passenger flow videos, and establishing a section video number statistical model; finally, according to input section video samples and collected point position video time, position, site and the like information, required point position predicted daily passenger flow, monthly daily average passenger flow and annual monthly daily average passenger flow and the like passenger flow result set can be predicted. The application provides a simple, efficient and accurate point position passenger flow prediction method for merchant store expansion and activity site selection, and reduces manpower and material resources input.

Description

Technical Field

[0001] This invention belongs to the field of passenger flow prediction technology, specifically relating to a method for predicting passenger flow at specific locations based on location segments. Background Technology

[0002] With the emergence of online traffic growth bottlenecks and consumers upgrading their consumption habits, the demand for offline marketing by businesses and offline experiences by consumers is increasing, making offline traffic increasingly important. When businesses hold pop-up events or select locations for other store expansions, they need to conduct traffic assessments of the venues in order to find the best locations for pop-ups and store openings.

[0003] Currently, businesses assess foot traffic by manually observing and recording foot traffic at the sites under consideration for extended periods. They rely on experience to estimate approximate foot traffic or purchase foot traffic data directly from data service providers. This approach has several drawbacks. Firstly, long-term manual observation is time-consuming and labor-intensive, especially for chain brands with nationwide site selection needs, incurring significant manpower costs. Furthermore, manual estimations are only simple rough calculations without the support of large-scale data analytics. Secondly, even when purchasing data from service providers, existing foot traffic data covers a wide range and lacks location-level data. Foot traffic varies considerably between different locations within the same area, thus requiring more granular foot traffic data to help businesses make better site selections.

[0004] Given the aforementioned problems, it is essential to provide a location-level customer flow prediction method to address the urgent need of businesses for precise, fine-grained customer flow prediction. Summary of the Invention

[0005] To address the problem of existing technologies being unable to predict location-level customer traffic, this invention provides a location-based customer traffic prediction method that helps businesses accurately grasp location-level customer traffic, save costs, and improve the efficiency of site selection and evaluation for store expansion and events.

[0006] The technical solution adopted in this invention is as follows: a method for predicting passenger flow at specific locations based on location segments, comprising the following steps:

[0007] Step 1: Collect offline training data: point-of-sale passenger flow data Flow(poi,t), location information Loc_info(poi), venue information Place_info(place), point information Poi_info(poi), surrounding information Sur_info(place), environmental information Context (place), and merchant brand information Brand_info(brand);

[0008] Step 2: Construct segment passenger flow samples: Break down and combine the continuous passenger flow details of the points to generate a large number of segment passenger flow samples, including specific time, specific duration, passenger flow within the duration, and daily passenger flow.

[0009] Step 3: Data processing, feature engineering, and model training:

[0010] Step 4: Collect and prepare a large number of passenger flow video clips: These passenger flow video clips can be obtained by splitting the passenger flow video data collected in Step 1 into a large number of video clip samples, as well as several minutes of clips collected at different locations using mobile phones or other video recording devices, where N is any value from 0 to 60 minutes.

[0011] Step 5: Train the passenger flow statistics model: Using video people tracking algorithms on the collected video samples and publicly available video datasets, train and tune the optimal model suitable for the current scenario.

[0012] Step 6: Online inference for passenger flow prediction based on video clips of passenger flow.

[0013] Preferably, the point-of-sale passenger flow data Flow(poi,t) is detailed data fixed at the point for more than one day. The details refer to the specific time points when people with different identifiers appear at the point. Through the detailed data, the number of passengers in a specific time period, minute, or day within the date of the point video statistics can be counted. The data can be counted by the people counting video equipment fixed at the point or detailed data directly provided by a third-party data provider.

[0014] Location information (Loc_info(poi)) includes the province, city, district, latitude and longitude, business district, city level, and city type of the shooting location;

[0015] The Place_info(place) information includes the venue address, venue type, venue number, venue area, construction time, venue floor, average venue rent, distribution of brand stores, customer traffic level, consumption level, venue tag profile, venue business type distribution, venue customer demographic profile, and venue historical transaction information. The venue type includes shopping malls, office buildings, scenic spots, street shops, and residential communities. The venue refers to the specific shopping center, office building, residential community, building name and number.

[0016] Location information Poi_info(poi) contains specific information about the location. The location refers to the location of a store to be opened or a business event to be held, and the location information specifically refers to the detailed information of the location, including the floor, door number, the business type distribution on the floor, the current store brand, the nearby brands and industry information, the location rent, the store type, the store area, and the customer traffic.

[0017] Surrounding information (Sur_info(place)) includes transportation facilities and public facilities (gas stations, charging stations) around the site;

[0018] Context information includes weather, season, holidays, day of the week, and promotional activities;

[0019] The Brand_info information includes industry, company, number of chain stores, brand positioning, target customer group, average product price, and competitor brands.

[0020] Preferably, the entire process depends on the evaluation results of the actual test set, and includes missing value handling, outlier handling, feature encoding, feature derivation, feature transformation, feature selection, feature combination, model selection, and model parameter tuning. The methods used to train and predict daily passenger flow can be statistical methods, clustering + statistics, machine learning regression algorithms, and deep learning algorithms. A set of model fusion methods {M} is also available. i}, select the optimal model A; after predicting the daily passenger flow, it can also be used according to the method mentioned above {M i Construct the optimal model B to predict the average daily passenger flow on weekdays and non-weekdays for the current month, and then use the method mentioned above to predict the average daily passenger flow for the month {M}. i Construct the optimal model C to estimate the annual passenger flow at each location.

[0021] Preferably, there are two ways to predict the target of Model A: Method 1 predicts the target T1 by taking the daily passenger flow as the target, and Method 2 predicts the target T2 by taking the hour to which the passenger flow segment belongs as the proportion of the total passenger flow in a day.

[0022] Preferably, there are two methods to choose from for sample feature construction and prediction:

[0023] Method S1: Convert the passenger flow of a segment into the passenger flow of an hour based on the proportion of time and the proportion coefficient, and then construct features based on the specific hour, the hourly passenger flow and other information described in step 1.

[0024] Method 2 Sample S2: Construct features directly from the sample in step 2 and other information described in step 1.

[0025] Preferably, the statistical method can be used to simply predict daily passenger flow based on the hourly distribution of historical data.

[0026] The clustering + statistical method involves first clustering the locations, and then predicting the daily passenger flow based on the statistical method described above.

[0027] The machine learning algorithms include linear regression, ridge regression, lasso regression, decision tree, random forest, XGBoost, LightGBM, and combinations of various model results, which predict the daily passenger flow based on the passenger flow of a segment.

[0028] Deep learning methods include neural networks of different depths and combinations thereof;

[0029] The model fusion method includes combining the results obtained by any combination of the methods mentioned above.

[0030] Preferably, the specific steps for online inference of passenger flow prediction based on video clips of passenger flow are as follows:

[0031] First, collect information for the site visit task: including location information, location information, venue information, surrounding information, environmental information, and brand information for the activities to be carried out;

[0032] Secondly: Record the passenger flow videos of the locations to be evaluated using mobile devices or other passenger flow video recording devices. Several video segments of N minutes each can be recorded, with the recording time distributed throughout the business hours of the locations. Analyze the passenger flow and number of people in each video segment to obtain the video recording time and the corresponding passenger flow.

[0033] Then: Based on the time-based passenger flow at several locations, output the predicted passenger flow result set according to models A, B, and C in step 2. Each location segment corresponds to a passenger flow sample and will have a corresponding model prediction result R1.

[0034] Finally: Multiple segments at the same location will result in multiple predictions, while the actual passenger flow at a location has only one value. As the number of input segments increases, different strategies will be used to combine the output results of different segments at the same location to output a final result R2. There are several ways to output the results. These include simply averaging the results of each segment, averaging multiple samples after removing outlier segments, flexibly setting parameter weights based on the importance of different segments, or adding the predicted passenger flow results of different segments to the historical data based on the method set {Mi} and then training a model to output model D. During online inference, the predicted results R1 of different segments are added to other features of the passenger flow samples and then fed into model D, and the results are output again. Finally, the results of each segment are averaged to obtain the final result R2.

[0035] The beneficial effects of this invention are as follows: Addressing the current difficulties and pain points faced by businesses in selecting locations for store expansion and promotional activities, location-level customer traffic prediction provides businesses with more accurate customer traffic information, rather than being limited to large-scale customer traffic data, which can lead to information errors and subsequent investment losses. Furthermore, the method based on segmented customer traffic video reduces the manpower and effort required for location evaluation, making it a low-cost customer traffic prediction method. In summary, this invention helps businesses accurately grasp location-level customer traffic, save costs, and improve the efficiency of site evaluation for store expansion and promotional activities. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the specific method used in an embodiment of the present invention.

[0037] Figure 2 This is a diagram illustrating the specific results of using this method in an embodiment of the present invention. Detailed Implementation

[0038] The present invention will become more apparent from the following detailed description with reference to the accompanying drawings and tables.

[0039] A method for predicting passenger flow at specific locations based on location segments, comprising the following steps:

[0040] Step 1: Collect offline training data: passenger flow data Flow(t), location information Loc_info(poi), venue information Place_info(place), point information Poi_info(poi), surrounding information Sur_info(place), environmental information Context (place), merchant brand information Brand_info, etc.

[0041] Among them, the passenger flow data Flow(t) is detailed data fixed at the location for more than one day. The details refer to the specific time when people with different identifiers appeared at the location. Through the detailed data, the number of passengers in a specific time period, minute, or day within the video statistics date of the location can be counted. The data can be counted by video equipment that counts the number of people fixed at the location or detailed data directly provided by the data provider.

[0042] Location information Loc_info(poi) includes the province, city, district, latitude and longitude, business district, city level, city type, etc. of the location;

[0043] The Place_info(place) information includes the venue address, venue type, venue number, venue area, construction time, venue floor, average venue rent, distribution of brand stores, customer traffic level, consumption level, and other venue tag profiles, venue business distribution, venue customer profiles, and venue historical transaction information. Among them, the venue type includes shopping malls, office buildings, scenic spots, street shops, residential communities, etc., and the venue refers to the specific shopping center, office building, residential community, building, etc. name and number;

[0044] Location information Poi_info(poi) contains specific information about the location. A location refers to the location of a store to be opened or a commercial event to be held, and the location's foot traffic needs to be evaluated. The specific location information refers to detailed information about the location, including the location, floor, door number, business distribution on the floor, current store brands, nearby brands and industry information, location rent, store type, store area, and foot traffic.

[0045] Surrounding information (Sur_info(place)) includes transportation facilities and public facilities (gas stations, charging stations, etc.) around the site;

[0046] Context information includes weather, season, holidays, days of the week, promotional activities, etc.

[0047] The Brand_info information includes industry, company, number of chain stores, brand positioning, target customer group, average product price, and competitor brands.

[0048] Step 2: Construct passenger flow samples: Break down and combine the continuous passenger flow details of the points to generate a large number of fragmented passenger flow samples, including specific time (date and hour), specific duration (unit / second), passenger flow within the duration, and passenger flow for the day;

[0049] Step 3: Data processing, feature engineering, and model training:

[0050] Model A has two prediction objectives: Method 1 uses the daily passenger flow as the prediction objective T1, and Method 2 uses the hour to which the passenger flow segment belongs to the total passenger flow in a day as the prediction objective T2.

[0051] There are two ways to construct and predict sample features:

[0052] Method S1: Convert the passenger flow of a segment into the passenger flow of an hour based on the proportion of time and the proportion coefficient, and then construct features based on the specific hour, the hourly passenger flow and other information described in step 1.

[0053] Method 2 Sample S2: Construct features directly from the samples in step 2 and other information described in step 1;

[0054] The entire process depends on the evaluation results on the actual test set and includes missing value handling, outlier handling, feature encoding, feature derivation, feature transformation, feature selection, feature combination, model selection, and model parameter tuning. Methods used to train the prediction of daily passenger flow can include statistical methods, clustering + statistics, machine learning regression algorithms, deep learning algorithms, and model fusion, among others. i}, select the optimal model A.

[0055] The statistical method can be used to simply predict daily passenger flow based on the hourly distribution of historical data.

[0056] The clustering + statistical method involves first clustering the locations, and then predicting the daily passenger flow based on the statistical method described above.

[0057] The machine learning algorithms include linear regression, ridge regression, lasso regression, decision tree, random forest, XGBoost, LightGBM, and combinations of various model results, which predict the daily passenger flow based on the passenger flow of a segment.

[0058] Deep learning methods include neural networks of different depths and combinations thereof;

[0059] The model fusion method includes combining the results obtained by any combination of the methods mentioned above.

[0060] After predicting the daily passenger flow, we can also use the methods mentioned above {M i Construct the optimal model B to predict the average daily passenger flow on weekdays and non-weekdays for the current month, and then use the method mentioned above to predict the average daily passenger flow for the month {M}. i Construct the optimal model C to estimate the annual passenger flow at each location.

[0061] Step 4: Collect and prepare a large number of passenger flow video clips: These passenger flow video clips can be obtained by splitting the passenger flow video data from Step 1 into a large number of video clip samples, as well as several minutes of clips collected at different locations using mobile phones or other video recording devices, where N is any value from 0 to 60 minutes.

[0062] Step 5: Train the passenger flow statistics model: By tracking people in videos, train and tune the parameters of the collected video samples and the publicly available video dataset to find the optimal model suitable for the current scenario.

[0063] Step 6: Online inference for passenger flow prediction based on video clips:

[0064] First, collect information for the site visit task: including location information, location information, venue information, surrounding information, environmental information, and brand information for the activities to be carried out;

[0065] Secondly: Record the passenger flow videos of the locations to be evaluated using mobile devices or other passenger flow video recording devices. Several N-minute videos can be recorded, with the recording time distributed within the business hours of the location (videos of duration from 1 to 60 minutes within any hour during the activity period); analyze the passenger flow and number of people in each video segment in turn to obtain the video recording time and its corresponding passenger flow.

[0066] Then: Based on the time-based passenger flow at several locations, output the predicted passenger flow result set according to models A, B, and C in step 3. Each location segment corresponds to a passenger flow sample and will have a corresponding model prediction result R1.

[0067] Finally: Multiple segments at the same location will result in multiple predictions, while the actual passenger flow at a location has only one value. As the number of input segments increases, different strategies will be used to combine the output results of different segments at the same location to output a final result R2. There are several ways to output the results. These include simply averaging the results of each segment, averaging multiple samples after removing outlier segments, flexibly setting parameter weights based on the importance of different segments, or adding the predicted passenger flow results of different segments to the historical data based on the method set {Mi} and then training a model to output model D. During online inference, the predicted results R1 of different segments are added to other features of the passenger flow samples and then fed into model D, and the results are output again. Finally, the results of each segment are averaged to obtain the final result R2.

[0068] This invention is not limited to the specific embodiments described above. Those skilled in the art can implement this invention using various other specific embodiments based on the content disclosed herein. Therefore, any design that adopts the design structure and concept of this invention with some simple changes or modifications falls within the protection scope of this invention.

Claims

1. A method for predicting passenger flow at specific locations based on location segments, characterized in that: The steps are as follows: Step 1: Collect offline training data: flow of customers at locations, location information Loc_info, venue information Place_info and location information Poi_info, surrounding information Sur_info, environmental information Context, and merchant brand information Brand_info; Step 2: Construct segment passenger flow samples: Break down and combine the continuous passenger flow details of the points to generate segment passenger flow samples, including specific time, specific duration, passenger flow within the duration, and passenger flow for the day; Step 3: Data processing, feature engineering, and model training: Step 4: Collect and prepare passenger flow video clips: The passenger flow video clips can be obtained by splitting the passenger flow video data collected in Step 1 into video clip samples, as well as several minutes of clips collected at different locations by mobile phones or other video recording devices, where N is any value from 0 to 60 minutes. Step 5: Train the passenger flow statistics model: Using video people tracking algorithms on the collected video samples and publicly available video datasets, train and tune the optimal model suitable for the current scenario. Step 6: Online inference for passenger flow prediction based on video clips of passenger flow; The point-of-sale passenger flow data is detailed data fixed at the point for more than one day. The detailed data refers to the specific time when people with different identifiers appeared at the point. Through the detailed data, the number of passengers in a specific time period, minute, or day within the date of the video statistics of the point can be counted. The data can be counted by video counting equipment that is fixed at the point or detailed data directly provided by a third-party data provider. Location information (Loc_info) includes the province, city, district, latitude and longitude, business district, city level, and city type of the shooting location; The Place_info information includes the venue address, venue type, venue number, venue area, construction time, venue floor, average venue rent, distribution of brand stores, customer traffic level, consumption level, venue tag profile, venue business type distribution, venue customer demographic profile, and venue historical transaction information. The venue type includes shopping malls, office buildings, scenic spots, street shops, and residential communities. The venue refers to the specific shopping center, office building, residential community, building name and number. Location information Poi_info contains specific information about the location. The location refers to the location of a store to be opened or a business event to be held, and the location information specifically refers to the detailed information of the location, including the floor, door number, the business type distribution on the floor, the current store brand, nearby brands and industry information, location rent, store type, store area, and customer traffic. Surrounding information (Sur_info) includes transportation and public facilities around the site; Context information includes weather, season, holidays, day of the week, and promotional activities; The Brand_info information includes industry, company, number of chain stores, brand positioning, target customer group, average product price, and competitor brands; The entire processing includes missing value handling, outlier handling, feature encoding, feature derivation, feature transformation, feature selection, feature combination, model selection, and model parameter tuning. The methods used to train the passenger flow statistics model include at least one of statistical methods, clustering + statistics, machine learning regression algorithms, and deep learning algorithms; the model fusion method set { }, select the optimal model A; after predicting the daily passenger flow, also based on the method set { Construct the optimal model B to predict the average daily passenger flow on weekdays and non-weekdays of the current month, and then, based on the average daily passenger flow for the month, use the method set { Construct the optimal model C to estimate the annual passenger flow at each location.

2. The method for predicting passenger flow at a location based on location segment passenger flow according to claim 1, characterized in that: Model A has two prediction objectives: Method 1 uses the daily passenger flow as the prediction objective T1, and Method 2 uses the hour to which the passenger flow segment belongs to the total passenger flow in a day as the prediction objective T2.

3. The method for predicting passenger flow at a location based on location segment passenger flow according to claim 1, characterized in that: There are two ways to construct a segment of passenger flow samples: Method S1: Convert the passenger flow of a segment into the passenger flow of an hour based on the proportion of time and the proportion coefficient, and then construct features based on the specific hour, the hourly passenger flow and other information described in step 1. Method 2 Sample S2: Construct features directly from the sample in step 2 and other information described in step 1.

4. The method for predicting passenger flow at a location based on location segment passenger flow according to claim 3, characterized in that: The statistical method can be used to simply predict daily passenger flow based on the hourly distribution of historical data. The clustering + statistical method involves first clustering the locations, and then predicting the daily passenger flow based on the construction method of sample S1 described above. The machine learning algorithms include linear regression, ridge regression, lasso regression, decision tree, random forest, XGBoost, LightGBM, and combinations of various model results, which predict the daily passenger flow based on the passenger flow of a segment. Deep learning methods include neural networks of different depths and combinations thereof; The model fusion method includes combining the results obtained from any combination of statistical methods, clustering + statistical methods, machine learning algorithms and deep learning methods.

5. The method for predicting passenger flow at a location based on location segment passenger flow according to claim 1, characterized in that: The specific steps for online inference for passenger flow prediction based on video clips are as follows: First, collect information for the site visit task: including location information, venue information, location information, surrounding information, environmental information, and merchant brand information from step 1; Secondly: Record the passenger flow videos of the locations to be evaluated using mobile devices or other passenger flow video recording devices. Several video segments of N minutes each can be recorded, with the recording time distributed throughout the business hours of the locations. Analyze the passenger flow and number of people in each video segment to obtain the video recording time and the corresponding passenger flow. Then: Based on the time-based passenger flow at several locations, output the predicted passenger flow result set according to models A, B, and C respectively. One segment of a location corresponds to one passenger flow sample, which corresponds to one prediction result R1 of the model. Finally: Multiple segments at the same location will result in multiple predictions, while the actual passenger flow at a location has only one value. As the number of input segments increases, different strategies will be used to combine the output results of different segments at the same location to output a final result R2. There are several ways to output the results. These include simply averaging the results of each segment, averaging multiple samples after removing outlier segments, flexibly setting parameter weights based on the importance of different segments, or adding the predicted passenger flow results of different segments to the historical data based on the method set {Mi} and then training a model to output model D. During online inference, the predicted results R1 of different segments are added to other features of the passenger flow samples and then fed into model D, and the results are output again. Finally, the results of each segment are averaged to obtain the final result R2.

Images