CN-122024152-A - Deep learning-based passenger flow statistics method, system and storage medium

CN122024152ACN 122024152 ACN122024152 ACN 122024152ACN-122024152-A

Abstract

The invention discloses a passenger flow statistical method, a passenger flow statistical system and a storage medium based on deep learning, and relates to the technical field of passenger flow statistics. The method comprises the steps of firstly, setting up different monitoring ends in a passenger flow region, extracting passenger flow characteristics, analyzing the passenger flow characteristics, constructing a characteristic database based on analysis results, secondly, configuring an implementation database, generating a corresponding solution and constructing a solution database through the problems generated in the implementation process of the implementation database statistics, thirdly, constructing a passenger flow statistical model, carrying out statistics on passenger flows, intensively collecting statistical data generated in the statistical process, establishing a collection storage library for storage, and simultaneously, intensively collecting the problems generated in the implementation process, and carrying out feedback adjustment based on the problems at a later stage.

Inventors

HUANG YIWEN
LIU DONG
HUANG ZUHONG

Assignees

苏州万店掌网络科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251201

Claims (10)

1. The passenger flow statistical method based on deep learning is characterized by comprising the following steps of: Firstly, setting up different monitoring ends in a passenger flow circulation area, extracting passenger flow characteristics, analyzing the passenger flow characteristics, and constructing a characteristic database based on analysis results; Step two, configuring an embodiment database, wherein the embodiment database comprises a first embodiment, a second embodiment and a third embodiment, generating corresponding solutions and constructing a solution database through the problems generated in the implementation process of the embodiment database statistics; The third step is to build a passenger flow statistical model to count passenger flow, collect the statistical data generated in the statistical process in a concentrated way, build a collection storage library to store, collect the problems generated in the implementation process in a concentrated way, and perform feedback adjustment based on the problems in the later period.
2. The deep learning based passenger flow statistics method as recited in claim 1, wherein the passenger flow features comprise appearance features, motion features and aggregation features, wherein the appearance features comprise global features, local features and semantic attribute features, the motion features comprise position and speed and motion trajectories, and the aggregation features comprise flow features, density features and behavior and time sequences; The extraction method of the appearance features comprises the steps that global features are depth feature vectors extracted from a whole detection frame, a convolutional neural network pre-trained on a large pedestrian re-recognition data set is adopted to extract high-dimensional vectors of head and shoulder appearance information of a coding pedestrian, local features are extracted by focusing on specific local areas of the body of the pedestrian, specifically, features of all areas are extracted after key points of the pedestrian are positioned firstly by using a gesture estimation model, or features are extracted respectively after a pedestrian image is horizontally divided into a plurality of stripes, semantic attribute features are used for recognizing high-level semantic attributes of the pedestrian, and classification networks are adopted to judge the sex, age bracket, wearing of a hat, color of a jacket, style of a lower garment, knapsack or the like of the pedestrian; The motion track is a track formed by connecting continuous position points of a target in a period of time; The method for extracting the aggregation features comprises the steps of flow features including incoming and outgoing passenger flows, total number of passengers in a specific time period, throughput, passenger flow passing through a certain section in unit time, conversion rate and core indexes of retail industry, density features including regional real-time density, number of people in unit area, average density and average density in a period of time, and behavior and time sequence features including average residence time, go sight-seeing routes and space-time distribution diagrams.
3. The deep learning-based passenger flow statistical method according to claim 2, wherein in the first step, association rule analysis, predictive analysis, cluster analysis and root cause analysis are adopted when analyzing features, and each analysis method specifically comprises: The method comprises the steps of association rule analysis, predictive analysis, clustering analysis, root cause analysis, multi-dimensional cross analysis and problem root positioning, wherein the association rule analysis is used for mining hidden relations among passenger flow characteristics, the predictive analysis is used for predicting short-term or long-term passenger flow in the future by adopting a long-term memory network model, the cluster analysis is used for grouping customers by adopting an unsupervised learning algorithm, and corresponding strategies are formulated for different groups, and the root cause analysis is used for checking reasons through drill-down analysis when abnormal data occur, and the multi-dimensional cross analysis is used for combining external data.
4. The passenger flow statistical method based on deep learning of claim 3, wherein the problems generated in the implementation process of statistics in the second step include shielding problems, illumination change problems, model precision and speed balancing problems and perspective effect problems, and the problems are specifically as follows: The method comprises the steps of screening, namely screening a plurality of people when people are dense, so that a detection frame is incomplete or lost, illuminating, namely, changing light in the daytime, at night, on cloudy days, on sunny days and indoors, so that the difference of image quality is large, balancing model precision and speed, namely, the high-precision model is low in running speed and cannot process multiple paths of videos in real time, and perspective effect, namely, when a camera looks down to shoot, imaging of a far person is smaller than that of a near person, so that deviation occurs in detection and counting.
5. The deep learning-based passenger flow statistics method as set forth in claim 4, wherein the solution database in the second step is constructed by the following concrete scheme: The shielding problem is solved by utilizing the appearance characteristic re-identification capability of a depth simple online real-time tracking algorithm and adopting representative points to detect key points; the illumination change problem is solved by adding data under illumination condition during model training through data enhancement processing, and adopting the characteristic of insensitivity to illumination; the model precision and speed balance solution is that a lightweight model is selected, combined with a model pruning technology, and deployed in edge computing equipment; The perspective effect problem solving scheme is that image coordinates are mapped to real world coordinates through perspective transformation or camera calibration, and people in different positions are weighted or normalized.
6. The deep learning-based passenger flow statistical method of claim 5, wherein the passenger flow statistical model in the third step is constructed by the following method: The monocular vision statistical method based on traditional image processing comprises the steps of adopting a single camera to identify human shapes and counting through virtual lines or designated areas; The monocular vision statistical method based on deep learning comprises the steps of adopting a YOLO deep learning target detection model, accurately identifying each person in a video frame, distributing unique IDs to each person through a depth simple online real-time tracking algorithm and forming a motion track, and finally judging the entering and exiting directions of the persons through a virtual line and finishing counting; the binocular vision or 3D stereoscopic vision statistical method comprises calculating target depth information by parallax by adopting two cameras to realize passenger flow statistics; The statistical method of the 3D depth camera is that infrared speckle or light pulse is emitted and return light is received, so that a depth map of a scene is directly obtained, and passenger flow statistics can be completed without depending on visible light.
7. The deep learning-based passenger flow statistics method as set forth in claim 6, wherein in the third step, a collection storage bank dedicated for storing passenger flow statistics is established, and then the passenger flow statistics collected in a centralized manner are transmitted to the collection storage bank in real time, and the collection storage bank manages the passenger flow statistics in a classified archiving and safe storage manner.
8. The deep learning-based passenger flow statistical method of claim 7, wherein the implementing feedback process in the third step is characterized in that passenger flow characteristics are analyzed first, a feedback adjustment scheme is generated based on analysis results, the feedback adjustment scheme comprises a first feedback adjustment scheme, a second feedback adjustment scheme and a third feedback adjustment scheme, wherein the adjustment force of the first feedback adjustment scheme is the largest, and the adjustment force of the third feedback adjustment scheme is the smallest.
9. The passenger flow statistical system based on deep learning is characterized in that the passenger flow statistical system is suitable for the passenger flow statistical method based on deep learning as claimed in claim 8, and comprises a feature extraction analysis unit, an implementation scheme formulation unit and a passenger flow statistical unit, wherein the functions of the units are as follows: The characteristic extraction and analysis unit is used for setting up different monitoring ends in the passenger flow circulation area, extracting passenger flow characteristics and analyzing the passenger flow characteristics, and constructing a characteristic database based on analysis results; an embodiment formulation unit for configuring an embodiment database including the first embodiment, the second embodiment and the third embodiment, counting the problems generated during the implementation by the database, generating corresponding solutions and constructing a solution database; The passenger flow statistics unit is used for constructing a passenger flow statistics model to count passenger flows, collecting statistics data generated in the statistics process in a concentrated manner, establishing a collection storage library to store, collecting problems generated in the implementation process in a concentrated manner, and carrying out feedback adjustment based on the problems in the later period.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when executed by a processor, implements the deep learning based passenger flow statistics method according to any of claims 1-8.

Description

Deep learning-based passenger flow statistics method, system and storage medium Technical Field The invention relates to the technical field of passenger flow statistics, in particular to a passenger flow statistics method, a passenger flow statistics system and a passenger flow statistics storage medium based on deep learning. Background Currently, passenger flow statistics refers to the process of systematically collecting, recording, analyzing and producing reports on the number of people flowing, behavior characteristics and moving direction in a specific area and a specific time period by using various technical means. However, the conventional passenger flow statistical method has limitations in characteristic limitation that on one hand, various passenger flow characteristics generated in the statistical process cannot be covered comprehensively, on the other hand, most conventional statistical methods can only be limited according to basic characteristics of human bodies, and are difficult to adapt to statistical requirements in complex scenes, and finally, deviation between statistical results and actual conditions is caused. Disclosure of Invention The invention aims to provide a passenger flow statistics method, a passenger flow statistics system and a passenger flow statistics storage medium based on deep learning so as to solve the problems in the background technology. In order to achieve the above purpose, the invention provides a passenger flow statistical method, a passenger flow statistical system and a storage medium based on deep learning, which comprises the following steps: Firstly, extracting passenger flow characteristics by setting different monitoring ends in passenger flow circulation areas, analyzing the characteristics after extraction is completed, and making a characteristic database after analysis is completed; Step two, configuring an embodiment database, wherein the embodiment database comprises a first embodiment, a second embodiment and a third embodiment, and the embodiment database is used for counting the problems generated in the implementation process, generating a solution after counting, and forming a solution database; and thirdly, constructing a passenger flow statistical model, counting passenger flows, collecting generated statistical data in a centralized manner in the statistical process, formulating a collection storage library for storage, collecting generated problems in a centralized manner in the implementation process, and performing feedback adjustment in the later period. Preferably, the passenger flow characteristics comprise appearance characteristics, movement characteristics and aggregation characteristics, wherein the appearance characteristics comprise global characteristics, local characteristics and semantic attribute characteristics, the movement characteristics comprise position, speed and movement track, and the aggregation characteristics comprise flow characteristics, density characteristics and behavior and time sequence; Extracting global features, namely extracting depth feature vectors from the whole detection frame, and extracting a high-dimensional vector by using a convolutional neural network pre-trained on a large pedestrian re-identification data set, wherein the high-dimensional vector encodes pedestrian head and shoulder appearance information; extracting local characteristics, namely focusing on a specific local area of a pedestrian body, positioning key points of the pedestrian by using a gesture estimation model, and then respectively extracting the characteristics of the areas or horizontally dividing a pedestrian image into a plurality of stripes to respectively extract the characteristics; the motion characteristic extraction method comprises the steps of extracting position and speed, representing the position by the coordinates of the central point of a target detection frame, extracting a motion track, namely connecting continuous position points of a target in a period of time to form a track; The aggregate feature extraction method comprises the steps of extracting flow features, namely, incoming/outgoing passenger flows, total passenger numbers in a specific time period, throughput, passenger flows passing through a certain section in unit time, conversion rate, retail industry core indexes, extracting density features, namely, regional real-time density and number in unit area, average density and average density in a period of time, extracting behaviors and time sequences, namely, average residence time, go sight-seeing routes and space-time distribution diagrams. Preferably, in the first step, association rule analysis, predictive analysis, cluster analysis and root cause analysis are adopted when the characteristics are analyzed, the analysis method specifically comprises association rule analysis, predictive analysis, clustering analysis and root cause analysis, wherein the association rule analysis comprises