CN114169659A - Criminal spatiotemporal risk prediction and decision scheduling method and device - Google Patents

Criminal spatiotemporal risk prediction and decision scheduling method and device Download PDF

Info

Publication number
CN114169659A
CN114169659A CN202110076296.5A CN202110076296A CN114169659A CN 114169659 A CN114169659 A CN 114169659A CN 202110076296 A CN202110076296 A CN 202110076296A CN 114169659 A CN114169659 A CN 114169659A
Authority
CN
China
Prior art keywords
database
predicted
road section
road
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110076296.5A
Other languages
Chinese (zh)
Inventor
顾海硕
陈鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
PEOPLE'S PUBLIC SECURITY UNIVERSITY OF CHINA
Original Assignee
PEOPLE'S PUBLIC SECURITY UNIVERSITY OF CHINA
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by PEOPLE'S PUBLIC SECURITY UNIVERSITY OF CHINA filed Critical PEOPLE'S PUBLIC SECURITY UNIVERSITY OF CHINA
Priority to CN202110076296.5A priority Critical patent/CN114169659A/en
Publication of CN114169659A publication Critical patent/CN114169659A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0635—Risk analysis of enterprise or organisation activities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • G06Q50/265—Personal security, identity or safety

Landscapes

  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Engineering & Computer Science (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Tourism & Hospitality (AREA)
  • Entrepreneurship & Innovation (AREA)
  • General Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Theoretical Computer Science (AREA)
  • Development Economics (AREA)
  • Educational Administration (AREA)
  • Physics & Mathematics (AREA)
  • Quality & Reliability (AREA)
  • Game Theory and Decision Science (AREA)
  • Operations Research (AREA)
  • Computer Security & Cryptography (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a method and a device for criminal spatiotemporal risk prediction and decision scheduling, which comprise the following steps: s100, acquiring data resources and establishing a corresponding database; s200, carrying out data management to obtain output data; s300, predicting the probability of a future one-week occurrence of the road section to be predicted based on the CTR prediction model and the output data of the deep learning; s400, regularizing the probability of the future one-week occurrence of the road section to be predicted, a receptive field screener, a district screener, a hotspot migration rule base and a historical case event database to obtain the corrected probability of the future one-week occurrence of the road section to be predicted; and S500, carrying out police dispatching and patrol route planning. The invention can enhance the flexibility of the input of the prediction model, improve the mobility of the system, enhance the interpretability of the prediction result, increase the prediction probability correction module, improve the flexibility of the correction of the prediction result, increase the scheduling module and simplify the research and judgment and scheduling process of the prediction result.

Description

Criminal spatiotemporal risk prediction and decision scheduling method and device
Technical Field
The invention relates to the field of spatiotemporal information processing, in particular to a method and a device for criminal spatiotemporal risk prediction and decision scheduling.
Background
The mainstream crime space-time risk prediction system mainly comprises two modules of data management and future case probability prediction. The data management module can process the input multi-source heterogeneous data into a format which can be input by the probability estimation module. After receiving the processed data, the future issue probability estimation module carries out online or offline training, the prediction model is online and provides implementation prediction service after reaching proper precision, and the parameters of the online model can be updated in real time according to the data or can adapt to the change of the data by adopting an offline updating mode.
In the prior art, the mainstream technology of criminal spatiotemporal prediction is to predict the probability of a crime in a certain future space and time by a grid-based prediction model. Although such a grid-based pattern is simple in preprocessing, spatial heterogeneity is not considered, and in contrast, a more reasonable analysis vehicle is road network data. The road network is one of the important carriers for transmitting space-time information outside a network space, criminal activities are generated in space-time, the movement of a committing main body in space-time is directly related to the distribution of the road network, and the grids cannot accurately reflect the spatial adjacency relation.
The prediction scale adopted by the existing criminal spatiotemporal prediction system is relatively large, and is not beneficial to actual decision-making command and scheduling. In order to prevent the data sparseness problem caused by the fact that the grid is too small, most prediction systems adopt the grid with the side length of dozens of meters or even hundreds of meters as a prediction unit, and the prediction unit has the advantages that the sparseness degree of input data is reduced, so that the prediction accuracy is improved, but the prediction unit has the defect that the prediction range is too large, and the specific patrol and control is difficult to achieve when the prediction unit is actually applied to the ground.
The conventional crime space-time prediction system takes the number of issued cases as a prediction index for judging the risk level of a region, the index is intuitive but is not reasonable for the space-time position of a zero sample, because crime hotspots can migrate, and the migration rule is related to space-time heterogeneity but is not necessarily adjacent to space-time. Therefore, for risk prediction of a zero sample spatiotemporal position, the existing crime spatiotemporal prediction system has no better solution, and even if the predicted value of the model appears in a zero sample area, reasonable explanation cannot be given. The risk in space and time is actually an embodiment of the size of the crime chance, and the number of cases is only one expression form of the crime chance and cannot represent the size of the crime chance. For example, although the number of cases in some areas is not large, even no case records exist in historical data, the concealment of the environment and the laggard security conditions greatly improve the success rate of crimes, and once a criminal and a potential victim target appear simultaneously, the criminal is probably caused. Thus, in addition to environmental factors, criminal and potential target activities also determine the occurrence of crimes. For the travel mode of a criminal, a criminal travel theory exists in environmental criminal science, and the basic idea is that the travel distance of criminal crime presents an exponential decay trend within a certain range. This description is similar to the "gravitation model" in the traffic domain. However, considering the influence of space-time opportunities on human trip in modern society, the distance cannot be the only factor limiting human trip, and for criminal trip, the distribution of criminal opportunities on space-time is also an important factor influencing criminal trip. In 2019, the Yan courage and Liu Bi Fang propose a unified human travel opportunity model which models human travel behaviors from the perspective of space-time opportunities, and compared with a gravitation model (similar to a distance attenuation model in criminal travel), the model has the advantages of better prediction effect and higher interpretability. Therefore, the simple case rate is corrected by the time-space environment information, the potential travel routes of criminals and the activity information of people, and the method is more reasonable compared with the method of representing the time-space risk by only using the case quantity.
The existing crime space-time prediction system does not fully consider the heterogeneity of the space, such as the influence of distribution of the district in charge on the case issuing condition, the influence of crime hotspot transfer on the case issuing probability, the influence of activities of people on the space-time on crime opportunities, and the like. The crime mode theory of environmental crime provides three concepts of a crime attraction area, a crime generation area and a crime neutral area, the probability of the collision of the accumulation and dispersion activities of people on time and space, criminals and potential targets is considered as a crime generation mechanism, and the conventional crime prediction method or system does not have a reasonable quantitative model for the activities of people and crime opportunities. An effective distance measurement method based on the flow size between space-time positions is proposed by Dirk Brockmann and Dirk Helbin in Science of 2013, and the method can effectively predict the origin of infectious diseases. For environmental criminals, the effective distance measurement method can model the influence of crowd dispersion activities on crime opportunities in space-time, and the distances from other areas to areas with high aggregation are relatively shorter, while the accessibility of areas with close spatial aggregation but relatively smaller inter-area traffic is relatively poor.
The comprehensive data dimension of the existing crime spatiotemporal prediction system is not comprehensive enough, the issue risk influencing a spatiotemporal position is related to various factors, the data adopted by the existing system are mostly population grid density, POI, land type and the like, the data can reflect the local characteristics of an area from the macro and mesoscopic levels, but are not detailed into the actual scene, such as the description of the landscape, the description of the building, the subjective feeling of people on the scene and the like. Zhang Sail et al, computer, environmental and Urban Systems, 2018, combo 64 types of common objects of city street views, propose to use scene semantic trees to quantitatively analyze street view content and categories, and they also use Place Pulse data sets to perform migration learning to obtain the representation of other street views. The Place Pulse data set labels six common human feelings such as safety, depression and vitality of the Google street view, and although the Place Pulse does not have too much data of the Chinese area, the trained model can be applied to cognitive feeling prediction of the street view in other areas through a transfer learning technology. Street view can potentially reflect the social and economic conditions, and compared with null point data such as POI (point of interest) and the like, the implied information can more directly reflect the cognition of people to a scene.
The predictive model in the existing predictive system has poor mobility, and for the grid-based model, the dimension of input data is changed due to the change of the predicted spatial range, or the source of the input data is changed due to the change of the predicted spatial position, and the like, so that the trained model can be redesigned and trained.
The existing prediction system has no reasonable interpretability of the prediction result and cannot correct prediction errors caused by the self deviation of a prediction model. Most of the existing prediction models are models based on deep learning, although the prediction accuracy is generally not too low after a large amount of data training, no system can judge whether the prediction result is reasonable or not when the prediction result is obtained, and the mechanism behind the data cannot be explained, or the error result caused by the self deviation of the model cannot be corrected.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a method and a device for forecasting the crime space-time risk and scheduling a decision, which can enhance the flexibility of input of a forecasting model, improve the mobility of a system, enhance the interpretability of a forecasting result, increase a pre-estimation probability regular module, improve the flexibility of correcting the forecasting result and increase a scheduling module to simplify the research and judgment and scheduling processes of the forecasting result.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
a crime spatiotemporal risk prediction and decision scheduling method comprises the following steps:
s100, acquiring data resources and establishing a corresponding database, wherein the data resources comprise case event space-time data, space-time background environment data and main body behavior data, and the data resources comprise: establishing a historical case event database and a suspect foothold database based on the case event time-space data, establishing a time-space environment database, a road network database and a street view database based on the time-space background environment data, and establishing a time-space track database and a real-time police resource distribution database based on the main body behavior data;
s200, based on the road section to be predicted, the suspect foothold database, the time-space environment database, the road network database, the street view database, the time-space trajectory database, a road network calculation engine, a graph calculation engine and a receptive field screener, performing data management to obtain output data, wherein the output data comprises: crime opportunities from potential foot-falling points of all criminals in a receptive field of a road section to be predicted to the road section to be predicted, social structure characteristics of the criminals, direction entropies of all road sections in the receptive field, cognitive feature vectors of the road section to be predicted, and road network structure characteristics of the road section to be predicted in the receptive field;
s300, predicting the probability of occurrence of a future week of the road section to be predicted based on a CTR prediction model for deep learning, the output data and the accumulated risk value of the road section to be predicted, wherein the accumulated risk value of the road section to be predicted is calculated based on crime opportunities from potential foothold of all criminals in the receptive field of the road section to be predicted to the road section to be predicted;
s400, regularizing the probability of the future one-week occurrence of the road section to be predicted, the receptive field screener, the jurisdiction screener, the hotspot migration rule base and the historical case event database to obtain the corrected probability of the future one-week occurrence of the road section to be predicted;
s500, carrying out police dispatching and routing planning based on the latest position information of the police, the receptive field screener, the probability of the future one-week case of all road sections in each jurisdiction, the road network database, the road network calculation engine and the real-time police force resource distribution database.
Further, in the method as described above, in S100,
for the case event space-time data, storing the space-time position data in the case event space-time data into a geographic database, and storing other attribute information into a relational database;
for the space-time environment database and the road network database, storing the space position data in the space-time environment database and the road network database into a geographic database, and storing other attribute information into a relational database;
for the street view database, storing spatial position data in the street view database into a geographic database, storing images into a file system, and storing vectorized images into a vector retrieval engine;
and storing the subject behavior data into a time sequence database.
Further, the method as described above, in S200, acquiring crime opportunities from potential foothold of all criminals in the receptive field of the road segment to be predicted to the road segment to be predicted, includes:
the method comprises the steps of obtaining a road section jump effective distance map with space-time risks based on a unified opportunity travel model and a map calculation engine, obtaining potential footfall points of all criminals from a suspect footfall point database in a receptive field obtained based on a road section to be predicted and a receptive field screener, and calculating travel probability from the potential footfall points to the road section to be predicted, namely crime opportunities based on the road section jump effective distance map.
Further, in the method as described above, in S200, obtaining social structural features of all criminals in the receptive field of the road segment to be predicted includes:
and constructing a potential social relationship based on the crime relations, the household registration relations and the space approach relations of all the criminals in the receptive field of the road section to be predicted, and acquiring the social structure characteristics of the criminals by using a node2vec model based on the potential social relationship.
Further, in the method as described above, in S200, acquiring a road network structure characteristic of the road segment to be predicted in the receptive field includes:
and screening corresponding road network data based on the road network database in the receptive field obtained based on the road section to be predicted and the receptive field screener, carrying out graph embedding operation on the screened road network data through the graph calculation engine, and obtaining the road network structure characteristics of the road section to be predicted in the receptive field by using a node2vec model.
Further, in the method as described above, in S200, acquiring the cognitive feature vector of the road segment to be predicted includes:
based on the street view database and a pre-training deep learning model, carrying out panoramic segmentation on street view data of the road section to be predicted, based on an entity and a scene semantic tree obtained by segmentation, obtaining a street view map corresponding to the road section to be predicted, taking the street view map corresponding to the road section to be predicted as an objective cognitive representation of a scene, and taking a six-dimensional cognitive vector as a subjective cognitive representation of the scene;
the method comprises the steps of sampling corresponding road network data to form a data set based on a road section to be predicted and a road network database, sequentially passing objective cognitive features of each group of objective cognitive features and subjective cognitive features through a first graph neural network layer, a graph pooling layer, a second graph neural network layer and a graph reading layer to obtain corresponding objective cognitive feature vectors when the data set is trained, sequentially performing splicing operation and linear regression prediction on the corresponding objective cognitive feature vectors and subjective cognitive feature vectors corresponding to the subjective cognitive features to obtain a road section risk value, and finally outputting the cognitive feature vectors of the road section to be predicted.
Further, the method as described above, S400 includes:
and merging the total range of the district in which the road section to be predicted is located and the adjacent district thereof screened by the district screener with the receptive field range screened by the receptive field screener to obtain a merged range, acquiring all crime high-incidence road sections in the merged range based on the historical case event database, calculating the transfer probability of all crime high-incidence road sections to the road section to be predicted based on all crime high-incidence road sections in the merged range and the hotspot transfer rule in the hotspot transfer rule base, taking the maximum value of all transfer probabilities, and correcting the case distribution probability of the road section to be predicted for one week in the future by a Bayesian formula.
Further, the method as described above, S500 includes:
when the latest alarm condition occurs, based on the occurrence position information of the latest alarm condition and the receptive field screener, a preset time circle is obtained by taking a street corresponding to the latest alarm condition as a starting point, the preset time circle and the area range to which the latest alarm condition belongs are intersected to obtain an intersection range, all on-duty patrol control alarm forces in the intersection range are obtained based on the real-time alarm force resource distribution database, patrol control schemes with the states not being the highest priority are screened out from the current patrol control schemes and added into a scheduling candidate list, the road network data in the calculation range in the road network database are called through the road network calculation engine, the patrol control scheme with the shortest time consumption is screened out from the scheduling candidate list, the states are set as the highest priority, and corresponding patrol control route information is obtained.
Further, in the method as described above, S500 further includes:
and when the latest warning condition does not appear, calculating to obtain the patrol control schemes of the police force respectively conforming to three patrol control targets with the maximum patrol control range, the maximum accumulated patrol control issuing area, the maximum risk and the minimum patrol control overlap ratio among the police forces in a certain time by taking the position of the police force dispatched by each jurisdiction as a starting point through the road network calculation engine based on the issue probability of all road sections in each jurisdiction in one week in the future, and obtaining the patrol control route information corresponding to each patrol control scheme.
A crime spatiotemporal risk prediction and decision scheduling device, comprising:
the acquisition module is used for acquiring data resources and establishing a corresponding database, wherein the data resources comprise case event space-time data, space-time background environment data and main body behavior data, and the acquisition module comprises: establishing a historical case event database and a suspect foothold database based on the case event time-space data, establishing a time-space environment database, a road network database and a street view database based on the time-space background environment data, and establishing a time-space track database and a real-time police resource distribution database based on the main body behavior data;
the data management module is used for performing data management on the basis of a road section to be predicted, the suspect foothold database, the time-space environment database, the road network database, the street view database, the time-space track database, a road network calculation engine, a graph calculation engine and a receptive field screener to obtain output data, wherein the output data comprises: crime opportunities from potential foot-falling points of all criminals in a receptive field of a road section to be predicted to the road section to be predicted, social structure characteristics of the criminals, direction entropies of all road sections in the receptive field, cognitive feature vectors of the road section to be predicted, and road network structure characteristics of the road section to be predicted in the receptive field;
the future case probability estimation module is used for estimating the case probability of the road section to be predicted for one week in the future based on a CTR estimation model of deep learning, the output data and the accumulated risk value of the road section to be predicted, wherein the accumulated risk value of the road section to be predicted is obtained by calculation based on crime opportunities from potential foothold of all criminals in the receptive field of the road section to be predicted to the road section to be predicted;
the prediction probability regularization module is used for regularizing the probability of a future one-week occurrence of the road section to be predicted, the receptive field screener, the jurisdiction screener, the hotspot migration rule base and the historical case event database to obtain the corrected probability of the future one-week occurrence of the road section to be predicted;
and the scheduling module is used for scheduling police force and planning an itinerant route based on the latest position information of the police condition, the receptive field screener, the probability of the future one-week case of all road sections in each jurisdiction, the road network database, the road network calculation engine and the real-time police force resource distribution database.
The invention has the beneficial effects that: the method is based on a historical case event database, a suspect foothold database, a space-time environment database, a road network database and a street view database, a space-time track database and a real-time police resource distribution database, and can obtain crime opportunities from potential footholds of all criminals in a receptive field of a road section to be predicted to the road section to be predicted, social structural characteristics of the criminals, direction entropies of all road sections in the receptive field, cognitive characteristic vectors of the road section to be predicted and road network structural characteristics of the road section to be predicted in the receptive field; based on the CTR estimation model of deep learning and the data, the probability of a future one-week occurrence of the road section to be predicted can be estimated, and the probability of a future one-week occurrence of the road section to be predicted can be normalized to obtain the corrected probability of a future one-week occurrence of the road section to be predicted. Meanwhile, police dispatching and patrol route planning can be realized. Compared with a common crime spatio-temporal prediction model based on grids, the method has the advantages that the road network is taken as an analysis carrier road section as an analysis unit, the problem of data sparsity is relieved, the problem of selection of grid division scale size is solved, and spatial heterogeneity is effectively adapted. Meanwhile, in order to facilitate the execution scheduling of the prediction result, the patrol route can be corrected in real time according to the prediction result, and the patrol coverage rate and patrol frequency of the high-risk road section are improved.
Drawings
Fig. 1 is a schematic flowchart illustrating a crime spatiotemporal risk prediction and decision scheduling method according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of a crime spatiotemporal risk prediction and decision scheduling system provided in an embodiment of the present invention;
FIG. 3 is a schematic diagram illustrating relationships between pictures of a street view mapping module according to an embodiment of the present disclosure;
FIG. 4 is a schematic view of a scene semantic tree of the street view mapping module provided in the embodiment of the present invention;
FIG. 5 is a schematic diagram of model training of a microfeature fusion module provided in an embodiment of the invention;
fig. 6 is a schematic structural diagram of a crime spatiotemporal risk prediction and decision scheduling apparatus according to an embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the drawings and the detailed description.
As shown in fig. 1, a method for criminal spatiotemporal risk prediction and decision scheduling includes the following steps:
s100, acquiring data resources and establishing a corresponding database, wherein the data resources comprise case event space-time data, space-time background environment data and main body behavior data, and the data resources comprise: a historical case event database and a suspect foothold database are established based on case event time-space data, a time-space environment database, a road network database and a street view database are established based on time-space background environment data, and a time-space track database and a real-time police force resource distribution database are established based on main body behavior data.
And for the case event space-time data, storing the space-time position data in the case event space-time data into a geographic database, and storing other attribute information into a relational database. The space-time environment database and the road network database belong to structured data, the space-time environment database comprises POI (point of interest) and other space environment data, the road network database comprises OSM and other road network data, the space position data in the structured data is stored in the geographic database, and other attribute information is stored in the relational database. The street view database belongs to unstructured data, spatial position data in the street view database are stored in a geographic database, images are directly stored in a file system, and the images can be vectorized and then stored in a Milvus isovector retrieval engine, so that later retrieval is facilitated. And storing the subject behavior data into a time sequence database.
It should be noted that the data resources may be obtained through an online public channel, or may be obtained through other legal channels.
S200, based on the road section to be predicted, the suspect foothold database, the time-space environment database, the road network database, the street view database, the time-space track database, the road network calculation engine, the graph calculation engine and the receptive field screener, performing data management to obtain output data, wherein the output data comprises: the method comprises the steps of obtaining crime opportunities from potential foot-dropping points of all criminals in the receptive field of a road section to be predicted to the road section to be predicted, social structure characteristics of the criminals, direction entropies of all road sections in the receptive field, cognitive feature vectors of the road section to be predicted and road network structure characteristics of the road section to be predicted in the receptive field.
After the data resources in S100 are obtained, data management needs to be performed on the data resources, which specifically includes preprocessing and performing feature engineering on the data by integrating multiple models, and finally obtaining crime opportunities from potential footfall points of all crimes in the receptive field of the road section to be predicted to the road section, social structure features of the crimes, direction entropies of all road sections in the receptive field, cognitive feature vectors of the road section, and road network structure features of the road section to be predicted in the receptive field.
As shown in fig. 2, for the receptive field screener 1, the receptive fields are a certain range that is continuous and correlated spatiotemporally at the current spatiotemporal location to be processed. The receptive field screener 1 takes the geographic coordinates to be inquired as input, calculates 30-minute walking time circles with the geographic coordinates as a starting point to be used as a data screener, and acquires data in the data screener to be used as a return value. The receptive field of the road section to be predicted refers to selecting a section of road section as the road section to be predicted, inputting the geographic coordinates of the road section to be predicted into the receptive field screener 1, and obtaining the receptive field of the road section to be predicted.
In S200, obtaining crime opportunities from potential foothold of all criminals in the receptive field of the road section to be predicted to the road section to be predicted includes:
the method comprises the steps of obtaining a road section jump effective distance map with space-time risks based on a unified opportunity travel model and a map calculation engine, obtaining potential footfall points of all criminals from a suspect footfall point database in a receptive field obtained based on a road section to be predicted and a receptive field screener, and calculating travel probability from the potential footfall points to the road section to be predicted, namely crime opportunities, based on the road section jump effective distance map.
Road networks are one of the carriers of space-time trajectories, but the accessibility of places in space-time cannot be directly reflected by geographical distances. The crime mode theory of environmental crime refers to the fact that the distribution of people flow in space and time can cause the generation and disappearance of crime attraction places and crime generation places, and the distribution of people flow in space and time cannot be directly reflected by geographic distance under the current traffic convenient scene. Crime opportunities of criminals are closely related to the spatial-temporal pedestrian flow distribution, and a unified opportunity travel model (intervention opportunity calculation module 15) can be better modeled by converting the geographical distance into the effective distance through the distance conversion based on the flow. Fig. 2 also includes an effective transfer distance conversion module 11, a road network matching module 12 and a spatiotemporal trajectory database 10. And calling a space-time trajectory database 10 to obtain a space-time trajectory, obtaining road sections corresponding to the trajectory through a road network matching module 12, and supplementing missing road sections according to an actual road network. And (3) dividing the space-time trajectory into 6 time slices with 4-hour span from 2 hours to 2 hours on the next day every day, and calculating the frequency of the inter-path jump in each slice. Calculating The effective distance between road sections in The road network by using The effective distance in The network, wherein The effective distance is defined by Brockmann in The high geometry of complex, network-drive contact mapping:
dmn＝(1-logPmn)≥1
wherein P ismnThe ratio of m to n to m total flow, i.e. the probability of a hop from node m to node n. Storing the effective distances as the directed weight of the relation between the road sections according to different time slices, and outputting the road section jump effective distances of different time slicesAnd (5) leaving the figure. The effective distance transfer module 11 weights the space-space distance according to the aggregation degree of the crowd in space-time, so that the weighted distance of the area with relatively large space-time aggregation degree is smaller than that of the area with small aggregation degree in the same geographic distance, and the reachability of the crime risk area is better simulated.
As shown in fig. 2, in order to implement the intervention opportunity calculation module 15, it is necessary to use the link jump effective distance map 13, the map calculation engine 14, the receptive field screener 1, and the suspect foothold database 16. Specifically, the intervention opportunity calculation module 15 obtains a road section jump effective distance map with space-time risks through a map calculation engine by using a unified opportunity travel model, obtains potential foothold data from a suspect foothold database within a receptive field range obtained based on a receptive field screener, and calculates a travel probability from the potential foothold to a road section to be predicted, namely a crime opportunity. The unified opportunistic travel model adopts the following models:
Figure BDA0002907603040000101
wherein m isiAs a starting opportunity value, mjFor the endpoint chance values, α and β are hyperparameters (0 ≦ α + β ≦ 1) to measure exploratory and cautious trends, s, respectively, of the subjectijIs the intervention opportunity value. For a suspect with only 1 case, α ═ β ═ 0.5; for suspects with a crime greater than 1, α ═ n/(n + m), β ═ m/(n + m); n is the number of plans outside the jurisdiction of the foothold, and m is the number of plans inside the jurisdiction of the foothold.
In S200, obtaining social structure characteristics of all criminals in a receptive field of a road section to be predicted, wherein the social structure characteristics comprise:
and constructing a potential social relationship based on the crime relations, the household registration relations and the space approach relations of all the criminals in the receptive field of the road section to be predicted, and acquiring the social structure characteristics of the criminals by using a node2vec model based on the potential social relationship.
In S200, obtaining road network structure characteristics of the road segment to be predicted in the receptive field includes:
screening corresponding road network data based on a road network database in a receptive field obtained based on a road section to be predicted and the receptive field screener, carrying out graph embedding operation on the screened road network data through a graph calculation engine, and obtaining road network structure characteristics of the road section to be predicted in the receptive field by using a node2vec model.
The structural feature calculation module 7 in fig. 2 is mainly divided into two blocks, one block is used for obtaining the social network structural features of criminals, and the other block is used for obtaining the structural features of road sections to be predicted in a road network. For a road network, calling a receptive field screener 1, carrying out graph embedding operation on the screened road network, and acquiring a feature vector of a road section to be predicted by using a node2vec model; for the social network structure characteristics, a potential social relationship (the crime relationship is a partnership relationship, the household relationship is the same in source province, and the space proximity relationship is the space proximity of the foothold) needs to be constructed from the crime relationship, the household relationship and the space proximity relationship, the social network structure characteristics of criminals are obtained by using the node2vec model, and the social network structure characteristics of the criminals corresponding to the foothold within the receptive field range are returned.
In S200, obtaining the cognitive feature vector of the road segment to be predicted includes:
s200a, carrying out panoramic segmentation on street view data of the road section to be predicted based on a street view database and a pre-training deep learning model, obtaining a street view map corresponding to the road section to be predicted based on an entity and a scene semantic tree obtained by segmentation, taking the street view map corresponding to the road section to be predicted as an objective cognitive representation of a scene, and taking a six-dimensional cognitive vector as a subjective cognitive representation of the scene.
The method comprises the steps of carrying out mapping processing by using an open data set of a Place Pulse project to obtain a corresponding street view map, and training the corresponding street view map by using a graph volume network to form a six-dimensional cognitive vector.
Firstly, the street view mapping module 19 in fig. 2 performs panoramic segmentation on street view data of a corresponding road section collected in the street view database 18 by using a pre-trained deep learning model, and then forms a map from entities obtained by the segmentation according to a scene semantic tree.
Specifically, the scene semantic tree is composed of objects commonly found in street views, and is divided according to the states, attributes and types of the objects. The invention adopts a panoramic segmentation model pre-trained by a COCO data set to process images, a scene semantic tree of the model is shown in figure 4, wherein, the numbers 0-2 of leaf nodes respectively represent the space positions (the ground, the ground but not separated from the ground and the sky) where the object frequently appears, in the same space position, the static objects are in an incidence relation, and the dynamic objects are in a dependency relation (weak in the incidence relation); between adjacent spatial positions, the object is in a dependency relationship; if both the adjacent objects in the image are static objects, the relationship is related, and if any one of the objects is a dynamic object, the relationship is dependent. As shown in fig. 3, if streetscapes of multiple angles are collected at the same spatial position, the relationship of objects between pictures follows the principle of "no cross-layer" processing, that is, the relationship is established between objects at the same spatial position, the relationship is not established between objects at different spatial positions, in the same layer, the father nodes of the objects of two different images in the scene semantic tree are the same, if both the objects are static objects, the association relationship is established, and if there is a dynamic object in both the objects, the dependency relationship is established; if the grandfather nodes of the objects of the two different images are the same, the association relationship is established between the static objects, and the establishment dependency relationship of the dynamic objects exists.
Then, the same mapping process is carried out by using an open data set of the Place Pulse project, and a street view map is trained by using a graph convolution network to form a six-dimensional cognitive vector. And transferring the trained model to a system to form a cognitive vector for the street view atlas of each street view.
And finally, taking the street view map of the road section to be predicted as an objective cognitive representation of the scene, and taking the six-dimensional cognitive vector as a subjective cognitive representation of the scene.
S200b, sampling corresponding road network data to form a data set based on a road section to be predicted and a road network database, when the data set is trained, for each group of objective cognitive features and subjective cognitive features, enabling the objective cognitive features to sequentially pass through a first graph neural network layer, a graph pooling layer, a second graph neural network layer and a graph reading layer to obtain corresponding objective cognitive feature vectors, sequentially performing splicing operation and linear regression prediction on the corresponding objective cognitive feature vectors and the subjective cognitive feature vectors corresponding to the subjective cognitive features to obtain road section risk values, and finally outputting the cognitive feature vectors of the road section to be predicted.
As shown in fig. 5, the microscopic cognitive feature fusion module 20 in fig. 2 trains and fuses to generate a cognitive feature vector of the road segment to be predicted. The method comprises the steps of firstly sampling a road network to form a data set, passing objective cognitive characterization through a graph neural network layer and a graph pooling layer for each group of objective cognitive characteristics and subjective cognitive characteristics in the process of training the data set, performing readout operation (reading), performing splicing operation on the readout characteristic vectors and the subjective cognitive vectors, and then performing linear regression to predict a road segment risk value. And for the condition that one road section has one street view acquisition point, all subjective feature vectors are directly summed, and the objective features are summed after readout. For the road sections without streetscapes but sampled into the data set, the subjective and objective feature vectors are initialized randomly and spliced directly to perform linear regression prediction, namely only the parameters of the linear regression part are updated during training. And finally outputting the subjective and objective characteristic vectors of the road section to be predicted after the training is finished.
The double-exponential trend fitting module 3 is further included in fig. 2, and the double-exponential trend fitting module 3 is configured to perform trend fitting on the historical case events belonging to the road segment according to a time sequence by using a double-exponential model, and return to the next stage of case issuing opportunity. The double-exponential model adopts a collective memory decay model proposed by Cristian Candia et al in Nature 2018, and is defined as follows:
Figure BDA0002907603040000131
wherein, N is an initial opportunity value and is set to be 1; p, q, r are attenuation parameters and can be obtained by fitting data in the historical case database 2. The cases belonging to each road section are arranged according to a time sequence, the number of the cases is subjected to sliding average by using a time window with a certain length (default is one week), the time when the maximum sliding average value in the last month is t is 0 is taken, and the subsequent opportunity value is the ratio of the sliding average value of each time to the sliding average value of the initial time. For the case of the current time, i.e. the maximum value, the next stage of the trial will be defaulted to 50%.
In consideration of the near-repetitive characteristic of criminal activities, the first-order road segment propagation module 4 in fig. 2 needs to call the graph calculation engine 14 to obtain corresponding data from the road segment jump effective distance graph 13, calculate the risk of neighboring road segment spread of the road segment to be predicted in the road network, and define the propagation range as the first-order neighbors of the road segment already developed in the last month, and the opportunity of developing the neighbors is half of the opportunity of developing the current road segment. The remaining road sections with no recorded issue will be set to 20%.
Since the risk of crime in space-time is related to spatial heterogeneity, the characterization of heterogeneity can be calculated from the POI, time-space environment data. According to historical issue data statistics, specific type space-time environment data corresponding to the road section where the issue is located can be obtained, and accordingly, specific type space-time data meeting conditions are screened out. And then converting the case-making opportunities of the specific type of space-time data and the road sections into feature vectors by using feature engineering, marking the road sections with case occurrence in one month as 1, otherwise marking the road sections with case occurrence as 0, and adopting the case-making opportunities predicted by a Bayesian regression model as opportunity values of the road sections with no case-making opportunities after the calculation of the road section propagation module. Due to the characteristic that the spatio-temporal environment data cannot be changed in a short period, the Bayesian regression model adopts an offline updating mode, and parameters do not need to be updated in real time.
Besides, the direction entropy calculation module 17 in fig. 2 is configured to calculate the direction entropy of all road segments within the receptive field of the road segment to be predicted. The directional entropy is an index for measuring the degree of Network misordering in an area, and is published by Geoff Boeing in Applied Network Science in 2019, and is defined as follows:
Figure BDA0002907603040000132
wherein P (o)i) Is the ratio of the total number of the road sections in the area occupied by the road section in the ith direction (the due north direction is set to be 0 degrees, the initial position is 5 degrees, and 360 degrees are divided into 36 equal parts at an interval of every 10 degrees), the invention adopts the direction entropyI.e. taking into account the ratio of the length of the ith direction to the total length of the road segments in the area. The area here uses receptive field segmentation.
S300, predicting the probability of a future one week of the road section to be predicted based on the CTR prediction model of deep learning, output data and the accumulated risk value of the road section to be predicted, wherein the accumulated risk value of the road section to be predicted is calculated based on crime opportunities from potential foothold of all criminals in the receptive field of the road section to be predicted to the road section to be predicted.
The future issue probability prediction module 21 in fig. 2 is different from the conventional neural network prediction model using matrix as input, and here, a deep click rate prediction model (CTR prediction model) based on neural network improvement is used for future issue probability prediction. The outputs of the structural feature calculation module 7, the intervention opportunity calculation module 15, the direction entropy calculation module 17 and the microscopic cognitive feature fusion module 20 are used as the inputs of the click rate estimation model, and the final output result is the probability of the future one week of the road section to be predicted. The CTR prediction model is originally applied to an advertisement click rate prediction task in an E-commerce scene in the field of a recommendation system, and models the probability of crime of a potential criminal from a footfall point to a road section to be predicted within a certain range by the aid of the method. The present invention adopts 30 min walking time circle as the screening condition. For the definition of the foot drop points of potential criminals, the historical foot drop points of the prior personnel are selected as high-risk foot drop points, and the positions of residential areas such as residential areas, villages and hotels in equal time circles are selected as the potential-risk foot drop points in consideration of the problems of changes of external population and the foot drop points. And (4) obtaining the traveling probability from each landing point to the road section to be predicted through an intervention opportunity calculation module, and then calculating the cumulative probability in a weighting mode. The method comprises the steps of firstly giving a weight to a high-risk foot-drop point of 1, giving a weight to a potential-risk foot-drop point of 0.1, normalizing all foot-drop point weights in a peer-to-peer time circle to a decimal between 0 and 1, then multiplying the corresponding trip probability by each foot-drop point weight to calculate an accumulated risk value of a road section to be predicted, wherein the accumulated risk value is a dynamically changed value because the opportunity value changes along with the updating of case and event data. The same weighting operation is also adopted for the social structure feature vector of the predecessor with the historical foothold within the range of the isochronous ring (the weight is the same as above, and the feature vector of the foothold with the potential risk is the 0 vector). And finally, inputting the CTR estimation model into an accumulated risk value, the summed social structure characteristic vector of the predecessor, the structure characteristic vector of the road section to be estimated, the direction entropy of the region where the road section to be estimated is located and the subjective and objective characteristic vector of the road section to be estimated.
In recent years, the field of recommendation systems is rapidly developed, the types of CTR prediction models are various, and common models include CCPM, PNN, Wide & Deep, Deep FM, MLR, NFM and the like. If the timing needs to be considered comprehensively, DIN, DIEN, etc. can be selected as the model.
S400, regularization is carried out on the basis of the future one-week-occurrence probability of the road section to be predicted, the receptive field screener, the jurisdiction screener, the hotspot migration rule base and the historical case event database, and the corrected future one-week-occurrence probability of the road section to be predicted is obtained.
S400 includes: the method comprises the steps of merging a total range of a district where a road section to be predicted is located and adjacent districts of the road section to be predicted, which is screened out by a district screener, with a receptive field range screened out by the receptive field screener to obtain a merged range, obtaining all crime high-incidence road sections in the merged range based on a historical case event database, calculating transfer probabilities of all crime high-incidence road sections to be potentially transferred to the road section to be predicted based on all the crime high-incidence road sections in the merged range and hotspot transfer rules in a hotspot transfer rule base, taking the maximum value of all the transfer probabilities, and correcting the case distribution probability of the road section to be predicted in the future one week through a Bayesian formula.
As shown in fig. 2, the prediction probability regularization module 22 is implemented by using a jurisdiction filter 24, a receptive field filter 1, a hotspot migration rule base 23 and a historical case event database 2. Specifically, after the future case probability of the road section to be predicted is obtained through the future case probability estimation module 21, the probability needs to be regularized through the estimation probability regularization module 22, and the total range of the area in which the road section to be predicted is located and the adjacent area thereof is screened out through the area screener 24 and is merged with the range screened out by the receptive field screener 1. All crime high-incidence road sections in the range are obtained through the historical case event database 2, the risk of potential transfer to the road section is calculated through the hot spot transfer rules in the hot spot transfer rule base 23, the maximum value of all transfer probabilities is taken, and the future one-week incidence probability of the current road section is corrected through a Bayesian formula.
Through S100-S400, the corrected probability of the future one week of the road section to be predicted can be obtained. Through the method of S100-S400, the future one-week occurrence probability of all road sections in each jurisdiction can be obtained. Police dispatching and patrol route planning can be realized based on the probability of occurrence of a future week of all road sections in each jurisdiction.
S500, carrying out police dispatching and routing planning based on the latest police position information, the receptive field screener, the future one-week occurrence probability of all road sections in each jurisdiction, the road network database, the road network computing engine and the real-time police resource distribution database.
S500 comprises:
s500a, when the latest alarm condition appears, based on the occurrence position information and the receptive field screener of the latest alarm condition, a predetermined time circle is obtained by taking a street corresponding to the latest alarm condition as a starting point, the predetermined time circle and the region range to which the latest alarm condition belongs are intersected to obtain an intersection range, all on-duty patrol control alarm forces in the intersection range are obtained based on a real-time alarm force resource distribution database, patrol control schemes with the states not being the highest priority are screened out from the current patrol control schemes and added into a scheduling candidate list, road network data in a calculation range in a road network database are called through a road network calculation engine, the patrol control scheme with the shortest time consumption is screened out from the scheduling candidate list, the state of the patrol control scheme is set as the highest priority, and corresponding patrol control route information is obtained.
S500b, when the latest warning situation does not appear, calculating to obtain the patrol control schemes of which the police force respectively accords with the patrol control targets with the maximum patrol control range, the maximum accumulative patrol control scheme issuing area, the maximum risk and the minimum patrol control coincidence degree among the police forces within a certain time by using the position of the police force dispatched by each jurisdiction as the starting point through a road network calculation engine based on the scheme issuing probability of one week in the future of all road sections in each jurisdiction, and obtaining the patrol control route information corresponding to each patrol control scheme.
As shown in fig. 2, in order to implement the police dispatch module 25, a real-time police resource distribution database 25, a latest position monitoring module 26, a road network calculation engine 6 and a road network database 5 are required. Specifically, the police dispatch module 25 divides the next phase of the route planning scheme into three levels: the method comprises a first level, a second level and a third level, wherein the first level is the highest priority. When the latest warning situation generating position is input, the receptive field screening module 1 is called to obtain a time circle of 10 minutes of vehicle running and the like by taking a street corresponding to the warning situation as a starting point and to intersect with a region range to which the warning situation belongs, all on-duty patrol control warning powers in the region are obtained from the real-time warning power resource distribution database 25, a scheduling candidate list in which the current patrol control scheme state is not the first level is screened out, road network data in a calculation range in the road network database 5 is called through the road network calculation engine 6, the warning power which takes the shortest time is screened out from the candidate list, the state is set as the first level, and navigation line information is returned. And for the condition that the system has no latest warning condition, acquiring the future one-week case issuing probability of all road sections in each jurisdiction, calculating three patrol targets with the largest patrol range, the largest accumulative patrol case issuing area and risk and the lowest patrol control coincidence degree among police forces in a certain time from the position where the district is dispatched by the road network calculation engine 6, and returning to the patrol control route plans corresponding to the patrol control targets.
In the prior art, the crime mode is mainly identified and predicted by using a grid-based reasoning mode, and the rationality of the prediction mode and the interpretability of a prediction result cannot be fully ensured. The method integrates data in various forms and scales such as images, texts, tables and the like, and adopts a road network-based reasoning mode to identify and predict the crime mode. Compared with a common crime space-time prediction model based on grids, the method has the advantages that the road network is used as an analysis carrier road section as an analysis unit, the problem of data sparsity is relieved, meanwhile, the problem of selection of grid division size is solved, and the spatial heterogeneity is effectively adapted.
Aiming at the risk estimation problem of zero sample space-time, the method disclosed by the invention jointly estimates the potential space-time risk probability of the zero sample area by utilizing the environmental element statistical indexes in a local range, the crime data statistical indexes of the issued area and the existing risk hotspot potential migration information in a certain range based on a logistic regression method.
Aiming at the modeling problems of crime space-time opportunities and crime travel based on the space-time opportunities, the invention utilizes the crime generation mechanism in the theory of the road network data modeling crime mode and utilizes environmental element data such as POI and the like to combine with the number of issued cases to represent the crime opportunities based on the effective distance. And on the basis, a unified opportunity travel model is used for predicting the potential lines of the criminals for travel.
The street view data are applied to a crime prediction system, and a method for constructing the cognitive feature vector is provided to represent the cognition of a person on the current road section environment.
Considering the approaching repetition phenomenon of criminal activities and the mode characteristics of criminal travel, the invention provides a trend fitting model based on a double-exponential model to fit the change trend of the space-time risk of the published case area. And reducing the calculated amount of the system while restoring the theoretical model based on the receptive field analysis method.
Meanwhile, the prediction model part adopted by the invention uses the space-time characteristic vector with the context information for input, so that the problem of input dimension change caused by prediction range change is avoided, and the method has stronger mobility.
In addition, the invention adds a prediction probability regular module to assist the decision of the user, and corrects the prediction result based on the hot spot migration rule of the crime pattern theory in the environmental crime theory, thereby improving the interpretability of the prediction result and increasing the flexibility of deviation correction.
Finally, in order to facilitate the execution scheduling of the prediction result, the invention provides the police force scheduling module which corrects the patrol route in real time according to the prediction result and improves the patrol coverage rate and patrol frequency of the high-risk road section.
As shown in fig. 6, a criminal spatiotemporal risk prediction and decision scheduling apparatus includes:
the obtaining module 601 is configured to obtain data resources and establish a corresponding database, where the data resources include case and event spatio-temporal data, spatio-temporal background environment data, and body behavior data, and include: a historical case event database and a suspect foothold database are established based on case event time-space data, a time-space environment database, a road network database and a street view database are established based on time-space background environment data, and a time-space track database and a real-time police force resource distribution database are established based on main body behavior data.
A data management module 602, configured to perform data management based on a road segment to be predicted, a suspect foothold database, a time-space environment database, a road network database, a street view database, a time-space trajectory database, a road network calculation engine, a graph calculation engine, and a receptive field filter, to obtain output data, where the output data includes: the method comprises the steps of obtaining crime opportunities from potential foot-dropping points of all criminals in the receptive field of a road section to be predicted to the road section to be predicted, social structure characteristics of the criminals, direction entropies of all road sections in the receptive field, cognitive feature vectors of the road section to be predicted and road network structure characteristics of the road section to be predicted in the receptive field.
The future occurrence probability estimation module 603 is configured to estimate occurrence probability of a future week of the road segment to be predicted based on the CTR estimation model of the deep learning, the output data, and the accumulated risk value of the road segment to be predicted, where the accumulated risk value of the road segment to be predicted is calculated based on crime opportunities from potential foothold of all criminals in the reception field of the road segment to be predicted to the road segment to be predicted.
And the estimated probability regularization module 604 is used for regularizing the probability of a future one-week occurrence of the road section to be predicted, the receptive field screener, the jurisdiction screener, the hotspot migration rule base and the historical case event database to obtain the corrected probability of the future one-week occurrence of the road section to be predicted.
And the scheduling module 605 is configured to perform police dispatching and patrol route planning based on the latest position information of the police situation, the receptive field screener, the probability of occurrence of a future one week of all road segments in each jurisdiction, the road network database, the road network calculation engine, and the real-time police resource distribution database.
In this embodiment, a criminal spatiotemporal risk prediction and decision scheduling apparatus may be a computer, and includes: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the aforementioned crime spatiotemporal risk prediction and decision scheduling methods via execution of the executable instructions. The memory and the processor may be connected by a bus. The memory unit may include a readable medium in the form of a volatile memory unit, such as a random access memory unit (RAM) and/or a cache memory unit, and may further include a read only memory unit (ROM). The computer also includes a display unit connected to the bus. The display unit can display the potentially important information of the patient and the like.
A computer-readable storage medium, on which a computer program is stored, which, when executed by the above-mentioned processor, implements the above-mentioned crime spatiotemporal risk prediction and decision scheduling method.
It should be noted that, the crime spatiotemporal risk prediction and decision scheduling apparatus and method in this embodiment are the same inventive concept, and the functions of the crime spatiotemporal risk prediction and decision scheduling apparatus can be described in detail in the embodiment of the crime spatiotemporal risk prediction and decision scheduling method.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is intended to include such modifications and variations.

Claims (10)

1. A crime spatiotemporal risk prediction and decision scheduling method is characterized by comprising the following steps:
s100, acquiring data resources and establishing a corresponding database, wherein the data resources comprise case event space-time data, space-time background environment data and main body behavior data, and the data resources comprise: establishing a historical case event database and a suspect foothold database based on the case event time-space data, establishing a time-space environment database, a road network database and a street view database based on the time-space background environment data, and establishing a time-space track database and a real-time police resource distribution database based on the main body behavior data;
s200, based on the road section to be predicted, the suspect foothold database, the time-space environment database, the road network database, the street view database, the time-space trajectory database, a road network calculation engine, a graph calculation engine and a receptive field screener, performing data management to obtain output data, wherein the output data comprises: crime opportunities from potential foot-falling points of all criminals in a receptive field of a road section to be predicted to the road section to be predicted, social structure characteristics of the criminals, direction entropies of all road sections in the receptive field, cognitive feature vectors of the road section to be predicted, and road network structure characteristics of the road section to be predicted in the receptive field;
s300, predicting the probability of occurrence of a future week of the road section to be predicted based on a CTR prediction model for deep learning, the output data and the accumulated risk value of the road section to be predicted, wherein the accumulated risk value of the road section to be predicted is calculated based on crime opportunities from potential foothold of all criminals in the receptive field of the road section to be predicted to the road section to be predicted;
s400, regularizing the probability of the future one-week occurrence of the road section to be predicted, the receptive field screener, the jurisdiction screener, the hotspot migration rule base and the historical case event database to obtain the corrected probability of the future one-week occurrence of the road section to be predicted;
s500, carrying out police dispatching and routing planning based on the latest position information of the police, the receptive field screener, the probability of the future one-week case of all road sections in each jurisdiction, the road network database, the road network calculation engine and the real-time police force resource distribution database.
2. The method of claim 1, wherein, in S100,
for the case event space-time data, storing the space-time position data in the case event space-time data into a geographic database, and storing other attribute information into a relational database;
for the space-time environment database and the road network database, storing the space position data in the space-time environment database and the road network database into a geographic database, and storing other attribute information into a relational database;
for the street view database, storing spatial position data in the street view database into a geographic database, storing images into a file system, and storing vectorized images into a vector retrieval engine;
and storing the subject behavior data into a time sequence database.
3. The method of claim 1, wherein the step S200 of obtaining the crime opportunities from the potential foot points of all criminals in the receptive field of the road section to be predicted to the road section to be predicted comprises the following steps:
the method comprises the steps of obtaining a road section jump effective distance map with space-time risks based on a unified opportunity travel model and a map calculation engine, obtaining potential footfall points of all criminals from a suspect footfall point database in a receptive field obtained based on a road section to be predicted and a receptive field screener, and calculating travel probability from the potential footfall points to the road section to be predicted, namely crime opportunities based on the road section jump effective distance map.
4. The method according to claim 1, wherein the step S200 of obtaining the social structure characteristics of all criminals in the receptive field of the road section to be predicted comprises the following steps:
and constructing a potential social relationship based on the crime relations, the household registration relations and the space approach relations of all the criminals in the receptive field of the road section to be predicted, and acquiring the social structure characteristics of the criminals by using a node2vec model based on the potential social relationship.
5. The method according to claim 1, wherein in S200, acquiring the road network structure characteristics of the road segment to be predicted in the receptive field comprises:
and screening corresponding road network data based on the road network database in the receptive field obtained based on the road section to be predicted and the receptive field screener, carrying out graph embedding operation on the screened road network data through the graph calculation engine, and obtaining the road network structure characteristics of the road section to be predicted in the receptive field by using a node2vec model.
6. The method according to claim 1, wherein in step S200, acquiring the cognitive feature vector of the road segment to be predicted comprises:
based on the street view database and a pre-training deep learning model, carrying out panoramic segmentation on street view data of the road section to be predicted, based on an entity and a scene semantic tree obtained by segmentation, obtaining a street view map corresponding to the road section to be predicted, taking the street view map corresponding to the road section to be predicted as an objective cognitive representation of a scene, and taking a six-dimensional cognitive vector as a subjective cognitive representation of the scene;
the method comprises the steps of sampling corresponding road network data to form a data set based on a road section to be predicted and a road network database, sequentially passing objective cognitive features of each group of objective cognitive features and subjective cognitive features through a first graph neural network layer, a graph pooling layer, a second graph neural network layer and a graph reading layer to obtain corresponding objective cognitive feature vectors when the data set is trained, sequentially performing splicing operation and linear regression prediction on the corresponding objective cognitive feature vectors and subjective cognitive feature vectors corresponding to the subjective cognitive features to obtain a road section risk value, and finally outputting the cognitive feature vectors of the road section to be predicted.
7. The method of claim 1, wherein S400 comprises:
and merging the total range of the district in which the road section to be predicted is located and the adjacent district thereof screened by the district screener with the receptive field range screened by the receptive field screener to obtain a merged range, acquiring all crime high-incidence road sections in the merged range based on the historical case event database, calculating the transfer probability of all crime high-incidence road sections to the road section to be predicted based on all crime high-incidence road sections in the merged range and the hotspot transfer rule in the hotspot transfer rule base, taking the maximum value of all transfer probabilities, and correcting the case distribution probability of the road section to be predicted for one week in the future by a Bayesian formula.
8. The method according to any one of claims 1-7, wherein S500 comprises:
when the latest alarm condition occurs, based on the occurrence position information of the latest alarm condition and the receptive field screener, a preset time circle is obtained by taking a street corresponding to the latest alarm condition as a starting point, the preset time circle and the area range to which the latest alarm condition belongs are intersected to obtain an intersection range, all on-duty patrol control alarm forces in the intersection range are obtained based on the real-time alarm force resource distribution database, patrol control schemes with the states not being the highest priority are screened out from the current patrol control schemes and added into a scheduling candidate list, the road network data in the calculation range in the road network database are called through the road network calculation engine, the patrol control scheme with the shortest time consumption is screened out from the scheduling candidate list, the states are set as the highest priority, and corresponding patrol control route information is obtained.
9. The method according to any one of claims 1-7, wherein S500 further comprises:
and when the latest warning condition does not appear, calculating to obtain the patrol control schemes of the police force respectively conforming to three patrol control targets with the maximum patrol control range, the maximum accumulated patrol control issuing area, the maximum risk and the minimum patrol control overlap ratio among the police forces in a certain time by taking the position of the police force dispatched by each jurisdiction as a starting point through the road network calculation engine based on the issue probability of all road sections in each jurisdiction in one week in the future, and obtaining the patrol control route information corresponding to each patrol control scheme.
10. A criminal spatiotemporal risk prediction and decision scheduling device is characterized by comprising:
the acquisition module is used for acquiring data resources and establishing a corresponding database, wherein the data resources comprise case event space-time data, space-time background environment data and main body behavior data, and the acquisition module comprises: establishing a historical case event database and a suspect foothold database based on the case event time-space data, establishing a time-space environment database, a road network database and a street view database based on the time-space background environment data, and establishing a time-space track database and a real-time police resource distribution database based on the main body behavior data;
the data management module is used for performing data management on the basis of a road section to be predicted, the suspect foothold database, the time-space environment database, the road network database, the street view database, the time-space track database, a road network calculation engine, a graph calculation engine and a receptive field screener to obtain output data, wherein the output data comprises: crime opportunities from potential foot-falling points of all criminals in a receptive field of a road section to be predicted to the road section to be predicted, social structure characteristics of the criminals, direction entropies of all road sections in the receptive field, cognitive feature vectors of the road section to be predicted, and road network structure characteristics of the road section to be predicted in the receptive field;
the future case probability estimation module is used for estimating the case probability of the road section to be predicted for one week in the future based on a CTR estimation model of deep learning, the output data and the accumulated risk value of the road section to be predicted, wherein the accumulated risk value of the road section to be predicted is obtained by calculation based on crime opportunities from potential foothold of all criminals in the receptive field of the road section to be predicted to the road section to be predicted;
the prediction probability regularization module is used for regularizing the probability of a future one-week occurrence of the road section to be predicted, the receptive field screener, the jurisdiction screener, the hotspot migration rule base and the historical case event database to obtain the corrected probability of the future one-week occurrence of the road section to be predicted;
and the scheduling module is used for scheduling police force and planning an itinerant route based on the latest position information of the police condition, the receptive field screener, the probability of the future one-week case of all road sections in each jurisdiction, the road network database, the road network calculation engine and the real-time police force resource distribution database.
CN202110076296.5A 2021-01-20 2021-01-20 Criminal spatiotemporal risk prediction and decision scheduling method and device Pending CN114169659A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110076296.5A CN114169659A (en) 2021-01-20 2021-01-20 Criminal spatiotemporal risk prediction and decision scheduling method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110076296.5A CN114169659A (en) 2021-01-20 2021-01-20 Criminal spatiotemporal risk prediction and decision scheduling method and device

Publications (1)

Publication Number Publication Date
CN114169659A true CN114169659A (en) 2022-03-11

Family

ID=80476351

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110076296.5A Pending CN114169659A (en) 2021-01-20 2021-01-20 Criminal spatiotemporal risk prediction and decision scheduling method and device

Country Status (1)

Country Link
CN (1) CN114169659A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115114753A (en) * 2022-05-30 2022-09-27 中铁二院工程集团有限责任公司 Intelligent design method for railway line plane in complex and difficult mountain area
CN116433051A (en) * 2023-06-09 2023-07-14 中国人民公安大学 Urban area police strategy dynamic adjustment method and system
CN115114753B (en) * 2022-05-30 2024-11-05 中铁二院工程集团有限责任公司 Intelligent design method for plane of railway line in complicated difficult mountain area

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115114753A (en) * 2022-05-30 2022-09-27 中铁二院工程集团有限责任公司 Intelligent design method for railway line plane in complex and difficult mountain area
CN115114753B (en) * 2022-05-30 2024-11-05 中铁二院工程集团有限责任公司 Intelligent design method for plane of railway line in complicated difficult mountain area
CN116433051A (en) * 2023-06-09 2023-07-14 中国人民公安大学 Urban area police strategy dynamic adjustment method and system
CN116433051B (en) * 2023-06-09 2023-08-18 中国人民公安大学 Urban area police strategy dynamic adjustment method and system

Similar Documents

Publication Publication Date Title
Tang et al. Multi-community passenger demand prediction at region level based on spatio-temporal graph convolutional network
Ali et al. Exploiting dynamic spatio-temporal graph convolutional neural networks for citywide traffic flows prediction
US11270579B2 (en) Transportation network speed foreeasting method using deep capsule networks with nested LSTM models
Phithakkitnukoon et al. Taxi-aware map: Identifying and predicting vacant taxis in the city
CN106488405B (en) A kind of position predicting method of fusion individual and neighbour's movement law
CN112200351A (en) Urban area passenger flow volume prediction method based on mobile phone signaling data
CN113313947A (en) Road condition evaluation method of short-term traffic prediction graph convolution network
Huang et al. A GAN framework-based dynamic multi-graph convolutional network for origin–destination-based ride-hailing demand prediction
Ma et al. Multi-attention graph neural networks for city-wide bus travel time estimation using limited data
Rajeh et al. Modeling multi-regional temporal correlation with gated recurrent unit and multiple linear regression for urban traffic flow prediction
Shao et al. Incorporating lstm auto-encoders in optimizations to solve parking officer patrolling problem
CN116307152A (en) Traffic prediction method for space-time interactive dynamic graph attention network
Turukmane et al. Multispectral image analysis for monitoring by IoT based wireless communication using secure locations protocol and classification by deep learning techniques
Nasiboglu Dijkstra solution algorithm considering fuzzy accessibility degree for patch optimization problem
CN113159371B (en) Unknown target feature modeling and demand prediction method based on cross-modal data fusion
CN118133084A (en) Method and device for predicting air quality of area without monitoring station based on hierarchical graph convolutional network
Han et al. Deep-AIR: A hybrid CNN-LSTM framework for air quality modeling in metropolitan cities
CN114169659A (en) Criminal spatiotemporal risk prediction and decision scheduling method and device
Ueda et al. Spatio-temporal multidimensional collective data analysis for providing comfortable living anytime and anywhere
CN114925994A (en) Urban village risk assessment and risk factor positioning method based on deep learning
CN112749209A (en) Method for constructing movement behavior map facing to space-time data
Chang et al. New travel time prediction algorithms for intelligent transportation systems
Geiß et al. Anticipating a risky future: LSTM models for spatiotemporal extrapolation of population data in areas prone to earthquakes and tsunamis in Lima, Peru
Huang et al. Deep ConvLSTM-inception network for traffic prediction in smart cities
Yuan et al. A novel learning approach for citywide crowd flow prediction

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination