A crime space-time risk prediction and decision scheduling method and device

Abstract

The application discloses a kind of criminal space-time risk prediction and decision scheduling method and device, comprising the following steps: S100, obtain data resources and establish corresponding database;S200, data governance is carried out, and output data is obtained;S300, the CTR estimation model based on deep learning and output data, estimate the probability of occurrence of the future week of the to-be-predicted section;S400, based on the probability of occurrence of the future week of the to-be-predicted section, receptive field filter, jurisdiction filter, hotspot migration rule base and historical case event database, regularization is carried out, and the probability of occurrence of the future week of the to-be-predicted section after correction is obtained;S500, police scheduling and patrol control route planning are carried out.The application can enhance the flexibility of the input of the prediction model, improve the migratability of the system, enhance the explainability of the prediction result, increase the flexibility of the prediction result correction by increasing the estimation probability correction module, and simplify the research and judgment of the prediction result and the scheduling process by increasing the scheduling module.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spatiotemporal information processing, in particular to a crime spatiotemporal risk prediction and decision-making scheduling method and device. BACKGROUND

[0002] The current mainstream crime spatiotemporal risk prediction system mainly includes two modules of data management and future crime probability estimation. The data management module processes the input multi-source heterogeneous data into a format that can be input into the probability estimation module. The future crime probability estimation module receives these processed data, performs online or offline training, and provides prediction services after the prediction model reaches the appropriate accuracy. The parameters of the online model can be updated in real time or in offline mode to adapt to data changes.

[0003] In the prior art, the mainstream technology of crime spatiotemporal prediction is to use a grid-based prediction model to predict the crime probability in a certain spatiotemporal future. Although this grid-based mode is simple to preprocess, it does not consider the heterogeneity of space. Compared with it, a more reasonable analysis carrier is road network data. Road network is one of the important carriers of spatiotemporal information transmission outside network space. Crime activities occur in space-time, and the movement of the perpetrators in space-time is directly related to the distribution of road network. Grids cannot accurately reflect the adjacency relationship in space.

[0004] The prediction scale of the existing crime spatiotemporal prediction system is relatively large, which is not conducive to actual decision-making command and scheduling. In order to prevent the problem of sparse data caused by too small grid, most prediction systems use a grid with a side length of tens of meters or even hundreds of meters as the prediction unit. This has the advantage of reducing the sparsity of input data and improving the prediction accuracy. However, the disadvantage is that the prediction range is too large, making it difficult to apply to specific patrol and control.

[0005] The existing crime spatiotemporal prediction system uses the number of cases as the prediction index to judge the risk level of an area. This index is intuitive, but it is not reasonable for zero-sample spatiotemporal locations because crime hotspots migrate, and this migration is related to spatiotemporal heterogeneity, but it is not necessarily adjacent in space and time. Therefore, for the risk prediction of zero-sample spatiotemporal locations, the existing crime spatiotemporal prediction system does not have a good solution, and even if the model's prediction value appears in the zero-sample area, it cannot give a reasonable explanation. The risk in space and time is actually a manifestation of the size of the crime opportunity. This number of cases is only a form of crime opportunity and cannot represent the size of the crime opportunity. For example, some areas may have few cases or even no historical records, but their environmental concealment and poor security conditions greatly increase the success rate of crime. Once the criminal and the potential target appear at the same time, there will be a high probability of crime. Therefore, in addition to environmental factors, the activities of criminals and potential targets also determine the occurrence of crime. For the mode of travel of criminals, there is a crime travel theory in environmental criminology, which describes that the distance of the criminal's travel to commit a crime presents an exponential decay trend within a certain range. This description is similar to the "gravity model" in the field of transportation. However, considering the influence of spatiotemporal opportunities on human travel in modern society, distance is no longer the only factor limiting human travel. For crime travel, the distribution of crime opportunities in space and time is also an important factor affecting the travel of criminals. In 2019, Yan Xiaoyong and Liu Erjian proposed a unified opportunity model for human travel, which models human travel behavior from the perspective of spatiotemporal opportunities. Compared with the gravity model (similar to the distance decay model in crime travel), this model has better prediction results and higher interpretability. Therefore, it is more reasonable to use spatiotemporal environmental information, potential travel routes of criminals, and human activity information to modify the simple number of cases than to use only the number of cases to represent the risk in space and time.

[0006] The existing crime spatiotemporal prediction system does not fully consider the spatial heterogeneity, such as the influence of the distribution of police stations on the crime situation, the influence of the crime hotspot transfer on the crime probability, and the influence of the human activities in space and time on the crime opportunity. The crime pattern theory of environmental criminology proposes three concepts of crime attraction area, crime generation area and crime neutral area, and believes that the collision probability of criminals and potential targets is the mechanism of crime generation. There is no reasonable quantitative model for human activities and crime opportunities in the existing crime prediction methods or systems. Dirk Brockmann and Dirk Helbing proposed an effective distance measurement method based on the flow size between space-time positions in Science in 2013. This method can effectively predict the source of infectious diseases. For environmental criminology, the effective distance measurement method can model the influence of human activities on crime opportunities in space and time. The distance from other areas to high-concentration areas will be relatively short, and the accessibility of areas with relatively small flow proportion will be relatively poor.

[0007] The existing crime spatiotemporal prediction system does not fully consider the spatial heterogeneity, such as the influence of the distribution of police stations on the crime situation, the influence of the crime hotspot transfer on the crime probability, and the influence of the human activities in space and time on the crime opportunity. The crime pattern theory of environmental criminology proposes three concepts of crime attraction area, crime generation area and crime neutral area, and believes that the collision probability of criminals and potential targets is the mechanism of crime generation. There is no reasonable quantitative model for human activities and crime opportunities in the existing crime prediction methods or systems. Dirk Brockmann and Dirk Helbing proposed an effective distance measurement method based on the flow size between space-time positions in Science in 2013. This method can effectively predict the source of infectious diseases. For environmental criminology, the effective distance measurement method can model the influence of human activities on crime opportunities in space and time. The distance from other areas to high-concentration areas will be relatively short, and the accessibility of areas with relatively small flow proportion will be relatively poor.

[0008] The existing prediction system has poor model transferability. For grid-based models, changes in input data dimensions caused by changes in predicted spatial range, or changes in input data sources caused by changes in predicted spatial position, may require the model to be redesigned and trained again.

[0009] The existing prediction system has no reasonable explainability for the prediction result, and cannot correct the prediction error caused by the bias of the prediction model. The existing prediction model is mostly a model 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 when the prediction result is obtained. Not only the mechanism behind the data cannot be explained, but also the wrong result caused by the bias of the model itself cannot be corrected. SUMMARY

[0010] In view of the defects in the prior art, the purpose of the present application is to provide a crime space-time risk prediction and decision scheduling method and device, which can enhance the flexibility of the prediction model input, improve the system migration, enhance the explainability of the prediction result, increase the estimation probability regular module to improve the flexibility of the prediction result correction, and increase the scheduling module to simplify the research and judgment of the prediction result and the scheduling process.

[0011] To achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0012] A crime space-time risk prediction and decision scheduling method, comprising the following steps:

[0013] S100, acquiring data resources and establishing a corresponding database, wherein the data resources include case event space-time data, space-time background environment data and subject behavior data, comprising: establishing a historical case event database and a suspect landing point database based on the case event space-time data, establishing a space-time environment database, a road network database and a street view database based on the space-time background environment data, and establishing a space-time trajectory database and a real-time police resource distribution database based on the subject behavior data;

[0014] S200, based on the to-be-predicted road section, the suspect landing point database, the space-time environment database, the road network database, the street view database and the space-time trajectory database, and the road network calculation engine, the graph calculation engine and the receptive field filter, data governance is performed to obtain output data, the output data includes: the crime opportunity of all potential landing points of criminals in the receptive field of the to-be-predicted road section to the to-be-predicted road section, the social structure characteristics of the criminals, the direction entropy of all road sections in the receptive field, and the cognitive feature vector of the to-be-predicted road section, and the road network structure characteristics of the to-be-predicted road section in the receptive field;

[0015] S300, based on the CTR estimation model of deep learning, the output data and the cumulative risk value of the to-be-predicted road section, estimating the case occurrence probability of the to-be-predicted road section in the future one week, wherein the cumulative risk value of the to-be-predicted road section is calculated based on the crime opportunity of all potential landing points of criminals in the receptive field of the to-be-predicted road section to the to-be-predicted road section;

[0016] S400, based on the case probability of the to-be-predicted road segment in the future one week, the receptive field filter, the jurisdiction filter, the hotspot migration rule library and the historical case database, regularization is performed to obtain a corrected case probability of the to-be-predicted road segment in the future one week.

[0017] S500, based on the occurrence location information of the latest police case, the receptive field filter, the case probability of all road segments in each jurisdiction in the future one week, the road network database, the road network computing engine and the real-time police resource distribution database, police scheduling and patrol route planning are performed.

[0018] Further, in the method described above, in S100,

[0019] For the spatio-temporal data of the case, the spatio-temporal location data is stored in the geographic database, and other attribute information is stored in the relational database.

[0020] For the spatio-temporal environment database and the road network database, the spatial location data is stored in the geographic database, and other attribute information is stored in the relational database.

[0021] For the street view database, the spatial location data is stored in the geographic database, the image is stored in the file system, and the vectorized image is stored in the vector retrieval engine.

[0022] For the subject behavior data, it is stored in the time series database.

[0023] Further, in the method described above, in S200, the crime opportunity of all potential landing points of criminals in the receptive field of the to-be-predicted road segment to the to-be-predicted road segment is obtained, including:

[0024] Based on the unified opportunity travel model and the graph computing engine, an effective distance graph of road segments with spatio-temporal risk is obtained, and based on the receptive field obtained based on the to-be-predicted road segment and the receptive field filter, all potential landing points of criminals in the receptive field are obtained from the suspect landing point database, and the travel probability from the potential landing point to the to-be-predicted road segment, i.e. the crime opportunity, is calculated based on the effective distance graph of the road segments.

[0025] Further, in the method described above, in S200, the social structure features of all criminals in the receptive field of the to-be-predicted road segment are obtained, including:

[0026] Based on the criminal relationship, the household relationship and the spatial proximity relationship of all criminals in the receptive field of the to-be-predicted road segment, a potential social relationship is constructed, and based on the potential social relationship, the social structure features of the criminals are obtained using a node2vec model.

[0027] Further, in the method as described above, in S200, the road network structure features of the to-be-predicted road section in the receptive field are obtained, including:

[0028] In the receptive field obtained based on the to-be-predicted road section and the receptive field filter, corresponding road network data is filtered out based on the road network database, graph embedding operation is performed on the filtered road network data by the graph computing engine, and the road network structure features of the to-be-predicted road section in the receptive field are obtained using a node2vec model.

[0029] Further, in the method as described above, in S200, the cognitive feature vector of the to-be-predicted road section is obtained, including:

[0030] Based on the street view database and the pre-trained deep learning model, panoramic segmentation is performed on the street view data of the to-be-predicted road section, the street view graph corresponding to the to-be-predicted road section is obtained based on the entity and scene semantic tree obtained by segmentation, the street view graph corresponding to the to-be-predicted road section is taken as an objective cognitive representation of the scene, and a six-dimensional cognitive vector is taken as a subjective cognitive representation of the scene.

[0031] Based on the to-be-predicted road section and the road network database, the corresponding road network data is sampled to form a data set, and when the data set is trained, for each set of objective cognitive features and subjective cognitive features, the objective cognitive features therein are sequentially subjected to a first graph neural network layer, a graph pooling layer, a second graph neural network layer, and a graph readout layer to obtain a corresponding objective cognitive feature vector, and the corresponding objective cognitive feature vector and the subjective cognitive feature vector corresponding to the subjective cognitive features therein are sequentially subjected to splicing operation and linear regression prediction to obtain a road section risk value, and finally the cognitive feature vector of the to-be-predicted road section is output.

[0032] Further, in the method as described above, S400 includes:

[0033] The total range of the to-be-predicted road section and its adjacent jurisdictions filtered out based on the jurisdiction filter is combined with the receptive field range filtered out based on the receptive field filter to obtain a combined range, all high-crime road sections in the combined range are obtained based on the historical case event database, the transfer probability of all high-crime road sections potentially transferred to the to-be-predicted road section is calculated based on all high-crime road sections in the combined range and the hotspot transfer rule in the hotspot transfer rule library, the maximum value in all transfer probabilities is taken, and the case occurrence probability of the to-be-predicted road section in the future one week is corrected through the Bayes formula.

[0034] Further, in the method as described above, S500 includes:

[0035] When a latest police case occurs, based on occurrence position information of the latest police case and the receptive field filter, a predetermined time circle is obtained starting from a street corresponding to the latest police case, an intersection range is obtained by intersecting the predetermined time circle and a jurisdiction range to which the latest police case belongs, all on-duty patrol police forces in the intersection range are obtained based on the real-time police force resource distribution database, a patrol scheme whose state is not the highest priority is filtered out from a current patrol scheme to join a dispatch candidate list, the road network data in the calculation range in the road network database is called by the road network calculation engine, a patrol scheme with the shortest time consumption is filtered out from the dispatch candidate list, the state of the patrol scheme is set to the highest priority, and corresponding patrol route information is obtained.

[0036] Further, the method as described above, S500 further comprises:

[0037] When no latest police case occurs, based on the case occurrence probability of all road segments in each jurisdiction in the future one week, the road network calculation engine is used to obtain patrol schemes that respectively meet three patrol targets of the largest patrol range, the largest cumulative patrol case area and the largest risk, and the lowest patrol overlap degree among police forces within a certain time, starting from the location of each jurisdiction police station, and corresponding patrol route information is obtained.

[0038] A criminal spatiotemporal risk prediction and decision-making dispatch device, comprising:

[0039] An acquisition module is configured to acquire data resources and establish corresponding databases, wherein the data resources include case event spatiotemporal data, spatiotemporal background environment data, and subject behavior data, and the acquisition module includes: establishing a historical case event database and a suspect foot point database based on the case event spatiotemporal data, establishing a spatiotemporal environment database, a road network database, and a street view database based on the spatiotemporal background environment data, and establishing a spatiotemporal trajectory database and a real-time police force resource distribution database based on the subject behavior data;

[0040] A data management module is configured to perform data management based on a to-be-predicted road segment, the suspect foot point database, the spatiotemporal environment database, the road network database, the street view database, and the spatiotemporal trajectory database, as well as a road network calculation engine, a graph calculation engine, and a receptive field filter, to obtain output data, wherein the output data includes: a criminal opportunity of all potential foot points of criminals in the receptive field of the to-be-predicted road segment to the to-be-predicted road segment, a social structure feature of the criminals, a direction entropy of all road segments in the receptive field, a cognitive feature vector of the to-be-predicted road segment, and a road network structure feature of the to-be-predicted road segment in the receptive field;

[0041] a future crime occurrence probability estimation module configured to estimate a crime occurrence probability of the to-be-predicted road segment in a week based on a deep learning-based CTR estimation model, the output data, and a cumulative risk value of the to-be-predicted road segment, wherein the cumulative risk value of the to-be-predicted road segment is calculated based on crime opportunities from potential landing points of all criminals in a receptive field of the to-be-predicted road segment to the to-be-predicted road segment;

[0042] a probability regularization module configured to regularize the crime occurrence probability of the to-be-predicted road segment in a week based on the crime occurrence probability of the to-be-predicted road segment in a week, the receptive field filter, the jurisdiction filter, a hotspot migration rule library, and the historical case database, to obtain a corrected crime occurrence probability of the to-be-predicted road segment in a week;

[0043] a scheduling module configured to schedule police forces and plan patrol routes based on location information of a latest police case, the receptive field filter, crime occurrence probabilities of all road segments in each jurisdiction in a week, the road network database, the road network calculation engine, and the real-time police force resource distribution database.

[0044] The present application has the following advantages: Based on the historical case database and the suspect landing point database, the spatio-temporal environment database, the road network database and the street view database, the spatio-temporal trajectory database and the real-time police force resource distribution database, the crime opportunities from potential landing points of all criminals in a receptive field of a to-be-predicted road segment to the to-be-predicted road segment, the social structure characteristics of criminals, the directional entropy of all road segments in the receptive field, the cognitive feature vector of the to-be-predicted road segment, and the road network structure characteristics of the to-be-predicted road segment in the receptive field can be obtained. Based on the deep learning-based CTR estimation model and the above data, the crime occurrence probability of the to-be-predicted road segment in a week can be estimated. By regularizing the crime occurrence probability of the to-be-predicted road segment in a week, the corrected crime occurrence probability of the to-be-predicted road segment in a week can be obtained. Meanwhile, police force scheduling and patrol route planning can be realized. The present application adopts a reasoning method based on a road network to identify and predict crime patterns. Compared with a general crime spatio-temporal prediction model based on a grid, the present application uses a road network as an analysis carrier and a road segment as an analysis unit, which alleviates the problem of data sparsity and solves the problem of grid division scale selection, effectively adapting to spatial heterogeneity. Meanwhile, in order to facilitate the execution and scheduling of prediction results, the present application can also correct patrol routes in real time according to prediction results, thereby improving patrol coverage rate and patrol frequency of high-risk road segments. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 a flowchart of a crime spatio-temporal risk prediction and decision scheduling method provided in an embodiment of the present application;

[0046] Figure 2 a structural diagram of a crime spatio-temporal risk prediction and decision scheduling system provided in an embodiment of the present application;

[0047] Figure 3 A schematic diagram of the inter-picture relationship of the street mapping module provided in the embodiment of the present application is shown in the figure;

[0048] Figure 4 A schematic diagram of the scene semantic tree of the street mapping module provided in the embodiment of the present application is shown in the figure;

[0049] Figure 5 A schematic diagram of the model training of the micro-feature fusion module provided in the embodiment of the present application is shown in the figure;

[0050] Figure 6 A structural schematic diagram of the crime spatio-temporal risk prediction and decision scheduling device provided in the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0051] The present application will be further described in detail below in combination with the accompanying drawings and specific embodiments.

[0052] As shown in the figure, a crime spatio-temporal risk prediction and decision scheduling method comprises the following steps: Figure 1 S100, data resources are acquired and a corresponding database is established, wherein the data resources include case event spatio-temporal data, spatio-temporal background environment data and subject behavior data, comprising: a historical case event database and a suspect landing point database are established based on the case event spatio-temporal data, a spatio-temporal environment database, a road network database and a street view database are established based on the spatio-temporal background environment data, and a spatio-temporal trajectory database and a real-time police resource distribution database are established based on the subject behavior data.

[0053] For the case event spatio-temporal data, the spatio-temporal location data therein is stored in a geographic database, and other attribute information is stored in a relational database. For the spatio-temporal environment database and the road network database, they belong to structured data, the spatio-temporal environment database includes POI and other spatio-temporal environment data, and the road network database includes OSM and other road network data, the spatial location data in the structured data is stored in the geographic database, and other attribute information is stored in the relational database. For the street view database, it belongs to unstructured data, the spatial location data therein is stored in the geographic database, and the image is directly stored in the file system, and the image can also be vectorized and stored in a Milvus or other vector retrieval engine for convenient retrieval in the later stage. For the subject behavior data, it is stored in a time series database.

[0054] It should be noted that the above data resources can be acquired through online public channels or through other legal channels.

[0055]

[0056] ​S200, based on the databases of the road segment to be predicted, the suspect's hideout, the spatiotemporal environment, the road network, the street view, and the spatiotemporal trajectory, as well as the road network computing engine, the graph computing engine, and the receptive field filter, performs data governance to obtain output data. The output data includes: the potential hideout of all criminals in the receptive field of the road segment to be predicted, the crime opportunity of all criminals in the road segment to be predicted, the social structure characteristics of criminals, the directional entropy of all road segments in the receptive field, the cognitive feature vector of the road segment to be predicted, and the road network structure characteristics of the road segment to be predicted in the receptive field.

[0057] After acquiring the data resources in S100, data governance is required. This includes preprocessing and feature engineering the data using multiple integrated models to ultimately obtain the potential locations of all criminals within the receptive field of the road segment to be predicted, the criminal opportunities of that road segment, the social structure characteristics of the criminals, the directional entropy of all road segments within the receptive field, the cognitive feature vector of that road segment, and the road network structure characteristics of the road segment to be predicted within the receptive field.

[0058] like Figure 2 As shown, for receptive field filter 1, the receptive field is a spatiotemporally continuous and correlated range of the current spatiotemporal location to be processed. Receptive field filter 1 takes the geographic coordinates to be queried as input and calculates a 30-minute walking isochronous circle as the starting point as a data filter, obtaining the data within the data filter as the return value. The receptive field of the road segment to be predicted refers to selecting a road segment as the road segment to be predicted, inputting the geographic coordinates of the road segment to be predicted into receptive field filter 1, which yields the receptive field of the road segment to be predicted.

[0059] In S200, the potential locations of all criminals within the perception field of the predicted road segment are obtained, along with the crime opportunities on the predicted road segment, including:

[0060] Based on the unified opportunity travel model and graph computing engine, an effective distance map of road segment jumps with spatiotemporal risks is obtained. Within the receptive field obtained based on the road segment to be predicted and the receptive field filter, the potential locations of all criminals are obtained from the suspect location database. Based on the effective distance map of road segment jumps, the travel probability from the potential location to the road segment to be predicted, i.e., the crime opportunity, is calculated.

[0061] Road networks are one of the carriers of spatiotemporal trajectories, but the accessibility of a location in spatiotemporal space cannot be directly reflected by geographical distance. Environmental criminology's crime pattern theory mentions that the spatiotemporal distribution of people causes the emergence and disappearance of crime-attracting and crime-generating locations. This spatiotemporal distribution of people cannot be directly reflected by geographical distance in today's convenient transportation environment. A criminal's opportunity to commit a crime is closely related to the spatiotemporal distribution of people. By converting geographical distance into effective distance based on traffic flow, a unified opportunity travel model can be better modeled (Intervention Opportunity Calculation Module 15). Figure 2The system further comprises an effective transfer distance conversion module 11, a road network matching module 12 and a space-time trajectory database 10. The space-time trajectory database 10 is called to obtain the space-time trajectory, the road section corresponding to the trajectory is obtained through the road network matching module 12, and the missing road section is supplemented according to the actual road network. The space-time trajectory is divided by day, and each day from 2 o'clock to the next day 2 o'clock is divided into 6 time slices with a span of 4 hours. The frequency of jumping between road sections in each slice is calculated. The effective distance between road sections in the road network is calculated using the effective distance in the network. The effective distance uses the definition of Brockmann in The hidden geometry of complex, network-driven contagion phenomena:

[0062] d mn =(1-logP mn )≥1

[0063] where P mn is the ratio of the flow from m to n to the total flow of m, that is, the probability of jumping from node m to node n. The effective distance is stored as a directed weight of the relationship between road sections according to different time slices, and the road section jump effective distance graph of different time slices is output. The spatial distance is weighted according to the aggregation degree of the crowd in space-time by the effective distance transfer module 11, so that the weighted distance of the area with relatively large aggregation degree is smaller than that of the area with relatively small aggregation degree under the same geographical distance, in order to better simulate the accessibility of the crime risk area.

[0064] As Figure 2 shown, in order to realize the intervention opportunity calculation module 15, the road section jump effective distance graph 13, the graph calculation engine 14, the receptive field filter 1 and the suspect landing point database 16 are needed. Specifically, the intervention opportunity calculation module 15 uses the unified opportunity travel model to obtain the road section jump effective distance graph with space-time risk through the graph calculation engine, obtains the potential landing point data from the suspect landing point database within the receptive field range obtained based on the receptive field filter, and calculates the travel probability from the potential landing point to the to-be-predicted road section, that is, the crime opportunity. The unified opportunity travel model adopts the following model:

[0065]

[0066] where m i is the starting opportunity value, m j is the ending opportunity value, and α and β are hyperparameters (0≤α+β≤1) respectively measuring the exploratory tendency and cautiousness of the object, s ijThe intervention opportunity value is α=β=0.5 for a suspect with only one case; and α=n / (n+m), β=m / (n+m) for a suspect with more than one case; n is the number of cases outside the jurisdiction of the landing point, and m is the number of cases in the jurisdiction of the landing point.

[0067] In S200, the social structure features of all criminals in the receptive field of the to-be-predicted road section are obtained, including:

[0068] Based on the crime relationship, household relationship and spatial proximity relationship of all criminals in the receptive field of the to-be-predicted road section, a potential social relationship is constructed, and the social structure features of the criminals are obtained using the node2vec model based on the potential social relationship.

[0069] In S200, the road network structure features of the to-be-predicted road section in the receptive field are obtained, including:

[0070] In the receptive field obtained based on the to-be-predicted road section and the receptive field filter, the corresponding road network data is screened out based on the road network database, the graph embedding operation is performed on the screened road network data by the graph computing engine, and the road network structure features of the to-be-predicted road section in the receptive field are obtained using the node2vec model.

[0071] Figure 2 The structure feature calculation module 7 is mainly divided into two parts, one is to obtain the social network structure features of the criminals, and the other is to obtain the structure features of the to-be-predicted road section in the road network. For the road network, the receptive field filter 1 needs to be called, the graph embedding operation is performed on the screened road network, and the feature vector of the to-be-predicted road section is obtained using the node2vec model; for the social network structure features, the potential social relationship needs to be constructed from the crime relationship, the household relationship and the spatial proximity relationship (the crime relationship is the same as the relationship, the household relationship is the same as the source province, and the spatial proximity relationship is the landing point spatial proximity), and the node2vec model is also used to obtain the social network structure features of the criminals, and the social network structure features of the criminals corresponding to the landing points in the receptive field are returned.

[0072] In S200, the cognitive feature vector of the to-be-predicted road section is obtained, including:

[0073] In S200a, based on the street view database and the pre-trained deep learning model, the panoramic segmentation is performed on the street view data of the to-be-predicted road section, the street view atlas corresponding to the to-be-predicted road section is obtained based on the entity and scene semantic tree obtained by segmentation, the street view atlas corresponding to the to-be-predicted road section is taken as the objective cognitive representation of the scene, and the six-dimensional cognitive vector is taken as the subjective cognitive representation of the scene.

[0074] The corresponding street view atlas is obtained by using the open dataset of the Place Pulse project for atlas processing, and the six-dimensional cognitive vector is formed by using the graph convolution network to train the corresponding street view atlas.

[0075] first, Figure 2 The street view mapping module 19 in the middle performs panoramic segmentation on the street view data of the corresponding road segments collected in the street view database 18 using a pre-trained deep learning model, and then forms a map of the segmented entities based on the scene semantic tree.

[0076] Specifically, the scene semantic tree is composed of common objects in street scenes, segmented according to the object's state, attributes, and type. This invention uses a panoptic segmentation model pre-trained on the COCO dataset for image processing, and its scene semantic tree is as follows: Figure 4 As shown, the numbers 0-2 in the leaf nodes represent the spatial locations where the object frequently appears (ground, on the ground but not detached from the ground, and sky), respectively. Within the same spatial location, static objects have an associative relationship, while dynamic objects have a dependency relationship (weaker than an associative relationship). Objects in adjacent spatial locations have a dependency relationship. Adjacent objects in the image are associative if both are static, and dependent if either is a dynamic object. Figure 3 As shown, if street scenes from multiple angles are captured at the same spatial location, the relationship between objects in the images follows the principle of "no cross-layer processing." That is, objects at the same spatial location establish relationships, while objects at different spatial locations do not establish relationships. Within the same layer, if objects in two different images have the same parent node in the scene semantic tree, an association relationship is established if both are static objects, and a dependency relationship is established if one of them is a dynamic object. If objects in two different images have the same grandparent node, an association relationship is established between static objects, and a dependency relationship is established if dynamic objects exist.

[0077] Then, the same graphing process was performed using the open dataset from the Place Pulse project, and a graph convolutional network was used to train the street view atlas to form a six-dimensional cognitive vector. The trained model was then transferred to the system to generate a cognitive vector for each street view atlas.

[0078] Finally, the street view map of the road segment to be predicted is used as the objective cognitive representation of the scene, and the six-dimensional cognitive vector is used as the subjective cognitive representation of the scene.

[0079] S200b: Based on the road segment to be predicted and the road network database, the corresponding road network data is sampled to form a dataset. When training the dataset, for each set of objective cognitive features and subjective cognitive features, the objective cognitive features are sequentially passed through the first graph neural network layer, the graph pooling layer, the second graph neural network layer, and the graph readout layer to obtain the corresponding objective cognitive feature vector. The corresponding objective cognitive feature vector is then concatenated with the subjective cognitive feature vector corresponding to the subjective cognitive features and linear regression prediction is performed to obtain the road segment risk value. Finally, the cognitive feature vector of the road segment to be predicted is output.

[0080] likeFigure 5 As shown, Figure 2 The micro-cognitive feature fusion module 20 trains and fuses to generate cognitive feature vectors for the road segments to be predicted. First, the road network is sampled to form a dataset. During dataset training, for each set of objective and subjective cognitive features, the objective cognitive representation is passed through a graph neural network layer and a graph pooling layer, followed by a readout operation. The readout feature vector is then concatenated with the subjective cognitive vector before linear regression is used to predict the road segment risk value. For a road segment with one street view collection point, all subjective feature vectors are summed directly, while objective features are summed after readout. For road segments without a street view but still sampled into the dataset, their objective and subjective feature vectors are randomly initialized and directly concatenated for linear regression prediction; that is, only the parameters of the linear regression part are updated during training. After training, the final output is the objective and subjective feature vectors of the road segments to be predicted.

[0081] Figure 2 The system also includes a dual-exponential trend fitting module 3, which uses a dual-exponential model to fit the historical incidents of a road segment according to their time sequence, and returns the probability of incidents occurring in the next stage. The dual-exponential model adopts the collective memory decay model proposed by Cristian Candia et al. in Nature in 2018, and is defined as follows:

[0082]

[0083] Where N is the initial chance value, set to 1; p, q, and r are decay parameters, which can be obtained by fitting data from the historical case database 2. Cases belonging to each road segment are arranged chronologically, and a moving average of the number of cases is calculated using a certain time window (default one week). The time t=0 is taken as the moment with the largest moving average in the most recent month. Subsequent chance values ​​are the ratio of the moving average at each moment to the moving average at the initial moment. For cases where the current moment is the maximum, the default chance of a case occurring in the next stage is 50%.

[0084] Considering the recurring nature of criminal activities, it is necessary to utilize Figure 2 The first-order road segment propagation module 4 calls the graph calculation engine 14 to obtain corresponding data from the effective distance map of road segment jumps 13, calculates the risk of the predicted road segment being affected by neighboring road segments in the road network, and defines the propagation range as the first-order neighbors of road segments that have experienced incidents in the past month, whose incident probability is half that of the current road segment. The incident probability of the remaining road segments without incident records is set to 20%.

[0085] Since the crime risk in space-time is related to spatial heterogeneity, the characterization of heterogeneity can be calculated by POI and other space-time environment data. According to the historical crime data statistics, the specific type of space-time environment data corresponding to the crime place can be obtained, and the specific type of space-time data meeting the conditions is screened out. Then the feature engineering is used to convert the specific type of space-time data and the crime opportunity of the road section into a feature vector, and the road section where the crime has occurred within a month is marked as 1, otherwise 0. The crime opportunity of the road section calculated by the road section propagation module is used as its opportunity value if the road section has no crime opportunity. Due to the characteristics that the space-time environment data will not change in a short period of time, the Bayesian regression model adopts an offline update mode and does not need to update the parameters in real time.

[0086] In addition, Figure 2 The direction entropy calculation module 17 in the direction entropy calculation module 17 is used to calculate the direction entropy of all road sections in the receptive field of the road section to be predicted. The direction entropy is an index to measure the degree of confusion of the road network in a region. In 2019, Geoff Boeing published in Applied Network Science, which is defined as follows:

[0087]

[0088] Where P(o i ) is the ratio of the road section in the i-th direction to the total number of road sections in the region (set the north direction as 0°, the starting position as 5°, and divide 360° into 36 equal parts every 10°), and the weighted version of the direction entropy is used in the present application, that is, the length of the i-th direction is considered. The ratio of the total length of the road section in the region. The region here is divided using the receptive field.

[0089] S300, based on the CTR estimation model of deep learning, output data and the cumulative risk value of the road section to be predicted, estimate the crime probability of the road section to be predicted in the future one week, wherein the cumulative risk value of the road section to be predicted is calculated based on the crime opportunity of the potential landing point of all criminals in the receptive field of the road section to be predicted to the road section to be predicted.

[0090] Figure 2The future crime probability estimation module 21 in the future crime probability estimation module 21 is different from the past neural network prediction model with matrix as input, and the improved deep click rate estimation model (CTR estimation model) based on neural network is used for future crime probability prediction. The outputs of the structure feature calculation module 7, the intervention opportunity calculation module 15, the direction entropy calculation module 17 and the micro cognitive feature fusion module 20 are used as the input of the click rate estimation model, and the final output result is the crime probability of the to-be-predicted road section in the future one week. The CTR estimation model is originally applied to the advertisement click rate prediction task in the e-commerce scene in the recommendation system field, and the probability of the crime event occurring in space and time is modeled as the probability of the potential criminal committing a crime from the landing point to the to-be-predicted road section. The 30-minute isochronal circle is used as the screening condition. For the definition of the potential criminal landing point, the invention selects the historical landing point of the delinquent as a high-risk landing point, and considers the problems of external population and landing point change. The invention selects the location of the residential land such as residential area, village and hotel in the isochronal circle as the potential risk landing point. The travel probability of each landing point to the to-be-predicted road section is obtained by the intervention opportunity calculation module, and then the cumulative probability is obtained by using the weighted method. First, the high-risk landing point is assigned a weight of 1, the potential risk landing point is assigned a weight of 0.1, and the weight of all landing points in the isochronal circle is normalized to a decimal between 0 and 1. Then, the cumulative risk value of the to-be-predicted road section is calculated by multiplying the weight of each landing point by the corresponding travel probability. Since the opportunity value will change with the update of the case event data, the cumulative risk value is also a dynamically changing value. The same weighting operation is also performed on the social structure feature vector of the delinquent with historical landing points in the isochronal circle (the weight is the same as above, and the feature vector of the potential risk landing point is a 0 vector). Finally, the input of the CTR estimation model is the cumulative risk value, the summed social structure feature vector of the delinquent, the structure feature vector of the to-be-predicted road section, the direction entropy of the region where the to-be-predicted road section is located, and the subjective and objective feature vector of the to-be-predicted road section.

[0091] The recommendation system field has developed rapidly in recent years, and there are many types of CTR estimation models. The general models include CCPM, PNN, Wide&Deep, DeepFM, MLR, NFM, etc. If the time sequence needs to be considered, the models that can be selected include DIN, DIEN, etc.

[0092] S400, based on the future one-week crime probability of the to-be-predicted road section, the receptive field filter, the jurisdiction filter, the hot spot migration rule library and the historical case event database, the regularization is performed to obtain the corrected future one-week crime probability of the to-be-predicted road section.

[0093] S400 comprises: taking the union of the total range of the jurisdiction where the to-be-predicted road segment is located and its adjacent jurisdictions filtered based on the jurisdiction filter and the range of the receptive field filtered based on the receptive field filter, to obtain a union range, obtaining all crime-prone road segments in the union range based on the historical case event database, calculating the transfer probability of all crime-prone road segments potentially transferred to the to-be-predicted road segment based on all crime-prone road segments in the union range and the hotspot transfer rules in the hotspot transfer rule library, taking the maximum value in all transfer probabilities, and correcting the case occurrence probability of the to-be-predicted road segment in the future one week through the Bayes formula.

[0094] As shown in Figure 2 , the estimated probability regularization module 22 needs to use the jurisdiction filter 24, the receptive field filter 1, the hotspot transfer rule library 23, and the historical case event database 2. Specifically, after obtaining the case occurrence probability of the to-be-predicted road segment in the future one week through the future case occurrence probability estimation module 21, the estimated probability regularization module 22 is needed to regularize the probability. First, the total range of the jurisdiction where the to-be-predicted road segment is located and its adjacent jurisdictions is filtered through the jurisdiction filter 24 and taken the union with the range filtered by the receptive field filter 1. All crime-prone road segments in the range are obtained through the historical case event database 2, and the risk of potential transfer to the road segment is calculated through the hotspot transfer rules in the hotspot transfer rule library 23. The maximum value in all transfer probabilities is taken, and the case occurrence probability of the current road segment in the future one week is corrected through the Bayes formula.

[0095] Through S100-S400, the corrected case occurrence probability of the to-be-predicted road segment in the future one week can be obtained. Through the method of S100-S400, the case occurrence probability of all road segments in each jurisdiction in the future one week can be obtained. Based on the case occurrence probability of all road segments in each jurisdiction in the future one week, police scheduling and patrol route planning can be realized.

[0096] S500, based on the latest police situation occurrence location information, the receptive field filter, the case occurrence probability of all road segments in each jurisdiction in the future one week, the road network database, the road network calculation engine, and the real-time police resource distribution database, performs police scheduling and patrol route planning.

[0097] S500 comprises:

[0098] S500a, when the latest police case occurs, based on the occurrence position information of the latest police case and the receptive field filter, taking the street corresponding to the latest police case as the starting point to obtain a predetermined time circle, intersecting the predetermined time circle with the jurisdiction range of the latest police case to obtain an intersection range, obtaining all on-duty patrol police forces in the intersection range based on the real-time police force resource distribution database, screening out the patrol scheme whose state is not the highest priority from the current patrol scheme to join the dispatch candidate list, calling the road network data in the calculation range in the road network database through the road network calculation engine, screening out the patrol scheme with the shortest time from the dispatch candidate list, setting its state to the highest priority, and obtaining the corresponding patrol route information.

[0099] S500b, when the latest police case does not occur, based on the case occurrence probability of all road segments in each jurisdiction within a week, through the road network calculation engine, taking the location of each jurisdiction police station as the starting point, calculating the patrol scheme that meets the three patrol targets of the largest patrol range, the largest cumulative patrol case area and risk, and the lowest patrol overlap degree among police forces within a certain time, and obtaining the corresponding patrol route information.

[0100] As shown in Figure 2 To implement the police force dispatching module 25, the real-time police force resource distribution database 25, the latest case occurrence position monitoring module 26, the road network calculation engine 6 and the road network database 5 are needed. Specifically, the police force dispatching module 25 divides the next stage route planning scheme into three levels: first level, second level and third level, where the first level is the highest priority. When the latest police case occurrence position is input, the receptive field filtering module 1 is called first to take the street corresponding to the police case as the starting point to obtain a 10-minute car isochronal circle and intersect it with the jurisdiction range of the police case, obtain all on-duty patrol police forces in the range from the real-time police force resource distribution database 25, screen out the current patrol scheme whose state is not the first level into the dispatch candidate list, call the road network data in the calculation range in the road network database 5 through the road network calculation engine 6, screen out the patrol scheme with the shortest time from the candidate list, set its state to the first level, and return the navigation route information. For the case where the latest police case has not occurred in the system, obtain the case occurrence probability of all road segments in each jurisdiction within a week, calculate the patrol scheme that meets the three patrol targets of the largest patrol range, the largest cumulative patrol case area and risk, and the lowest patrol overlap degree among police forces within a certain time from the location of the jurisdiction police station through the road network calculation engine 6, and return the corresponding patrol route planning.

[0101] Existing technologies primarily utilize grid-based reasoning for crime pattern identification and prediction, but the rationality of the prediction method and the interpretability of the prediction results cannot be fully guaranteed. This invention integrates data of various forms and scales, including images, text, and tables, and employs a road network-based reasoning approach for crime pattern identification and prediction. Compared to typical grid-based spatiotemporal crime prediction models, using the road network as the analysis carrier and road segments as the analysis unit alleviates the problem of data sparsity and solves the problem of choosing the grid division scale, effectively adapting to spatial heterogeneity.

[0102] To address the problem of risk prediction in zero-sample spatiotemporal areas, this invention uses a logistic regression method to jointly predict the potential spatiotemporal risk probability of zero-sample areas by utilizing statistical indicators of environmental factors in a local area, statistical indicators of crime data in areas where cases have occurred, and potential migration information of existing risk hotspots within a certain range.

[0103] Addressing the issues of spatiotemporal opportunities for crime and crime travel modeling based on these opportunities, this invention utilizes trajectory data and road network data to model the crime generation mechanism within crime pattern theory, based on effective distance. It also uses environmental factor data such as Points of Interest (POIs) combined with crime incidence rates to characterize crime opportunities. Furthermore, it employs a unified opportunity travel model to predict potential routes for criminals.

[0104] This invention applies street view data to a crime prediction system and proposes a method for constructing cognitive feature vectors to represent human cognition of the current road environment.

[0105] Considering the recurring nature of criminal activity and the patterns of criminal travel, this invention proposes a trend-fitting model based on a dual-exponential model to fit the changing trends of spatiotemporal risk in areas where crimes have occurred. Based on a receptive field analysis method, the theoretical model is restored while reducing the computational load of the system.

[0106] Meanwhile, the prediction model used in this invention partially uses spatiotemporal feature vectors with contextual information as input, avoiding the problem of input dimension changes caused by changes in the prediction range, and thus having stronger transferability.

[0107] In addition, this invention adds a prediction probability regularization module to assist users in decision-making. Based on the hotspot migration rules of crime pattern theory in environmental criminology, the prediction results are corrected, which improves the interpretability of the prediction results and increases the flexibility of correction.

[0108] Finally, to facilitate the execution and scheduling of the prediction results, this invention provides a police force scheduling module that can correct patrol routes in real time based on the prediction results, thereby increasing the patrol coverage and frequency of high-risk road sections.

[0109] like Figure 6 As shown, a crime spatiotemporal risk prediction and decision-making scheduling device includes:

[0110] The acquisition module 601 is used to acquire data resources and establish corresponding databases. The data resources include case and event spatiotemporal data, spatiotemporal background environment data, and subject behavior data, including: establishing a historical case and event database and a suspect location database based on the case and event spatiotemporal data; establishing a spatiotemporal environment database, a road network database, and a street view database based on the spatiotemporal background environment data; and establishing a spatiotemporal trajectory database and a real-time police resource distribution database based on the subject behavior data.

[0111] The data governance module 602 is used to perform data governance based on the database of the road segment to be predicted, the database of the suspect's location, the database of spatiotemporal environment, the database of road network, the database of street view, and the database of spatiotemporal trajectory, as well as the road network computing engine, the graph computing engine, and the receptive field filter, to obtain output data. The output data includes: the potential location of all criminals in the receptive field of the road segment to be predicted and the crime opportunity of the road segment to be predicted, the social structure characteristics of the criminals, the directional entropy of all road segments in the receptive field, the cognitive feature vector of the road segment to be predicted, and the road network structure characteristics of the road segment to be predicted in the receptive field.

[0112] The future crime probability prediction module 603 is used to predict the crime probability of the road segment to be predicted in the next week by using the deep learning-based CTR prediction model, output data and cumulative risk value of the road segment to be predicted. The cumulative risk value of the road segment to be predicted is calculated based on the potential location of all criminals in the perception field of the road segment to be predicted and the crime opportunity of the road segment to be predicted.

[0113] The probability prediction regularization module 604 is used to perform regularization based on the probability of incidents on the road segment to be predicted in the next week, the receptive field filter, the jurisdiction filter, the hotspot migration rule base and the historical case event database, to obtain the corrected probability of incidents on the road segment to be predicted in the next week.

[0114] The dispatch module 605 is used to dispatch police forces and plan patrol routes based on the latest location information of police incidents, the sensor field filter, the probability of incidents on all road sections in each jurisdiction for the next week, the road network database, the road network computing engine, and the real-time police resource distribution database.

[0115] In the embodiment, the crime space-time risk prediction and decision scheduling device can be a computer, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the aforementioned crime space-time risk prediction and decision scheduling method by executing the executable instructions. The memory and the processor can be connected through a bus. The storage unit can include a readable medium in the form of a volatile storage unit, such as a random access memory (RAM) and / or a cache memory unit, and can further include a read-only memory (ROM). The computer further comprises a display unit connected to the bus. The display unit can display the aforementioned patient's potential important information, etc.

[0116] A computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by the aforementioned processor to implement the aforementioned crime space-time risk prediction and decision scheduling method.

[0117] It should be noted that the crime space-time risk prediction and decision scheduling device and method in the embodiment are the same inventive concept, and the functions of the crime space-time risk prediction and decision scheduling device can be seen in the crime space-time risk prediction and decision scheduling method embodiment.

[0118] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications and variations.

Claims

1. A crime space-time risk prediction and decision scheduling method, characterized in that, The 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 subject behavior data, and the method comprises: establishing a historical case event database and a suspect landing point database based on the case event space-time data, establishing a space-time environment database, a road network database, and a street view database based on the space-time background environment data, and establishing a space-time trajectory database and a real-time police resource distribution database based on the subject behavior data; S200, based on a to-be-predicted road section, the suspect landing point database, the space-time environment database, the road network database, the street view database, and the space-time trajectory database, as well as a road network calculation engine, a graph calculation engine, and a receptive field filter, data governance is performed to obtain output data, the output data comprising: a crime opportunity of all potential landing points of criminals in a receptive field of the to-be-predicted road section to the to-be-predicted road section, a social structure feature of the criminals, a direction entropy of all road sections in the receptive field, and a cognitive feature vector of the to-be-predicted road section, and a road network structure feature of the to-be-predicted road section in the receptive field; the receptive field filter takes a to-be-queried geographic coordinate as input, and calculates a 30-minute isochrone as a data filter from the to-be-queried geographic coordinate as a starting point, and obtains data in the data filter as a return value; S300, based on a deep learning-based CTR estimation model, the output data, and a cumulative risk value of the to-be-predicted road section, a crime probability of the to-be-predicted road section in a future week is estimated, wherein the cumulative risk value of the to-be-predicted road section is calculated based on the crime opportunity of all potential landing points of criminals in the receptive field of the to-be-predicted road section to the to-be-predicted road section; S400, based on the crime probability of the to-be-predicted road section in the future week, the receptive field filter, a jurisdiction filter, a hotspot migration rule library, and the historical case event database, regularization is performed to obtain a corrected crime probability of the to-be-predicted road section in the future week; S500, based on occurrence position information of a latest police case, the receptive field filter, a crime probability of all road sections in each jurisdiction in a future week, the road network database, the road network calculation engine, and the real-time police resource distribution database, police scheduling and patrol route planning are performed.

2. The method of claim 1, wherein, In S100, For the case event space-time data, space-time position data therein is stored in a geographic database, and other attribute information is stored in a relational database; For the space-time environment database and the road network database, spatial position data therein is stored in a geographic database, and other attribute information is stored in a relational database; For the street view database, spatial position data therein is stored in a geographic database, images are stored in a file system, and vectorized images are stored in a vector retrieval engine; For the subject behavior data, it is stored in a time series database.

3. The method of claim 1, wherein, In S200, the crime opportunity of all potential landing points of criminals in a receptive field of the to-be-predicted road section to the to-be-predicted road section is acquired, comprising: Based on the uniform opportunity travel model and the graph computing engine, an effective distance graph of a road section jump with a space-time risk is obtained, and within a receptive field obtained based on the to-be-predicted road section and the receptive field filter, all potential landing points of criminals from a suspect landing point database are obtained, and a travel probability from the potential landing points to the to-be-predicted road section, i.e., a criminal opportunity, is calculated based on the effective distance graph of the road section jump.

4. The method of claim 1, wherein, In S200, the social structure features of all criminals within the receptive field of the to-be-predicted road section are obtained, including: Based on the criminal relationship, the household relationship and the spatial proximity relationship of all criminals within the receptive field of the to-be-predicted road section, a potential social relationship is constructed, and the social structure features of the criminals are obtained using a node2vec model based on the potential social relationship.

5. The method of claim 1, wherein, In S200, the road network structure features of the to-be-predicted road section within the receptive field are obtained, including: Within the receptive field obtained based on the to-be-predicted road section and the receptive field filter, corresponding road network data is filtered based on the road network database, graph embedding operation is performed on the filtered road network data by the graph computing engine, and the road network structure features of the to-be-predicted road section within the receptive field are obtained using a node2vec model.

6. The method of claim 1, wherein, In S200, the cognitive feature vector of the to-be-predicted road section is obtained, including: Based on the street view database and a pre-trained deep learning model, panoramic segmentation is performed on the street view data of the to-be-predicted road section, a street view atlas corresponding to the to-be-predicted road section is obtained based on the segmented entity and scene semantic tree, the street view atlas corresponding to the to-be-predicted road section is taken as an objective cognitive representation of a scene, and a six-dimensional cognitive vector is taken as a subjective cognitive representation of the scene; Based on the to-be-predicted road section and the road network database, corresponding road network data is sampled to form a data set, and when the data set is trained, for each set of objective cognitive features and subjective cognitive features, the objective cognitive features therein are sequentially subjected to a first graph neural network layer, a graph pooling layer, a second graph neural network layer and a graph readout layer to obtain a corresponding objective cognitive feature vector, and the corresponding objective cognitive feature vector and a subjective cognitive feature vector corresponding to the subjective cognitive features therein are sequentially subjected to splicing operation and linear regression prediction to obtain a road section risk value, and finally the cognitive feature vector of the to-be-predicted road section is output.

7. The method of claim 1, wherein, S400 includes: The total range of the jurisdiction of the to-be-predicted road section and its adjacent jurisdictions filtered based on the jurisdiction filter is combined with the receptive field range filtered based on the receptive field filter to obtain a combined range, all high-crime road sections within the combined range are obtained based on the historical case event database, a transfer probability of all high-crime road sections potentially transferred to the to-be-predicted road section is calculated based on all high-crime road sections within the combined range and the hotspot transfer rules in the hotspot transfer rule library, the maximum value in all transfer probabilities is taken, and the case occurrence probability of the to-be-predicted road section in the future week is corrected through a Bayesian formula.

8. The method according to any one of claims 1 to 7, characterized in that, S500 includes: When the latest police case occurs, based on the occurrence position information of the latest police case and the receptive field filter, a predetermined time circle is obtained starting from the street corresponding to the latest police case, the intersection range is obtained by intersecting the predetermined time circle and the jurisdiction range to which the latest police case belongs, all on-duty patrol police forces in the intersection range are obtained based on the real-time police force resource distribution database, the patrol scheme whose state is not the highest priority is filtered out from the current patrol scheme to join the dispatch candidate list, the road network data within the calculation range in the road network database is called by the road network calculation engine, the patrol scheme with the shortest time consumption is filtered out from the dispatch candidate list, the state thereof is set to the highest priority, and the corresponding patrol route information is obtained.

9. The method according to any one of claims 1 to 7, characterized in that, S500 further includes: When the latest police case does not occur, based on the case occurrence probability of all road segments in each jurisdiction within a week, the road network calculation engine is used to calculate patrol schemes that respectively meet three patrol targets of the largest patrol range, the largest cumulative patrol case area and the largest risk, and the lowest patrol overlap degree among police forces within a certain time, starting from the location of each jurisdiction police station, and corresponding patrol route information is obtained.

10. A crime spatio-temporal risk prediction and decision dispatching device, characterized in that, It includes: An acquisition module is configured to acquire data resources and establish corresponding databases, wherein the data resources include case event spatio-temporal data, spatio-temporal background environment data, and subject behavior data, including: establishing a historical case event database and a suspect landing point database based on the case event spatio-temporal data, establishing a spatio-temporal environment database, a road network database, and a street view database based on the spatio-temporal background environment data, and establishing a spatio-temporal trajectory database and a real-time police force resource distribution database based on the subject behavior data; A data management module is configured to perform data management based on a to-be-predicted road segment, the suspect landing point database, the spatio-temporal environment database, the road network database, the street view database, and the spatio-temporal trajectory database, as well as a road network calculation engine, a graph calculation engine, and a receptive field filter, to obtain output data, including: a crime opportunity of all criminals' potential landing points in the receptive field of the to-be-predicted road segment to the to-be-predicted road segment, a social structure feature of the criminals, a direction entropy of all road segments in the receptive field, and a cognitive feature vector of the to-be-predicted road segment, and a road network structure feature of the to-be-predicted road segment in the receptive field; the receptive field filter takes a to-be-queried geographic coordinate as input and calculates a 30-minute isochronal circle as a data filter with the to-be-queried geographic coordinate as a starting point, and obtains data in the data filter as a return value; A future case occurrence probability estimation module is configured to estimate a case occurrence probability of a to-be-predicted road segment within a week based on a deep learning-based CTR estimation model, output data, and a cumulative risk value of the to-be-predicted road segment, wherein the cumulative risk value of the to-be-predicted road segment is calculated based on a crime opportunity of all criminals' potential landing points in the receptive field of the to-be-predicted road segment to the to-be-predicted road segment. The estimated probability regularization module is configured to perform regularization based on the case occurrence probability of the to-be-predicted road segment in a future week, the receptive field filter, the jurisdiction filter, the hotspot migration rule library, and the historical case event database, to obtain a corrected case occurrence probability of the to-be-predicted road segment in a future week. The scheduling module is configured to perform police force scheduling and patrol route planning based on the occurrence location information of the latest police case, the receptive field filter, the case occurrence probability of all road segments in each jurisdiction in a future week, the road network database, the road network calculation engine, and the real-time police force resource distribution database.

Images