A city space disorder measurement method based on street view big data and deep learning

Abstract

The application discloses a kind of city space disarray measure method based on street view big data and deep learning, belong to city management and information technology field, including: obtaining original street view image sequence from street view database, extracting initial disarray element features in image, calculating the local disarray index of single image;High-risk image is identified, and the local disarray index of high-risk image is aggregated, to determine the area disarray measure value of block level;Extract common features, and induct typical disarray mode category, construct mode knowledge base;Based on the time series street view image sequence of the same place, get the evolution trend sequence, match with the typical disarray mode category in mode knowledge base, and generate disarray evolution prediction result when matching degree is higher than preset threshold value;Early warning mechanism is triggered for disarray evolution prediction result, and output early warning signal.The application realizes the whole-link accurate measurement of city space disarray from static identification, dynamic analysis to active early warning.

Description

Technical Field

[0001] This invention belongs to the field of urban management and information technology, and in particular relates to a method for measuring urban spatial disorder based on street view big data and deep learning. Background Technology

[0002] Urban spatial disorder, such as cluttered public environments, damaged facilities, and illegal occupation of roads, directly impacts residents' quality of life, community safety, and the city's image. It is a core issue requiring quantifiable identification and proactive intervention in the refined governance of cities. With the widespread adoption of street view data collection vehicles and the rapid development of online map services, large-scale street view image data covering the entire city and multiple time periods has been achieved, providing an unprecedented data foundation for objectively and efficiently assessing urban spatial conditions using computer vision technology.

[0003] However, existing urban spatial assessment methods based on street view images have significant limitations. First, most methods focus on identifying and labeling disordered elements in individual images, such as detecting litter, graffiti, or illegally parked vehicles. These methods only provide isolated, static observations and cannot effectively aggregate massive, scattered single-point identification results into systematic measurement indicators at macro-scales such as blocks or administrative districts. This makes it difficult to conduct horizontal comparisons and quantitative rankings of the severity of disorder in different areas under the same standard.

[0004] Secondly, due to the lack of a systematic summary of disorder phenomena, existing technologies struggle to differentiate between different types of disorder in different regions and with different causes. For example, disorder in older residential areas may primarily manifest as damaged building facades and cluttered debris, while disorder in commercial districts may stem more from street vendors and chaotic advertising. Using the same set of universal identification standards for assessment will lead to distorted results, failing to accurately reflect the actual problems in various areas and thus hindering the development of differentiated governance strategies.

[0005] A more critical drawback is that existing methods primarily rely on street view "snapshots" at a single moment, failing to effectively utilize long-term accumulated temporal street view image data of the same location. Urban disorder is a dynamic evolutionary process, often taking months or even years to accumulate from sporadic appearances to concentrated outbreaks. Due to the lack of continuous analysis capabilities for temporal data, existing technologies cannot capture the spatial diffusion patterns and temporal deterioration trends of disorder, leaving urban management departments only able to passively respond to exposed problems, unable to provide early warnings and interventions in the early stages of problems' emergence or accelerated spread.

[0006] In summary, the main shortcomings of existing technologies are: the inability to effectively aggregate microscopic image recognition results into macroscopic, comparable spatial metrics; the inability to identify and summarize differentiated regional disorder patterns; and the inability to utilize time-series data to analyze the dynamic evolution of disorder and provide risk warnings. These bottlenecks limit the in-depth application value of street view big data in urban governance. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method for measuring urban spatial disorder based on street view big data and deep learning, comprising the following steps: Based on the original street view image sequence obtained from the street view database, the initial disordered element features in the images are extracted, and the local disorder index of a single image is calculated accordingly. High-risk images are identified based on the local disorder indicators, and the local disorder indicators of the high-risk images are aggregated to determine the regional disorder measure value at the street level. Common features are extracted from the disorder measure values ​​of multiple regions, and typical disorder pattern categories are summarized based on the common features to construct a pattern knowledge base; Based on a time-series street view image sequence of the same location, we analyze its changing trends to capture the evolutionary trend sequence; The evolution trend sequence is matched with typical disordered pattern categories in the pattern knowledge base, and disordered evolution prediction results are generated when the matching degree is higher than a preset threshold. An early warning mechanism is triggered based on the predicted results of disordered evolution, which locates specific risk areas and outputs early warning signals.

[0008] Optionally, the process of extracting initial disordered element features from the original street view image sequence obtained from the street view database, and calculating the local disorder index of a single image accordingly, includes: Multiple raw street view images are acquired through a preset image acquisition interface to form an image sequence; A convolutional neural network is used to extract features from each image in the image sequence to obtain an initial out-of-order element feature vector; Based on the initial disordered element feature vector, calculate the disorder density and disorder intensity values; The disorder density and disorder intensity values ​​are fused using feature engineering methods to generate the local disorder index for a single image.

[0009] Optionally, the process of using a convolutional neural network to extract features from each image in the image sequence includes: An image dataset is constructed from the acquired raw street view image sequences; Each image in the image dataset is processed using a pre-trained convolutional neural network model to extract feature vectors containing texture and edge information, forming an initial feature vector set. Perform sequence analysis on the feature vectors in the initial feature vector set, and determine and adjust the arrangement order of the feature vectors; Based on the adjusted feature vector sequence, the contextual correlation and continuity features between images are analyzed, and a structured representation of the image sequence is constructed. The final serialized feature data is generated based on the structured representation, serving as the initial out-of-order element feature vector.

[0010] Optionally, the process of identifying high-risk images based on the local disorder indicators and aggregating the local disorder indicators of the high-risk images to determine the area disorder measure value at the street level includes: Individual images whose local disorder indicators exceed a preset threshold are marked as high-risk images, forming a set of high-risk images; From the set of high-risk images, all high-risk images belonging to the same block are selected to form a high-risk subset of the block; The average disorder intensity of the block is calculated based on the local disorder index of each high-risk image within the high-risk subset of the block. By using a weighted summation method, the local disorder indicators of the high-risk subset of the block and the number of high-risk images are aggregated to obtain the overall disorder measure of the block, which is used as the regional disorder measure.

[0011] Optionally, the process of extracting common features from the regional disorder measure values ​​of multiple regions and summarizing typical disorder pattern categories based on the common features to construct a pattern knowledge base includes: Based on the regional disorder measure value of each region, calculate the feature vector of each regional sample under the same dimension to form an initial feature set; The initial feature set is clustered using a clustering algorithm, and cluster center points are assigned to each region sample to obtain a group label sequence; Calculate the distance from each region sample to its cluster center; Region samples whose distance values ​​exceed a preset multiple of the average distance to their respective cluster center points are marked as out-of-order points; Based on the correspondence between the group label sequence and the cluster center point, a mapping table between the pattern category and the center feature vector is constructed to form a typical pattern description structure; For regional samples of non-abnormal disordered points, the number of samples covered by each cluster center and the regional identifier are counted to obtain the pattern coverage distribution; The pattern category label, cluster center point feature vector, and pattern coverage distribution are combined into a complete pattern record and stored in the knowledge base to construct the pattern knowledge base.

[0012] Optionally, the process of analyzing the changing trends of time-series street view image sequences based on the same location to capture the evolutionary trend sequence includes: Arrange time-series street view images collected at the same location in chronological order to construct a complete time-series image sequence; Hierarchical feature extraction is performed on each frame of the time-series image to obtain a visual content description of each frame. Based on the visual content description, analyze the differences in visual elements between adjacent images; If the difference between the visual elements exceeds a preset difference threshold, the corresponding time point is marked as a change point, forming a set of change points; Extract the distribution pattern over time from the set of change points to construct a preliminary model of the change trend; Statistical analysis of the intervals and frequencies of the set of change points is performed to determine the periodic characteristics of the change trend; Long Short-Term Memory (LSTM) networks are used to perform deep learning on the time-series image sequences and the periodic features to capture long-term dependencies and obtain prediction results of evolutionary patterns. Based on the prediction results of the aforementioned evolutionary laws, the evolutionary trend sequence is generated.

[0013] Optionally, the process of matching the evolutionary trend sequence with typical disordered pattern categories in the pattern knowledge base, and generating disordered evolution prediction results when the matching degree is higher than a preset threshold, includes: The evolutionary trend sequence is standardized to obtain a normalized trend dataset; The normalized trend dataset is compared with the typical disordered pattern categories in the pattern knowledge base to calculate the degree of matching. If the matching degree exceeds the preset matching threshold, preliminary evolutionary prediction data is generated based on the comparison results, and the disordered pattern category of the matching is recorded. The preliminary evolution prediction data is refined to obtain a more refined prediction output. The refined prediction output is associated with the matched disordered pattern category to generate the final disordered evolution prediction result. The disordered evolution prediction results are stored in the pattern knowledge base.

[0014] Optionally, the process of triggering an early warning mechanism based on the predicted disordered evolution results, and locating specific risk areas and outputting early warning signals, includes: The disordered evolution prediction results data are obtained and compared with a preset rule base to determine whether there are abnormal fluctuations; If abnormal fluctuations are detected, the corresponding evolutionary state information is recorded to form a preliminary abnormality marker; The preliminary anomaly markers are filtered based on preset triggering conditions. If a marker meets the triggering conditions, an early warning activation command is generated. Based on the warning activation command, extract potential risk characteristics from the abnormal data; The potential risk characteristics are analyzed using a risk localization algorithm to determine the specific risk area range; Detailed evolution status information within the risk area is obtained, and the priority of each risk area is determined by data comparison to obtain a list of high-priority risk areas; For the list of high-priority risk areas, early warning signal content is constructed and formatted to obtain standardized early warning signals; The standardized early warning signal is distributed to the corresponding risk area.

[0015] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0016] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0017] Compared with the prior art, the present invention has the following advantages and technical effects: This invention addresses the unique business scenario in urban management where it is difficult to detect disorder in streetscape environments in real time and predict their deterioration trends. It deeply integrates single street view image disorder detection with multi-temporal sequence evolution analysis. First, a convolutional neural network is used to extract image features and quantify local disorder indicators, aggregating them to form a street-level disorder measure. Then, a clustering algorithm is used to extract typical disorder patterns and construct a knowledge base. Subsequently, a long short-term memory network is used to capture the changing trends of temporal images of the same location and match them with patterns in the knowledge base to achieve accurate evolution prediction. Once the matching degree exceeds a threshold, risk location and early warning are triggered. This invention, through a closed-loop mechanism of feature-level disorder quantification, pattern knowledge base, and temporal trend matching, achieves a complete link from static disorder identification to dynamic evolution prediction, effectively improving urban governance's ability to early perceive and accurately intervene in disorder issues such as environmental degradation and public security risks. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the process for determining the street status according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the generation of a pattern knowledge base according to an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] Example 1 like Figure 1 As shown in the figure, this embodiment provides a method for measuring urban spatial disorder based on street view big data and deep learning, including the following steps: Based on the original street view image sequence obtained from the street view database, the initial disordered element features in the images are extracted, and the local disorder index of a single image is calculated accordingly. High-risk images are identified based on the local disorder indicators, and the local disorder indicators of the high-risk images are aggregated to determine the regional disorder measure value at the street level. Common features are extracted from the disorder measure values ​​of multiple regions, and typical disorder pattern categories are summarized based on the common features to construct a pattern knowledge base; Based on a time-series street view image sequence of the same location, we analyze its changing trends to capture the evolutionary trend sequence; The evolution trend sequence is matched with typical disordered pattern categories in the pattern knowledge base, and disordered evolution prediction results are generated when the matching degree is higher than a preset threshold. An early warning mechanism is triggered based on the predicted results of disordered evolution, which locates specific risk areas and outputs early warning signals.

[0022] The feasible process of acquiring multiple raw street view image sequences from a street view database through a pre-defined image acquisition interface, and then using a convolutional neural network to extract features from each image to obtain an initial out-of-order element feature vector includes: Multiple raw street view image sequences are acquired from a street view database through a pre-defined data interface to construct an initial image dataset. Based on the acquired raw street view image sequences, a convolutional neural network is used to extract features from each image, resulting in an initial feature vector set. For out-of-order elements in the initial feature vector set, sequence analysis is performed to determine the order of the feature vectors. If the order of the feature vectors is inconsistent, a pre-defined threshold is used to determine the distribution of out-of-order elements, resulting in an adjusted feature vector sequence. Based on the adjusted feature vector sequence, the contextual relationships between image sequences are analyzed to determine the continuity features of each image in the sequence. Using the analysis results of continuous features, a structured representation of the image sequence is constructed to determine the final image sequence feature description. After obtaining the final image sequence feature description, corresponding serialized feature data is generated, completing the feature processing flow of the street view images.

[0023] Feasible. In the street view image processing workflow, the first step is to obtain multiple raw image sequences from the street view database. Assume that 100 consecutive street view images of a certain street are obtained through an interface, each image with a resolution of 1920x1080 pixels. These images are initially sorted by timestamp, but some may be out of order or missing. After obtaining the images, an initial dataset is constructed to lay the foundation for subsequent feature extraction.

[0024] For feature extraction, a convolutional neural network is used to process each image. Assuming a pre-trained ResNet model is used, each image is transformed into a 2048-dimensional feature vector. After processing 100 images, an initial set of 100 feature vectors is obtained. These vectors capture visual information such as texture and edges in the images, providing data support for subsequent analysis. The advantage of this approach is that it efficiently extracts deep features and reduces the cost of manual annotation.

[0025] During the sequence analysis phase, out-of-order elements in the initial feature vector set are ordered using timestamps and spatial location information. For example, if the feature vector timestamps of the 10th and 11th images are found to be abnormal, possibly due to a temporary malfunction in the acquisition equipment, a time threshold, such as 5 seconds, is set. If the time difference between two images exceeds this threshold, they are marked as out of order. In this way, the adjusted feature vector sequence more closely matches the actual street scene order, helping to improve the accuracy of subsequent analysis.

[0026] When analyzing contextual relevance, the similarity of feature vectors between adjacent images is calculated based on the adjusted feature vector sequence. For example, if the cosine similarity of the feature vectors of images 20 and 21 is 0.95, it indicates high content continuity, while the similarity between images 30 and 31 is only 0.6, potentially indicating abrupt scene changes. By judging continuity features, breakpoints or anomalies in the sequence can be effectively identified, improving the quality of the structured representation of image sequences.

[0027] When constructing the structured representation, continuous features are transformed into a graph structure, where nodes represent images and edges represent the strength of continuity. It is assumed that adjacent images with a similarity greater than 0.8 are connected as strongly related edges, while those with a similarity less than 0.6 are labeled as weakly related edges. This representation intuitively reflects the logical relationships of street scene sequences, facilitating subsequent applications such as path planning or scene reconstruction.

[0028] When finally generating serialized feature data, the structured representation is encoded into a file containing temporal, spatial, and feature information. It is assumed that the file records the location coordinates and continuity labels of each image, facilitating rapid retrieval and retrieval by the system. This processing flow significantly improves the usability of street view data, providing reliable support for applications such as urban modeling and traffic analysis.

[0029] The feasible process of calculating disorder density and intensity values ​​based on the initial disorder element feature vectors, and then fusing these values ​​using feature engineering methods to generate a local disorder index for a single image includes: A basic dataset is constructed by obtaining feature vector data from initial elements. Preprocessing methods are used to standardize the data, resulting in a normalized set of feature vectors. Based on this normalized set, disorder density and intensity values ​​are calculated, and a pre-defined threshold is used for initial screening. If the calculated disorder density exceeds the threshold, the data is marked to identify anomalous feature regions. For these marked anomalous feature regions, feature engineering methods are used to integrate the disorder density and intensity values, obtaining comprehensive disorder parameters to determine if they meet pre-defined local indicator standards. By combining the comprehensive disorder parameters with pixel distribution data from a single image, a mapping model for local indicators is constructed. A support vector machine algorithm is used to classify the local indicators, resulting in a classified distribution of disordered regions. Based on this classified distribution, the changing trends of local indicators in a single image are analyzed, and disorder features of key regions are extracted to determine their priority in image analysis. Based on the priority ranking, targeted processing strategies are generated. If the disorder features of key regions are concentrated within a specific range, in-depth analysis is performed to obtain detailed descriptions of the disorder features. By using the disordered feature descriptions obtained through deep analysis, the final output data of image analysis is constructed, and it is determined whether it meets the preset analysis accuracy requirements to obtain complete local disorder index results.

[0030] It is feasible to start from the original feature vector set and first standardize the data through preprocessing methods in the out-of-order feature processing flow of street view image sequences.

[0031] Understandably, this process mainly unifies the feature vectors extracted from different street view images to the same scale, avoiding numerical deviations caused by differences in lighting and angle.

[0032] In one possible implementation, zero-mean standardization is used, which adjusts the mean of each feature vector to 0 and the variance to 1, thereby obtaining a normalized set of feature vectors. This operation helps to ensure the fairness of subsequent out-of-order density calculations.

[0033] Specifically, when calculating the disorder density and intensity values ​​of the feature vector sequence after normalization, the cosine similarity between adjacent vectors can be selected as the basic index.

[0034] For example, when the feature vector similarities of five consecutive street view images are 0.92, 0.89, 0.31, 0.87, and 0.91 respectively, there is a significant decrease between the third and fourth images. In this case, the disorder density can be defined as the number of jumps within a unit sequence where the similarity is below 0.5, and the intensity is the average jump amplitude. In this example, the disorder density is 1, and the intensity is approximately 0.58. If the preset threshold density is 0.3 and the intensity is 0.4, then this sequence exceeds the threshold range, and the system immediately marks it, forming an abnormal feature region.

[0035] It should be noted that the marked abnormal regions will enter the feature engineering integration stage.

[0036] In one embodiment, disorder density and intensity are weighted and fused into a comprehensive disorder parameter, for example, with weights of 0.6 and 0.4 respectively, resulting in a comprehensive parameter of 0.512. Subsequently, a local index mapping model is constructed by combining single-image pixel distribution data, such as edge density or color histogram differences. Support vector machines are then used for classification, classifying regions with a comprehensive parameter higher than 0.45 as high disorder regions and vice versa. Through this classification, the system accurately locates abnormal frames in street scene sequences caused by occlusion or abrupt viewpoint changes.

[0037] Preferably, based on the disordered regional distribution after classification, the trend of local indicator changes is further analyzed.

[0038] For example, high disorder areas are concentrated in frames 12 to 18 of the sequence. Pixel distribution shows that a large number of trees occlude this segment, causing abrupt changes in the foreground. At this point, disorder features of key areas are extracted, such as a 32% increase in texture entropy and a 28% increase in color dispersion. Based on this, priority ranking is determined, placing occlusion-related disorder with the highest priority. Targeted processing strategies are generated for the highest priority areas, such as performing deep analysis on this segment and re-representing it through multi-scale gradient features to obtain a more detailed description of the disorder, such as "foreground tree occlusion caused a 70% loss of main road features." The final image analysis output data meets the preset analysis accuracy requirements (such as an overall disorder recognition rate of over 96%), significantly improving the robustness and continuous understanding ability of street view sequences in tasks such as autonomous driving map updates and urban change detection, ensuring that subsequent path planning modules will not cause global misjudgments due to single-frame anomalies.

[0039] Implementable, if a local disorder index exceeds a preset threshold, the image is marked as high-risk. The process of determining the block-level regional disorder measure value by aggregating the local disorder indices of high-risk images includes: If the local disorder index of an image exceeds a preset threshold, the image is marked as a high-risk image, resulting in a high-risk image set. From this set, all high-risk images belonging to the same block are selected to form a high-risk subset for that block. The average local disorder index of each high-risk image within this subset is calculated to obtain the average disorder intensity of the block. The local disorder index of the high-risk subset is then aggregated with the number of high-risk images using a weighted summation method to obtain the overall disorder measure of the block. If the overall disorder measure of a block remains higher than the historical average disorder intensity for the same period for several days, the block is determined to have entered a high-disorder state.

[0040] Feasible, such as Figure 2 As shown, when analyzing local disorder indicators in images, we can start from the actual scenario of street surveillance to explore how to identify and manage high-risk areas. If the local disorder indicator of an image exceeds a preset threshold (assuming a threshold of 5.0), and the indicator reaches 6.2, then it is marked as a high-risk image. This marking method can help quickly identify potentially problematic images and form a set of high-risk images. In implementation, an automated system can batch scan the images collected daily, comparing the indicator values ​​one by one to filter out abnormal images.

[0041] To form a high-risk subset of images from a collection of high-risk images, we can imagine a specific block, such as Block A, where 10 images are marked as high-risk on a certain day. These images are categorized into Block A based on their geographic location information, forming a subset. In implementation, location tags from the image metadata can be used to quickly group images through database queries, ensuring accurate classification.

[0042] When calculating the average disorder intensity of a block, assuming the local disorder indices of 10 images within the high-risk subset of Block A are 6.2, 5.8, 6.0, 5.5, 6.1, 5.9, 6.3, 5.7, 6.4, and 5.6 respectively, the system automatically calculates the average, resulting in an average disorder intensity of approximately 6.0. This value reflects the overall degree of disorder in the block, providing foundational data for subsequent analysis. In practice, calculations can be performed using simple statistical tools, ensuring real-time updates of the results.

[0043] For the step of aggregating local disorder indicators and the number of high-risk images using a weighted summation method to obtain the overall disorder measure of a block, the weight of the local disorder indicator can be set to 0.7, and the weight of the number of high-risk images to 0.3. Taking Block A as an example, with an average disorder intensity of 6.0 and 10 images, the weighted overall disorder measure is approximately 5.1. This method can comprehensively consider the degree of disorder and the scope of impact, and the weight ratio can be adjusted according to the characteristics of the block during implementation.

[0044] If the overall disorder measure of a neighborhood remains higher than the historical average for the same period for several consecutive days, it is determined to be in a high disorder state. For example, if the historical average disorder measure of neighborhood A over the past 30 days is 4.5, and the overall disorder measure for the past 5 days has been above 5.1, the system will automatically trigger a high disorder state warning. In implementation, a time window can be set, such as a 5-day consecutive period, and dynamic comparisons can be made with historical data. This approach helps to promptly identify potential persistent problems within a neighborhood, providing early warning support for subsequent management. Through the above multi-faceted analysis and examples, it can be seen that the process from a single image to the overall disorder assessment of a neighborhood is logically rigorous, with each step supporting the others and jointly providing a basis for accurate judgment of the neighborhood's status. This method can effectively improve monitoring efficiency in practical applications and assist in the implementation of relevant management work.

[0045] The feasible process of extracting common features from regional disorder measures, using clustering algorithms to group similar features into typical disorder pattern categories, and storing them in a knowledge base to obtain a pattern knowledge base includes: like Figure 3As shown, feature vectors of all samples in a unified dimension are calculated based on the regional disorder measure value to obtain an initial feature set. K-means clustering algorithm is used to perform clustering operations on the initial feature set to determine the cluster centroid of each sample, resulting in a group label sequence. The distance from each sample to its cluster centroid is calculated using Euclidean distance to obtain a distance sequence. If the distance value of a single sample in the distance sequence exceeds a preset multiple of the average distance to the cluster centroids, the sample is marked as an outlier, resulting in an outlier set. Based on the correspondence between the group label sequence and the cluster centroids, a mapping table between pattern categories and centroid feature vectors is constructed to obtain a typical pattern description structure. For samples within the non-outlier set, the number of samples and region identifiers contained in each cluster centroid are counted to obtain the pattern coverage distribution. The pattern category labels, cluster centroid feature vectors, and coverage distribution together form a complete pattern record, which is stored in the pattern knowledge base to obtain the final pattern knowledge base.

[0046] It is feasible to use image data of a city block as a sample during the analysis of regional disorder measures and the construction of feature vectors. Each sample contains disorder indicators from different locations within the block, such as environmental clutter and facility damage rates. These indicators are quantified into numerical values; for example, a block might have a clutter level of 0.75 and a facility damage rate of 0.62, forming a multi-dimensional feature vector. In this way, the feature vectors of all block samples are aggregated to form an initial feature set, laying the foundation for subsequent analysis.

[0047] When using clustering algorithms to group a feature set, we can envision performing a clustering operation on 100 street samples, setting the number of cluster centroids to 5. After running the algorithm, each street sample is assigned a centroid, forming a group label sequence. Suppose one centroid represents a "low disorder pattern," with a feature vector that might have a noise level of 0.2 and a damage rate of 0.1, while another centroid represents a "high disorder pattern," with a feature vector that has a noise level of 0.8 and a damage rate of 0.7. This grouping method helps identify typical patterns of street disorder.

[0048] When calculating the distance from a sample to the center point and identifying outliers, anomalies can be determined by comparing the distance values ​​of each block sample to its respective center point. For example, if the average distance to a center point is 0.3, and the distance value of a block sample is 0.9, exceeding a preset multiple such as twice, it is marked as an outlier. These outliers may reflect sudden problems or data deviations within the block; marking them facilitates subsequent focused monitoring and analysis.

[0049] When constructing the mapping table between pattern categories and central feature vectors, the five cluster centers can be mapped to pattern categories such as "low disorder," "medium disorder," and "high disorder," and the feature vector value of each center point can be recorded. This mapping table forms a typical pattern description structure, intuitively showing the feature differences of different disorder patterns, which helps to quickly identify the status of a street.

[0050] When analyzing the coverage distribution of non-outlier samples, it was found that the center point of a certain "low disorder pattern" covered 30 street blocks, involving the northern part of the city, while the "high disorder pattern" covered 10 street blocks, mostly concentrated in older areas. This distribution analysis provides a basis for subsequent resource allocation, such as prioritizing high disorder areas.

[0051] When constructing and storing complete pattern records in the knowledge base, information such as pattern category labels (e.g., "high disorder pattern"), center point feature vectors (e.g., disorder level 0.8, damage rate 0.7), and coverage distribution (e.g., involving 10 city blocks) can be integrated into a single record. This recording method ensures the comprehensiveness and traceability of the knowledge base, provides data support for long-term tracking of changes in the status of city blocks, and helps improve management efficiency.

[0052] It is feasible to acquire time-series street view image sequences of the same location and use long short-term memory networks to analyze the changes in the time-series street view image sequences to capture the evolutionary trend sequence.

[0053] By batch reading street view data from the same location, the time-series images are arranged chronologically to construct a complete image sequence. For the arranged image sequence, hierarchical feature extraction is performed to obtain a visual content description for each frame. A pre-established feature extraction module is used to analyze the content of each image in the sequence, identifying its key visual elements. Based on the visual content description, the differences between adjacent images are analyzed to determine if significant changes exist. If the visual element differences between adjacent images exceed a preset threshold, they are marked as change points, resulting in a set of change points. The distribution patterns along the time dimension are extracted from the change point set to construct a preliminary model of the change trend. Statistical analysis of the intervals and frequencies of change points determines the periodicity of the change trend. For the periodicity of the change trend, a Long Short-Term Memory (LSTM) network is applied for deep learning to capture long-term dependencies, obtaining predictions of the evolutionary pattern. Through layer-by-layer processing of the time-series data, a dynamic representation of the evolutionary pattern is constructed. Based on this dynamic representation, street view changes over a future period are extrapolated, generating a prediction sequence. By comparing the prediction sequence with historical image sequences, the reliability of the prediction results is assessed, yielding the final description of the evolutionary trend.

[0054] It is feasible to monitor the disordered evolution of urban blocks by batch reading street scene time-series images captured by the same fixed camera. This allows for the precise sorting of 144 images taken every 10 minutes within a day according to their timestamps, forming a continuous visual sequence. This orderly arrangement lays the foundation for subsequent capture of dynamic changes.

[0055] Specifically, when using the hierarchical feature extraction module, a pre-trained visual model is first used to extract basic features such as low-level edges and color distribution from each frame of the image. Then, an attention mechanism is used to obtain high-level semantic descriptions, such as "the amount of litter piled up on the road," "the proportion of non-motorized vehicles parked haphazardly," and "pedestrian density," forming a standardized visual content vector. This multi-level description can more comprehensively represent the state of street order.

[0056] In one embodiment, the cosine similarity of visual vectors between adjacent frames is calculated. If the similarity suddenly drops from 0.92 to 0.67, which is far below the daily fluctuation threshold of 0.85, it is determined to be a significant change point. For example, between 8 pm and 9 pm on Friday, if the image shows a sudden increase in street vendors causing a sharp drop in similarity, it is marked as a change point, thereby accurately capturing the moment when the disorderly event occurs.

[0057] It should be noted that statistical analysis of the temporal distribution of all change points revealed a peak approximately every 72 hours in a certain old community, exhibiting a clear periodicity. Further inputting the change point interval sequence into a Long Short-Term Memory (LSTM) network allowed the network to learn a long-term dependency pattern of "Friday night market → Monday cleanup → Wednesday recovery," achieving a prediction accuracy of over 85% for the disorder intensity of the following week after training.

[0058] Preferably, forward extrapolation based on the learned dynamic representation can generate a virtual street view sequence for the next 48 hours. For example, it can predict that the garbage accumulation will increase from the current 12% to 38% by 10 pm on Friday. The prediction is then compared with the actual subsequent collected images, and the deviation is only 4.7%, thus providing the urban management department with an accurate early warning window of 6-12 hours in advance.

[0059] Understandably, the evolutionary trend description formed through the complete link from image sorting to deep temporal modeling and then to reliability verification not only reveals the periodic triggering mechanism of regional disorder, but also quantifies the future risk level, transforming passive patrols into proactive prevention, and significantly improving the efficiency of street governance and the accuracy of resource allocation.

[0060] The feasible process of matching evolutionary trend sequences with typical disordered pattern categories in a pattern knowledge base, and generating disordered evolution prediction results if the matching degree is higher than a preset threshold, includes: The system acquires input evolutionary trends and sequence data, which are then standardized using a data preprocessing module to obtain a normalized trend dataset. For this normalized dataset, out-of-order patterns and typical categories are extracted from a pre-established pattern knowledge base. A comparison algorithm is used for pattern matching to determine the degree of matching. If the degree of matching exceeds a preset threshold, preliminary evolutionary prediction data is generated based on the comparison process, recording the corresponding out-of-order pattern category. Based on the preliminary evolutionary prediction data and category judgment logic, the prediction results are refined to obtain a refined prediction output. Using the refined prediction output and a result generation module, out-of-order patterns are associated and mapped with evolutionary trends to generate the final out-of-order evolutionary prediction result. The final out-of-order evolutionary prediction result is archived in the knowledge base using a storage module for subsequent querying and updates.

[0061] Feasible, standardization in the data preprocessing module is particularly important when dealing with the evolutionary trends and sequence data of street view images. Standardization aims to unify street view data collected at different time points to a common scale range for subsequent analysis. For example, assuming that a city's street view data includes multiple indicators such as brightness, color saturation, and object density, standardization will adjust the numerical range of these indicators to between 0 and 1, eliminating dimensional differences. Specifically, this can be achieved by linearly normalizing the brightness values, mapping the original values ​​to a standard range, ensuring the comparability of data from different time periods.

[0062] For standardized trend datasets, when extracting out-of-order patterns and typical categories from a pattern knowledge base, we can envision a knowledge base containing multiple street scene change patterns, such as road widening, increased greenery, or building renovation. Pattern matching can be achieved through comparison algorithms.

[0063] Specifically, the standardized street view data is compared one by one with the pattern features in the knowledge base, and a similarity score is calculated. If the matching score between a street view data and the "increased greenery" pattern reaches 0.85, exceeding the preset threshold of 0.8, it is determined to be a preliminary match, and the pattern category is recorded. This method helps to quickly locate the potential patterns of street view changes.

[0064] After generating initial evolutionary prediction data, the data is refined using category-based judgment logic. This refinement process optimizes the prediction results by analyzing time intervals and frequency of change in historical street view data. For example, assuming a street view area experiences a significant change in its greening increase pattern every two years, the refinement process adjusts the prediction data according to this periodic pattern, ensuring the predictions more closely reflect actual trends. This refinement process improves prediction accuracy and provides a more reliable basis for subsequent decision-making.

[0065] When mapping the refined prediction output to disordered patterns, this can be achieved through a result generation module. For example, combining the "increased greening" pattern with street view evolution trends can generate a prediction report indicating potential changes in greening in the area over the next three years. This mapping can be implemented using a rule-based reasoning system, logically binding pattern categories with time trends to form intuitive prediction results.

[0066] When the final disordered evolution prediction results are archived to the knowledge base, the storage module can classify and store them according to time, region, and pattern category. For example, if the prediction results for a city's street view are archived as "2023 - Increased Greenery - A Street," subsequent queries can quickly retrieve them using keywords. This storage method facilitates data management and updates, providing historical references for future street view analyses. Through meticulous processing at each stage, the accuracy and practicality of street view evolution trend predictions can be effectively improved, providing crucial support for urban planning and environmental monitoring.

[0067] An feasible process for triggering an early warning mechanism based on disordered evolution prediction results, which involves locating specific risk areas and outputting early warning signals using a risk localization algorithm, includes: The system acquires predicted data of disordered evolution and compares the evolutionary state with a pre-established rule base to determine if abnormal fluctuations exist. If abnormal fluctuations are detected, the corresponding evolutionary state information is recorded, resulting in preliminary anomaly markers. These preliminary anomaly markers are then filtered using preset trigger conditions. If a marker meets the trigger conditions, an activation command for the early warning mechanism is generated, identifying the abnormal data requiring further processing. Based on the activation command, potential risk features are extracted from the abnormal data, and a risk localization algorithm is used to analyze the potential risks and determine the specific risk area range. For the determined risk area range, evolutionary state details within the area are acquired, and the priority of area identification is determined using a data comparison tool, resulting in a list of high-priority risk areas. For the high-priority risk area list, a signal generation module constructs early warning signal content, and a preset signal formatting tool generates standardized early warning signal outputs. Based on the standardized early warning signal outputs, an information distribution module maps the signal content to the corresponding risk areas, completing the accurate association and transmission of the early warning signals.

[0068] It is feasible to immediately acquire the result data and enter the abnormal fluctuation detection stage after the disordered evolution prediction result is generated. First, the state is compared through a pre-established rule base, which contains threshold templates for various historical disorder cases, such as a sudden change in trend slope exceeding 15% or a fluctuation amplitude reaching 2.5 times the normal value in a short period of time.

[0069] Specifically, when the prediction results show that the attention to a social event suddenly increases from an average daily growth of 8% to 45% over three consecutive cycles, the system quickly identifies it as an abnormal fluctuation, records the current evolution status information, and forms a preliminary abnormality marker.

[0070] In one possible implementation, the initial anomaly marker includes a timestamp, intensity level, and associated pattern category. Then, preset trigger conditions are used for filtering; for example, an alert mechanism activation command is only generated when the anomaly intensity level is higher than medium and the duration exceeds a preset number of cycles. This effectively filters noise and avoids false alarms.

[0071] If only a single small fluctuation occurs, such as a brief jump of 12%, but without a trend reversal, it will not trigger, further demonstrating the accuracy of the screening.

[0072] It should be noted that after the activation command is triggered, the system extracts potential risk characteristics from abnormal data, such as the number of nodes accelerating the spread and sudden changes in sentiment polarity. Through risk localization algorithms, typically based on density clustering, the risk area is then pinpointed to a specific segment of the propagation path.

[0073] In one instance of public opinion evolution, the algorithm discovered that the risk was mainly concentrated among users in second-tier cities, which contributed 72% of the abnormal increase, thus determining the specific risk area.

[0074] After identifying the risk areas, the system obtains detailed evolutionary status within those areas, such as the depth of the forwarding chain and the activity level of key opinion leaders. Priority scores are calculated using data comparison tools, with rules set to a weight of 0.6 for forwarding growth rate and 0.4 for the proportion of negative sentiment. Areas scoring above 85 are added to the high-priority list. In a real-world example, the scores for the three areas were 92, 78, and 66 respectively, with only the former ultimately making the list, demonstrating the efficiency of priority identification.

[0075] In one embodiment, for the high-priority list, the signal generation module automatically selects a template based on the disorder pattern category to construct an early warning signal that includes the risk level, expected peak time, and suggested intervention points. This is then standardized using a formatting tool to include fields such as event ID, risk coordinates, and 92% confidence level, ensuring cross-system compatibility. Finally, the standardized early warning signal is accurately pushed to the management interface or decision-making terminal of the corresponding risk area through the information distribution module.

[0076] In the aforementioned public opinion cases, the signals were specifically pushed to the regional content moderation teams, enabling them to intervene in key accounts six hours before the incident spiraled out of control. This effectively controlled the spread to within 63% of the expected peak, significantly reducing the negative impact of the disorderly evolution. This closed-loop mechanism, from detection to precise transmission, greatly improves the timeliness and targeting of early warnings.

[0077] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0078] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0079] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for measuring urban spatial disorder based on street view big data and deep learning, characterized in that, Includes the following steps: Based on the original street view image sequence obtained from the street view database, the initial disordered element features in the images are extracted, and the local disorder index of a single image is calculated accordingly. High-risk images are identified based on the local disorder indicators, and the local disorder indicators of the high-risk images are aggregated to determine the regional disorder measure value at the street level. Common features are extracted from the disorder measure values ​​of multiple regions, and typical disorder pattern categories are summarized based on the common features to construct a pattern knowledge base; Based on a time-series street view image sequence of the same location, we analyze its changing trends to capture the evolutionary trend sequence; The evolution trend sequence is matched with typical disordered pattern categories in the pattern knowledge base, and disordered evolution prediction results are generated when the matching degree is higher than a preset threshold. The predicted results of disordered evolution trigger an early warning mechanism, which locates specific risk areas and outputs early warning signals. The process of identifying high-risk images based on the local disorder indicators and aggregating the local disorder indicators of high-risk images to determine the area disorder measure value at the street level includes: Individual images whose local disorder indicators exceed a preset threshold are marked as high-risk images, forming a set of high-risk images; From the set of high-risk images, all high-risk images belonging to the same block are selected to form a high-risk subset of the block; The average disorder intensity of the block is calculated based on the local disorder index of each high-risk image within the high-risk subset of the block. By using a weighted summation method, the local disorder indicators and the number of high-risk images of the high-risk subset of the block are aggregated to obtain the overall disorder measure of the block, which is used as the disorder measure of the region. The process of extracting initial disordered element features from the original street view image sequence obtained from the street view database, and calculating the local disorder index of a single image accordingly, includes: Multiple raw street view images are acquired through a preset image acquisition interface to form an image sequence; A convolutional neural network is used to extract features from each image in the image sequence to obtain an initial out-of-order element feature vector; Based on the initial disordered element feature vector, calculate the disorder density and disorder intensity values; The disorder density and disorder intensity values ​​are fused using feature engineering methods to generate the local disorder index for a single image.

2. The method according to claim 1, characterized in that, The process of extracting features from each image in the image sequence using a convolutional neural network includes: An image dataset is constructed from the acquired raw street view image sequences; Each image in the image dataset is processed using a pre-trained convolutional neural network model to extract feature vectors containing texture and edge information, forming an initial feature vector set. Perform sequence analysis on the feature vectors in the initial feature vector set, and determine and adjust the arrangement order of the feature vectors; Based on the adjusted feature vector sequence, the contextual correlation and continuity features between images are analyzed, and a structured representation of the image sequence is constructed. The final serialized feature data is generated based on the structured representation, serving as the initial out-of-order element feature vector.

3. The method according to claim 1, characterized in that, The process of extracting common features from the disorder measure values ​​of multiple regions and summarizing typical disorder pattern categories based on the common features to construct a pattern knowledge base includes: Based on the regional disorder measure value of each region, calculate the feature vector of each regional sample under the same dimension to form an initial feature set; The initial feature set is clustered using a clustering algorithm, and cluster center points are assigned to each region sample to obtain a group label sequence; Calculate the distance from each region sample to its cluster center; Region samples whose distance values ​​exceed a preset multiple of the average distance to their respective cluster center points are marked as out-of-order points; Based on the correspondence between the group label sequence and the cluster center point, a mapping table between the pattern category and the center feature vector is constructed to form a typical pattern description structure; For regional samples of non-abnormal disordered points, the number of samples covered by each cluster center and the regional identifier are counted to obtain the pattern coverage distribution; The pattern category label, cluster center point feature vector, and pattern coverage distribution are combined into a complete pattern record and stored in the knowledge base to construct the pattern knowledge base.

4. The method according to claim 1, characterized in that, The process of analyzing the changing trends of the time-series street view image sequence based on the same location to capture the evolutionary trend sequence includes: Arrange time-series street view images collected at the same location in chronological order to construct a complete time-series image sequence; Hierarchical feature extraction is performed on each frame of the time-series image to obtain a visual content description of each frame. Based on the visual content description, analyze the differences in visual elements between adjacent images; If the difference between the visual elements exceeds a preset difference threshold, the corresponding time point is marked as a change point, forming a set of change points; Extract the distribution pattern over time from the set of change points to construct a preliminary model of the change trend; Statistical analysis of the intervals and frequencies of the set of change points is performed to determine the periodic characteristics of the change trend; Long Short-Term Memory (LSTM) networks are used to perform deep learning on the time-series image sequences and the periodic features to capture long-term dependencies and obtain prediction results of evolutionary patterns. Based on the prediction results of the aforementioned evolutionary laws, the evolutionary trend sequence is generated.

5. The method according to claim 1, characterized in that, The process of matching the evolutionary trend sequence with typical disordered pattern categories in the pattern knowledge base, and generating disordered evolution prediction results when the matching degree is higher than a preset threshold, includes: The evolutionary trend sequence is standardized to obtain a normalized trend dataset; The normalized trend dataset is compared with the typical disordered pattern categories in the pattern knowledge base to calculate the degree of matching. If the matching degree exceeds the preset matching threshold, preliminary evolutionary prediction data is generated based on the comparison results, and the disordered pattern category of the matching is recorded. The preliminary evolution prediction data is refined to obtain a more refined prediction output. The refined prediction output is associated with the matched disordered pattern category to generate the final disordered evolution prediction result. The disordered evolution prediction results are stored in the pattern knowledge base.

6. The method according to claim 1, characterized in that, The process of triggering an early warning mechanism based on the disordered evolution prediction results, and locating specific risk areas and outputting early warning signals, includes: The disordered evolution prediction results data are obtained and compared with a preset rule base to determine whether there are abnormal fluctuations; If abnormal fluctuations are detected, the corresponding evolutionary state information is recorded to form a preliminary abnormality marker; The preliminary anomaly markers are filtered based on preset triggering conditions. If a marker meets the triggering conditions, an early warning activation command is generated. Based on the warning activation command, extract potential risk characteristics from the abnormal data; The potential risk characteristics are analyzed using a risk localization algorithm to determine the specific risk area range; Detailed evolution status information within the risk area is obtained, and the priority of each risk area is determined by data comparison to obtain a list of high-priority risk areas; For the list of high-priority risk areas, early warning signal content is constructed and formatted to obtain standardized early warning signals; The standardized early warning signal is distributed to the corresponding risk area.

7. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-6.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-6.

Images