A method and system for monitoring dynamic changes of objects based on UAV images

Abstract

This invention provides a method and system for monitoring the dynamic changes of objects based on UAV images, belonging to the field of image analysis technology. The method includes: acquiring an image dataset of a target area and performing target tracking processing to obtain multiple trajectory segments; identifying multiple scene structural components and performing rhythm analysis on the target area to construct rhythmic profiles of the multiple scene structural components; collecting real-time monitoring data of the target area and performing anomaly detection to generate structural behavior anomaly detection data for the target area; constructing an interaction graph and dynamic flow constraint model for the target area; extracting behavior density data and behavior blockage data of the target area and identifying multiple local abnormal flow areas in the target area; identifying the diffusion impact of the local abnormal flow areas to determine multiple flow anomaly areas; and fusing the structural behavior anomaly detection data to generate dynamic change monitoring results for the target area. This invention improves the real-time perception capability of the gradual accumulation of risks in the dynamic behavior of objects.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image analysis, in particular to a method and system for monitoring dynamic changes of objects based on images captured by a UAV. BACKGROUND

[0002] The UAV device has important value in the security monitoring field of modern parks, large hotels and comprehensive commercial bodies due to its flexible maneuverability, wide field of view, and the ability to cover blind spots. The high-altitude images captured by the UAV can intuitively reflect the distribution and movement of targets such as personnel, service vehicles, and visitor vehicles within the region, providing convenient conditions for security situation awareness and abnormal behavior early warning. Existing UAV image analysis techniques mainly use detection and tracking as the main means, which are often used to identify the location and number of specific targets to determine whether there is an abnormal gathering, illegal parking or passage blocking phenomenon in the region at present. However, some methods are limited to static or short-time state identification, and the results are usually static conclusions such as whether there is a crowd gathering or whether there is a vehicle illegally parked, which is difficult to reflect the dynamic change trend of the target activity in the region and the potential gradual risk accumulation process.

[0003] In actual scenarios, the security state of a region often has a certain gradualness and rhythm. For example, at the entrance of a park or a large event site, even if there is no obvious congestion, there may be early signs of abnormal increase in personnel density, increase in the number of people staying in a specific area, sudden drop in vehicle flow speed, or imbalance in traffic flow in a specific passage within a short period of time. These phenomena may be precursors of potential abnormalities. If the monitoring system only focuses on whether the gathering or congestion has occurred significantly, and ignores such dynamic change characteristics, the risk accumulation process may be missed, resulting in delayed early warning or lagging monitoring results. SUMMARY

[0004] To solve the above technical problems, the present application provides a method and system for monitoring dynamic changes of objects based on images captured by a UAV, which extracts information reflecting the dynamic evolution of the region's state from continuous image data collected by the UAV, realizes early identification of abnormal change trends, and realizes intelligent monitoring of target dynamic changes.

[0005] The first aspect of the present application provides a method for monitoring dynamic changes of objects based on images captured by a UAV, comprising:

[0006] Obtaining an image data set collected by a UAV device on a target region, performing target tracking processing on the image data set based on a target tracker, and obtaining a plurality of track segments of the image data set;

[0007] Scene structure recognition is performed on the target region to identify multiple scene structure components. Based on the image dataset and multiple scene structure components, rhythm analysis is performed on the target region to construct rhythm profiles of multiple scene structure components.

[0008] Collect real-time monitoring data of the target area and extract real-time behavioral features of multiple scene structural components. Based on the rhythm profile, perform anomaly detection on the real-time behavioral features to generate structural behavior anomaly detection data of the target area.

[0009] Based on multiple trajectory segments of the image dataset, an interaction graph and dynamic flow constraint model of the target region are constructed. Behavioral density data and behavioral blockage data of the target region are extracted and fused to identify multiple local abnormal flow regions in the target region.

[0010] Based on the dynamic flow constraint model, the diffusion impact of local abnormal flow regions is identified, multiple abnormal flow regions are determined, and the dynamic change monitoring results of the target region are generated by fusing structural behavior anomaly detection data.

[0011] Preferably, scene structure recognition is performed on the target region to determine multiple scene structure components, and rhythm analysis is performed on the target region based on the image dataset and the multiple scene structure components to construct a rhythm profile of the multiple scene structure components, including:

[0012] The target area is segmented by boundary to obtain multiple candidate passage segments. The passage segments are then checked based on the multiple trajectory segments to generate multiple passage structure components.

[0013] The target area is divided into grid regions to generate multiple grid regions. The dwell data of each grid region is determined based on multiple trajectory segments, and multiple dwell structure components are generated. Multiple trajectory segments of the target area are fused to determine multiple channel structure components.

[0014] The image dataset is partitioned by load, multiple load intervals are determined, and corresponding interval profile data are extracted. From the interval profile data, the traffic distribution data of multiple passage structure components is extracted, the dwell distribution data and dwell time data of multiple dwell structure components are extracted, and the flow direction data and flow speed data of each channel structure component are extracted to obtain the rhythm profile of multiple scene structure components.

[0015] Preferably, the behavioral density data and behavioral blocking data of the target region are extracted, including:

[0016] Based on multiple trajectory segments, trajectory interaction analysis is performed on the target area within different sliding windows. The window interaction weight of any two trajectory segments within each sliding window is calculated, and the interaction graph of each sliding window is constructed based on multiple trajectory segments and window interaction weights.

[0017] Flow conflict detection is performed on multiple grid regions based on the interaction graph. This includes determining multiple flow target combinations in each grid region of each sliding window based on multiple trajectory segments, counting flow events in each grid region of the sliding window based on the interaction weights of multiple windows in the interaction graph, determining the behavior density data of each grid region under each interaction graph, and calculating the behavior blocking intensity parameter of each grid region based on the interaction weights of multiple windows in the interaction graph, thereby generating behavior blocking data for each grid region.

[0018] Preferably, for the dynamic flow constraint model, it further includes:

[0019] The flow network structure of the target area is constructed based on multiple scene structural components. Multiple channel diffusion combinations are extracted from the flow network structure, including the main channel and multiple branch channels. The flow diffusion sample set of each channel diffusion combination is constructed based on the image dataset and multiple trajectory segments.

[0020] For each flow diffusion sample set, hysteresis detection of the main channel with respect to multiple branch channels is performed to determine the diffusion hysteresis parameters and diffusion intensity data of the main channel in each branch channel. The diffusion hysteresis parameters and diffusion intensity data of multiple channel diffusion combinations are fused to generate a dynamic flow constraint model for the target area.

[0021] Preferably, the diffusion impact of local abnormal flow regions is identified based on the dynamic flow constraint model, and multiple abnormal flow regions are determined, including:

[0022] The channel diffusion combination to which each local abnormal flow region belongs is determined. Real-time diffusion data corresponding to the channel diffusion combination to which the local abnormal flow region belongs is extracted from real-time monitoring data. Diffusion anomaly detection is performed on each set of real-time diffusion data to obtain multiple sets of diffusion anomaly detection results. The diffusion hysteresis parameter and diffusion intensity data of the channel diffusion combination are matched with the diffusion anomaly detection results to determine the flow anomaly region corresponding to the channel diffusion combination to which the local abnormal flow region belongs.

[0023] Preferably, for each flow diffusion sample set, hysteresis detection of the main channel with respect to multiple branch channels is performed to determine the diffusion hysteresis parameters and diffusion intensity data of the main channel in each branch channel, including:

[0024] The cross-correlation coefficients between the main channel and multiple branch channels are calculated at multiple lag periods. The lag period with the largest cross-correlation coefficient is selected as the diffusion lag parameter of the main channel in the branch channel. The cross-correlation coefficients corresponding to the diffusion lag parameters are used as the diffusion intensity parameters of the main channel in the branch channels, thus generating the diffusion lag parameters and diffusion intensity data of the main channel in each branch channel.

[0025] A second aspect of the present invention provides a system for monitoring dynamic changes of objects based on UAV images, used to implement the above-mentioned method for monitoring dynamic changes of objects based on UAV images, comprising:

[0026] The trajectory analysis module is used to acquire image datasets collected by the UAV equipment on the target area, and to perform target tracking processing on the image dataset based on the target tracker to obtain multiple trajectory segments of the image dataset;

[0027] The rhythm profile construction module is used to identify scene structure in the target area to determine multiple scene structure components, perform rhythm analysis on the target area based on the image dataset and multiple scene structure components, and construct rhythm profiles of multiple scene structure components.

[0028] The structural behavior detection module is used to collect real-time monitoring data of the target area and extract real-time behavior features of multiple scene structural components. Based on the rhythm profile, it performs anomaly detection on the real-time behavior features and generates structural behavior anomaly detection data of the target area.

[0029] The flow anomaly identification module is used to construct an interaction graph and dynamic flow constraint model of the target region based on multiple trajectory segments of the image dataset, extract and fuse the behavior density data and behavior blockage data of the target region, and identify multiple local abnormal flow regions in the target region.

[0030] The diffusion impact analysis module is used to identify the diffusion impact of local abnormal flow areas based on the dynamic flow constraint model, determine multiple abnormal flow areas, and generate dynamic change monitoring results of the target area by fusing structural behavior anomaly detection data.

[0031] The present invention has the following beneficial effects:

[0032] This invention analyzes trajectory segments in UAV images to identify scene structure components in the target area and constructs a rhythmic profile to characterize the dynamic behavior patterns of the region. It combines interactive graphs and dynamic flow constraint models to quantify local flow conflicts and diffusion effects, and integrates the results of structural behavior anomaly detection and flow anomaly identification. This avoids the lag in anomaly warnings caused by ignoring dynamic rhythms and spatial correlations. By collaboratively analyzing target dwell density, behavioral blockage intensity, and channel diffusion characteristics, it can keenly capture early signs of anomalies before significant congestion occurs and accurately locate the source area of ​​flow anomalies. This significantly improves the real-time perception and warning accuracy of the gradual risk accumulation process and enhances the intelligent level of target dynamic change monitoring. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating a method for monitoring dynamic changes of objects based on UAV images, provided in an embodiment of the present invention.

[0034] Figure 2 This is a schematic diagram of the structure of an object dynamic change monitoring system based on UAV images, provided in an embodiment of the present invention. Detailed Implementation

[0035] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention.

[0036] This invention provides a method for monitoring the dynamic changes of objects based on UAV images. Please refer to [link to relevant documentation]. Figure 1 The method includes the following steps:

[0037] Step S1: Obtain the image dataset collected by the UAV equipment on the target area, and perform target tracking processing on the image dataset based on the target tracker to obtain multiple trajectory segments of the image dataset.

[0038] Specifically, image sequences or continuous video streams can be acquired using drones at a fixed frame rate to obtain image datasets of target areas, such as large integrated parks, hotel complexes, and other areas requiring dynamic monitoring. These areas involve complex dynamic behaviors involving multiple targets and types, such as personnel movement, vehicle traffic, and material transportation, necessitating comprehensive monitoring. The acquired image datasets are then processed for target detection and tracking. This invention employs existing mature multi-target tracking technologies to process continuous image frames, such as tracking algorithms based on detection results (e.g., DeepSORT, ByteTrack). This yields a set of trajectory segments containing the movement behaviors of numerous targets within the target area. Each trajectory segment includes time-series information, target center coordinates, direction of movement, velocity, acceleration, and other data, representing the movement path of one target, such as a vehicle or pedestrian. It should be noted that target detection and tracking algorithms are mature technologies well-known to those skilled in the art, and this invention does not specifically limit the implementation of this part.

[0039] Step S2: Perform scene structure recognition on the target area to determine multiple scene structure components. Based on the image dataset and multiple scene structure components, perform rhythm analysis on the target area to construct a rhythm profile of multiple scene structure components.

[0040] Specifically, the purpose of scene structure recognition is to identify multiple functional components within a region, such as entrance / exit areas, passageways, and dwell areas, thereby obtaining multiple scene structure components that represent key areas traversed by targets. Image datasets can be used to select representative data on the flow behavior of multiple targets within a target region under normal conditions. Combined with the image dataset, global rhythm analysis of the target region is performed to extract rhythmic data such as inflow / outflow volume, dwell time, and movement speed in different areas. Through statistical analysis, a rhythmic profile model of each structural component is constructed, reflecting the dynamic changes of the target region under normal conditions.

[0041] In this embodiment, the scene structure recognition process for multiple scene structure components may include:

[0042] The target area is segmented by boundary to obtain multiple candidate passage segments. Specifically, the boundary of the target area is determined and evenly divided into multiple candidate passage segments according to a pre-set length. These segments represent the location of the entrances and exits of the target area. Then, the multiple candidate passage segments are checked for passage based on multiple trajectory segments, and multiple passage structure components are generated. In this process, the candidate passage segments interspersed by each trajectory segment are determined, and the total number of trajectory segments interspersed by each candidate passage segment is counted. Several candidate passage segments that are frequently intersected by the trajectory segments are determined by the crossing count and marked as passage structure components, representing the entrance and exit structure of the target area.

[0043] The target area is divided into grid regions, generating multiple grid areas. Each grid area can be a fixed-size rectangular grid or a polygonal unit that adaptively adjusts according to terrain features. Then, based on multiple trajectory segments, the dwell time data for each grid area is determined, generating multiple dwell time structure components. For the dwell time data, the dwell time status of the target can be identified based on multiple trajectory segments. For example, if the dwell time in a grid area exceeds a pre-set dwell time threshold, it indicates that the target has stayed in that area once. By counting dwell times in different grid areas, multiple grid areas with high dwell frequency are selected and marked as dwell time structure components. This indicates that these areas exhibit short-term dwelling or clustering behavior under normal circumstances, such as waiting areas, waiting zones, or pickup points.

[0044] By fusing multiple trajectory segments from the target region to determine multiple channel structure components, the target region can be divided into multiple sub-grid regions with finer granularity. The number of times each sub-grid region is traversed by a trajectory segment is counted, and the frequently traversed sub-grid regions are identified and connected, thus obtaining multiple high-frequency passageways in the target region. These are marked as channel structure components, representing the channel skeleton of the target region. Multiple scene structure components in the target region are identified through the above method.

[0045] In this embodiment, rhythmic analysis of the target region based on the image dataset and multiple scene structure components is performed to construct a rhythmic profile of the multiple scene structure components, including:

[0046] The image dataset is partitioned by load, identifying multiple load intervals and extracting corresponding interval profile data for each. Load partitioning specifically refers to dividing historical image data based on the overall activity intensity of the target area. Specifically, it can be based on the total traffic data within the area, such as the maximum number of targets, to divide it into multiple load intervals, such as low-load, medium-load, and high-load intervals, or more granular division methods. Based on the total number of targets contained within different time windows in the image dataset, the dataset is segmented to obtain interval profile data for each load interval, representing the dynamic behavior changes of multiple moving targets in the target area under a specific traffic intensity.

[0047] For different scene structure components, statistical feature data of different dimensions are extracted, such as target entry and exit data, target movement data, target stay data and other multi-dimensional feature information.

[0048] Specifically, for access control components, traffic distribution data for multiple access control components is extracted from the interval profile data. For example, the average traffic volume of each access control component in the interval profile data represents the difference in traffic intensity at different entrances and exits under specific traffic flow conditions. By combining the average traffic volume data of multiple access control components, traffic distribution data for multiple access control components is obtained. This data is used to reflect the utilization rate changes and distribution characteristics of different entrance and exit channels under various load levels.

[0049] For dwell structure components, dwell distribution data and dwell time data for multiple dwell structure components are obtained from the interval profile data. Specifically, the average number of target dwells and the average dwell time in each dwell area are statistically analyzed and summarized into dwell time data. This data serves as the dwell behavior characteristics of each dwell structure component under a specific load traffic level for multiple targets. At the same time, the average number of target dwells of multiple dwell structure components are combined to obtain dwell distribution data for multiple dwell structure components, which is used to reflect the utilization rate changes of different dwell areas under various load levels.

[0050] For each channel structure component, the flow direction and velocity data are extracted. For example, the average flow direction and average velocity of different channel structure components within each load range are calculated. Alternatively, the velocity vector of a trajectory point in a trajectory segment can be projected onto the channel centerline of the corresponding local segment within the channel structure component to obtain the velocity distribution curve along the main direction. Furthermore, the velocity variation under different load levels is analyzed to characterize the flow rhythm and velocity fluctuation patterns of different mainstream moving channels, reflecting traffic efficiency and dynamic trends.

[0051] The statistical features extracted from various structural components are summarized to form rhythmic profile data corresponding to multiple load intervals, resulting in a rhythmic profile dataset for the target area. This dataset comprehensively describes the traffic patterns, dwell characteristics, and flow rhythm of the target area under different load conditions, reflecting the dynamic behavior patterns of the area under normal operating conditions. For example, when the load increases, the flow at some entrances and exits will increase preferentially, and the number of targets in some dwell areas will increase.

[0052] Step S3: Collect real-time monitoring data of the target area and extract real-time behavioral features of multiple scene structural components. Perform anomaly detection on the real-time behavioral features based on the rhythm profile to generate structural behavior anomaly detection data of the target area.

[0053] Specifically, for real-time monitoring data collected from the target area, the aforementioned target tracking processing method for the image dataset is used to obtain multiple real-time trajectories in the real-time monitoring data. For multiple scene structure components in the pre-constructed target area, real-time behavioral features of each scene structure component are extracted. Based on the target tracking processing results, the load interval matched to the real-time monitoring data is determined, and rhythmic profile data for anomaly detection of the real-time monitoring data is identified. The real-time behavioral features are compared with the rhythmic profiles of the corresponding structure components to analyze the differences between the current state and the rhythmic profile. Anomaly detection is performed using methods such as statistical deviation, dynamic threshold, or clustering offset. For example, different rhythmic features can be determined according to actual needs, such as traffic features, dwell features, and flow features, with corresponding deviation thresholds. The deviation data between the current state and the rhythmic profile regarding different rhythmic features is calculated. Finally, combined with the corresponding deviation thresholds, structural behavior anomaly detection data of the target area is identified from a global perspective. For example, some traffic structure components are abnormal, with traffic flow exceeding normal levels; some channel structure component areas have low average flow speeds and abnormal stagnation phenomena, thus identifying potential anomalies such as short-term aggregation, abnormal stagnation, or traffic fluctuations in the target area.

[0054] Step S4: Construct an interaction graph and dynamic flow constraint model for the target region based on multiple trajectory segments of the image dataset, extract and fuse the behavior density data and behavior blockage data of the target region, and identify multiple local abnormal flow regions in the target region.

[0055] Specifically, by analyzing the flow characteristic data between targets represented by various trajectory segments within the target area, the degree of mutual influence between different targets in space and time is quantified to construct an interaction diagram of the target area. Furthermore, behavioral density data and behavioral congestion data representing the overall dynamic state are extracted. These are used to comprehensively characterize the dynamic activity level or potential congestion level within the area, as well as the degree of mutual interference and movement restriction between targets.

[0056] In this embodiment, the process of constructing an interaction map of the target region based on multiple trajectory segments of the image dataset includes:

[0057] The target region is analyzed based on multiple trajectory segments within different sliding windows. The window interaction weight of any two trajectory segments within each sliding window is calculated. An interaction graph for each sliding window is constructed based on multiple trajectory segments and window interaction weights.

[0058] It is worth noting that the trajectory interaction analysis process specifically involves analyzing the flow influence state of targets represented by any two trajectory segments over a period of time. For the trajectory segments of two targets, sliding window processing is performed to extract the position and velocity information corresponding to each target in each sliding window, representing the flow state of each target, including the target's two-dimensional coordinates, velocity magnitude, and velocity direction. Based on the position and velocity information, it is determined whether the two targets collide in the sliding window according to their corresponding states, and the collision duration and initial position difference are determined. The ratio between the initial position difference and the collision duration is used as the window interaction weight of the two targets in a specific sliding window. The larger the initial position difference and the shorter the collision duration, the higher the interaction intensity of the two targets, and the larger the final window interaction weight value. In this way, the flow interaction influence characteristics between pairs of multiple targets in a specific sliding window can be constructed. Finally, using the targets represented by multiple trajectory segments as graph nodes, and combining multiple window interaction weights, an interaction graph of multiple trajectory segments in each sliding window is constructed.

[0059] Interaction graphs represent the changing interaction relationships between target individuals. Based on these graphs, the fluid interaction characteristics between targets are further mapped onto the real-world scene. Behavioral density data and behavioral congestion data for target regions are extracted from the interaction graphs. The specific implementation process includes:

[0060] Flow conflict detection is performed on multiple grid regions based on the interaction graph. This process first determines multiple combinations of flow targets within each grid region of each sliding window based on multiple trajectory segments. Specifically, it identifies multiple individual targets within each grid region of the sliding window and combines any two targets to obtain one flow target combination. Then, based on the interaction graph of the sliding window, the window interaction weight of each flow target combination is determined. Multiple flow events within these combinations are then identified based on a behavior impact threshold. If the window interaction weight of a flow target combination is greater than the behavior impact threshold, a flow event count is performed. This process of counting flow events in each grid region of the sliding window based on the interaction graph determines the total number of flow events in each grid region under each interaction graph, yielding behavior density data for different grid regions, representing the breadth of interaction occurrences in a specific region.

[0061] Furthermore, the behavior blocking intensity parameter of each grid region is calculated based on the interaction weights of multiple windows in the interaction graph. In this embodiment, the mean value corresponding to the window interaction weights of the combination of flowing targets is calculated for characterization, thereby generating behavior blocking data for each grid region, characterizing the degree of mutual interference caused by target interactions in a specific region.

[0062] During the real-time monitoring phase, real-time trajectories corresponding to multiple targets are extracted from the real-time monitoring image data. Real-time behavior density data and real-time behavior congestion data can be extracted from these trajectories using the methods described above. For the behavior density data and behavior congestion data extracted from the image dataset used as historical normal samples, these data are fused. In this embodiment, the product of the total number of flow events in each grid region and the behavior congestion intensity parameter is taken as the conflict intensity parameter for that grid region. For two regions with similar flow behavior density levels, a smaller behavior congestion intensity parameter indicates a smaller degree of behavioral interference caused by the actual target flow, and the overall conflict intensity may not have reached a congestion level.

[0063] By statistically analyzing the conflict intensity parameters of each grid region across multiple interaction maps—for example, calculating the mean and standard deviation—a reference conflict intensity range is determined based on the mean and standard deviation to assess the grid region's performance at a specific load level. For instance, the range corresponding to one standard deviation deviation centered on the mean is denoted as the reference conflict intensity range. A value above this range indicates potential congestion, while a value below it suggests possible anomalies in neighboring or upstream / downstream areas. By analyzing the reference conflict intensity ranges for different regions and the real-time conflict intensity parameters calculated from real-time monitoring data, the real-time anomaly status of different grid regions within the target area can be determined. Grid regions that do not conform to the reference conflict intensity range are marked as areas of localized abnormal flow.

[0064] Simultaneously, based on multiple scene structure components, a dynamic flow constraint model for the target region is further established to describe the flow patterns and temporal correlations of the target between different channels and regions. This model can be used for source tracing analysis of abnormal regions. In this embodiment, the construction process of the dynamic flow constraint model includes:

[0065] A flow network structure for the target region is constructed based on multiple scene structural components. Multiple channel diffusion combinations are extracted from the flow network structure. A flow diffusion sample set for each channel diffusion combination is constructed based on the image dataset and multiple trajectory segments.

[0066] Specifically, the flow network structure consists of several entrances / exits, channels, and rest areas corresponding to the aforementioned multiple scene structure components. It is used to depict the dynamic connection relationships between channels, rest areas, and entrances / exits within the target area. Based on the flow network structure, the intersections and rest areas between multiple channels are identified as channel nodes, and multiple entrances / exits are also recorded as channel nodes. The multiple channel structure components are then segmented according to the multiple channel nodes, that is, the local structure contained in two adjacent channel nodes in a channel structure component is segmented into one channel segment, thereby generating multiple channel segments.

[0067] Stability testing is performed on each channel segment. This process involves extracting traffic sequence data for different channel segments from interval profile data across different load ranges. Based on the traffic sequence data, the traffic intensity and traffic fluctuation parameters for each channel segment are calculated. Traffic intensity is characterized by the mean of the traffic sequence data, and the traffic fluctuation parameter is characterized by the standard deviation of the traffic sequence data. The traffic intensity is then corrected using the traffic fluctuation parameter to obtain the corrected traffic intensity for each channel segment. Multiple channel segments are then binary classified, and classification thresholds are set appropriately according to the actual scenario requirements to obtain multiple main channels and multiple branch channels.

[0068] The main channel reflects the primary flow direction in the target area, while branch channels typically correspond to secondary outflows or local convergence directions. From this, several channel diffusion combinations can be extracted, which are flow unit combinations centered on the main channel and branched by several branch channels connected to it. For each channel diffusion combination, a flow diffusion sample set is constructed based on the flow sequence data of each channel segment.

[0069] For each flow diffusion sample set, hysteresis detection of the main channel with respect to multiple branch channels is performed. The cross-correlation coefficient between the flow sequences of the main channel and each branch channel can be calculated. Specifically, it is calculated separately at different hysteresis periods. The hysteresis period with the largest cross-correlation coefficient is selected as the diffusion hysteresis parameter of the main channel in the branch channel. The cross-correlation coefficient corresponding to this hysteresis period is used as the diffusion intensity parameter of the main channel in the branch channel. Thus, the diffusion hysteresis parameter and diffusion intensity data of the main channel in each branch channel are obtained.

[0070] After obtaining the diffusion hysteresis parameters and diffusion intensity data for each diffusion combination, a local constraint model for each diffusion combination is established. These features are then fused to form a dynamic flow constraint model for the target area. This model describes the flow transmission path and response characteristics within the target area and can be used to determine flow anomalies in real-time data. For example, if abnormal congestion occurs in a branch channel in real-time monitoring data, while the flow in the corresponding main channel does not increase significantly, or the hysteresis response time far exceeds the normal range, it can be determined that there is a local flow anomaly in that branch channel.

[0071] Step S5: Identify the diffusion impact of local abnormal flow regions based on the dynamic flow constraint model, determine multiple abnormal flow regions, and fuse structural behavior anomaly detection data to generate dynamic change monitoring results for the target region.

[0072] Specifically, by analyzing the channel diffusion combinations associated with local abnormal flow regions, we can determine whether the dynamic flow change characteristics deviate from the dynamic flow constraint model in real-time scenarios, thereby identifying multiple abnormal flow regions. In this process, we determine the channel diffusion combination to which each local abnormal flow region belongs, that is, the channel diffusion combination of the main or branch channels covered by the local abnormal flow region.

[0073] For the identified channel diffusion combinations, corresponding real-time diffusion data is extracted from real-time monitoring data. This data, identical to the aforementioned flow diffusion sample set, describes the sequence of flow index changes in the main channel and each branch channel within a continuous time window in a real-time scenario. Using the aforementioned hysteresis detection method, diffusion anomaly detection is performed on each set of real-time diffusion data to obtain multiple sets of diffusion anomaly detection results. These results include the hysteresis characteristics and cross-correlation between the main channel and multiple branch channels within the channel diffusion combination under real-time data. Then, the diffusion hysteresis parameters and diffusion intensity data of the channel diffusion combination are matched with the diffusion anomaly detection results to determine the flow anomaly area corresponding to the channel diffusion combination to which the local abnormal flow area belongs.

[0074] The difference between real-time diffusion data and model data can be calculated based on the historical diffusion relationships of each channel combination in the dynamic flow constraint model. For example, if a local abnormal flow area covers one of the branch channels of the channel diffusion combination, and the change in the main channel flow should cause fluctuations in the flow of other normal branch channels within a certain lag time, and this response level is also observed in the real-time data, then it indicates that the local abnormal flow area is the source of the abnormality. Otherwise, it means that the main channel in the channel diffusion combination is the source of the diffusion abnormality. It is possible that there is an anomaly near the connection between the branch channel covered by the local abnormal flow area and the main channel, which has not yet affected other branch channels but needs to be addressed.

[0075] By performing the aforementioned hysteresis detection on real-time diffusion data for each channel diffusion combination, and combining it with a dynamic flow constraint model, multiple flow anomaly regions can be obtained. These include the original local anomaly flow regions and the source influence regions corresponding to some local anomaly flow regions. Each set of results records the anomaly regions and influence relationship chains for the corresponding combination. Finally, the structural behavior anomaly detection data and multiple flow anomaly regions are fused to generate dynamic change monitoring results for the target area. The flow anomaly regions mainly reflect local-level flow anomalies, including grid-level congestion, conflicts, or abnormal diffusion phenomena; while the structural behavior anomaly detection data reflects global-level structural changes, such as abnormal overall channel flow, excessive load in the dwell area, or imbalance in the flow direction of the gateway area. The results of the structural behavior anomaly detection data and the flow anomaly regions complement each other, jointly constituting comprehensive dynamic change monitoring of the target area. By fusing the two types of results, not only can overall structural operational anomalies be identified, but the spatial distribution and formation mechanism of anomalies can also be revealed at the local level, realizing dynamic monitoring from macro trends to micro behaviors. This can provide accurate dynamic references and early warning basis for traffic flow organization, crowd safety management, or on-site operation scheduling, significantly improving intelligent monitoring and management capabilities in complex scenarios.

[0076] Based on the same inventive concept, embodiments of the present invention also provide a system for monitoring dynamic changes of objects based on UAV images. Please refer to [link to relevant documentation]. Figure 2 The system includes:

[0077] The trajectory analysis module is used to acquire image datasets collected by the UAV equipment on the target area, and to perform target tracking processing on the image dataset based on the target tracker to obtain multiple trajectory segments of the image dataset;

[0078] The rhythm profile construction module is used to identify scene structure in the target area to determine multiple scene structure components, perform rhythm analysis on the target area based on the image dataset and multiple scene structure components, and construct rhythm profiles of multiple scene structure components.

[0079] The structural behavior detection module is used to collect real-time monitoring data of the target area and extract real-time behavior features of multiple scene structural components. Based on the rhythm profile, it performs anomaly detection on the real-time behavior features and generates structural behavior anomaly detection data of the target area.

[0080] The flow anomaly identification module is used to construct an interaction graph and dynamic flow constraint model of the target region based on multiple trajectory segments of the image dataset, extract and fuse the behavior density data and behavior blockage data of the target region, and identify multiple local abnormal flow regions in the target region.

[0081] The diffusion impact analysis module is used to identify the diffusion impact of local abnormal flow areas based on the dynamic flow constraint model, determine multiple abnormal flow areas, and generate dynamic change monitoring results of the target area by fusing structural behavior anomaly detection data.

[0082] The above are merely specific embodiments of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art. Parts not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A method for monitoring dynamic changes of objects based on images of unmanned aerial vehicles, characterized in that, The method comprises the following steps: acquiring image data sets collected by a UAV device on a target area, performing target tracking on the image data sets based on a target tracker to obtain multiple track segments of the image data sets; performing boundary segmentation on the target area to obtain multiple candidate passage segments, performing passage inspection on the multiple candidate passage segments based on the multiple track segments, and generating multiple passage structure components; performing grid partitioning on the target area to generate multiple grid areas, determining stay data of each grid area based on the multiple track segments, and generating multiple stay structure components, and fusing the multiple track segments of the target area to determine multiple channel structure components; performing load partitioning on the image data sets, determining multiple load intervals and extracting corresponding interval image data, extracting passage distribution data about the multiple passage structure components from the interval image data, extracting stay distribution data and stay time data about the multiple stay structure components, and extracting flow direction data and flow speed data of each channel structure component to obtain rhythm images of multiple scene structure components; acquiring real-time monitoring data of the target area and extracting real-time behavior features of the multiple scene structure components, performing anomaly detection on the real-time behavior features based on the rhythm images, and generating structure behavior anomaly detection data of the target area; performing trajectory interaction analysis on the target area based on the multiple track segments in different sliding windows, calculating window interaction weights of any two track segments in each sliding window, and constructing an interaction graph of each sliding window based on the multiple track segments and the window interaction weights; determining multiple flow target combinations of each grid area in each sliding window based on the multiple track segments, performing flow event counting on each grid area in the sliding window based on the interaction graph and the multiple window interaction weights, determining behavior density data of each grid area under each interaction graph, and calculating behavior congestion intensity parameters of each grid area based on the multiple window interaction weights in the interaction graph to generate behavior congestion data of each grid area, and fusing the behavior density data and the behavior congestion intensity parameters to identify multiple local abnormal flow areas of the target area; performing diffusion influence identification on the local abnormal flow areas based on a dynamic flow constraint model, determining multiple flow abnormal areas, and generating dynamic change monitoring results of the target area by fusing the structure behavior anomaly detection data. 2.The object dynamic change monitoring method based on UAV images of claim 1, wherein, For the dynamic flow constraint model, the method further comprises the following steps: constructing a flow network structure of the target area based on the multiple scene structure components, extracting multiple channel diffusion combinations from the flow network structure, including a main channel and multiple branch channels, and constructing a flow diffusion sample set of each channel diffusion combination based on the image data sets and the multiple track segments; performing lag detection of the main channel on the multiple branch channels on each flow diffusion sample set to determine diffusion lag parameters and diffusion intensity data of the main channel in each branch channel, fuse the diffusion lag parameters and diffusion intensity data of the multiple channel diffusion combinations, and generate a dynamic flow constraint model of the target area. 3.The object dynamic change monitoring method based on UAV images of claim 2, wherein, performing diffusion influence identification on the local abnormal flow areas based on the dynamic flow constraint model to determine multiple flow abnormal areas, comprising: The channel diffusion combination to which each local abnormal flow region belongs is determined, real-time diffusion data corresponding to the channel diffusion combination to which the local abnormal flow region belongs is extracted from the real-time monitoring data, a plurality of diffusion anomaly detection results are obtained by performing diffusion anomaly detection on each set of real-time diffusion data, and the diffusion lag parameter and diffusion intensity data of the channel diffusion combination are matched with the diffusion anomaly detection results to determine a flow abnormal region corresponding to the channel diffusion combination to which the local abnormal flow region belongs. 4.The object dynamic change monitoring method based on UAV images of claim 3, wherein, The lag detection of the main channel with respect to the plurality of branch channels is performed on each flow diffusion sample set, and the diffusion lag parameter and diffusion intensity data of the main channel in each branch channel are determined, including: The cross-correlation coefficients of the main channel and the plurality of branch channels are calculated at a plurality of lag periods, the lag period with the largest cross-correlation coefficient is selected as the diffusion lag parameter of the main channel in the branch channel, the cross-correlation coefficient corresponding to the diffusion lag parameter is taken as the diffusion intensity parameter of the main channel in the branch channel, and the diffusion lag parameter and diffusion intensity data of the main channel in each branch channel are generated.

5. An object dynamic change monitoring system based on images of unmanned aerial vehicles, characterized by, The system is used to implement the object dynamic change monitoring method based on the UAV image according to any one of claims 1-4, and includes: A trajectory analysis module is configured to acquire image data sets collected by a UAV device in a target region, perform target tracking on the image data sets based on a target tracker, and obtain a plurality of trajectory segments of the image data sets. A rhythm image construction module is configured to identify scene structure components in the target region, perform rhythm analysis on the target region based on the image data sets and the scene structure components, and construct rhythm images of the scene structure components. A structure behavior detection module is configured to acquire real-time monitoring data of the target region, extract real-time behavior features of the scene structure components, perform abnormal detection on the real-time behavior features based on the rhythm images, and generate structure behavior abnormal detection data of the target region. A flow anomaly recognition module is configured to construct an interaction graph and a dynamic flow constraint model of the target region based on the plurality of trajectory segments of the image data sets, extract and fuse behavior density data and behavior congestion data of the target region, and recognize a plurality of local abnormal flow regions of the target region. A diffusion influence analysis module is configured to perform diffusion influence recognition on the local abnormal flow regions based on the dynamic flow constraint model, determine a plurality of flow abnormal regions, and fuse the structure behavior abnormal detection data to generate a dynamic change monitoring result of the target region.

Images