River monitoring video-oriented floating object dynamic detection method and system

Abstract

The application provides a floating object dynamic detection method and system for river monitoring video. The method comprises: acquiring a river monitoring video, stabilizing consecutive frames, and determining a dynamic water area according to water surface boundary changes; extracting water surface motion information, correlating according to river position and time sequence, and generating a water surface basic motion field; predicting the evolution state of the water surface without floating objects based on the water surface basic motion field, analyzing the difference with the actual state, and generating a candidate trajectory by correlating the water surface abnormal area across time; establishing a water surface disturbance state hypothesis and a floating object transport state hypothesis for the candidate trajectory, predicting subsequent state changes, and obtaining counterfactual prediction results; matching the counterfactual prediction results with subsequent observation states, determining target attributes according to the matching difference; when the target attribute is a floating object, generating a floating object dynamic detection result based on the candidate trajectory and the water surface transport state. The application can reduce the false detection rate of complex water surface interference, improve the detection rate of small floating objects, and enhance the credibility of dynamic trajectory identification.

Description

Technical Field

[0001] This application relates to the fields of computer vision and intelligent water environment monitoring technology, and in particular to a method and system for dynamic detection of floating objects in river monitoring videos. Background Technology

[0002] With the development of river environment management and smart water management, the automatic detection of floating objects in rivers using monitoring equipment has become an important means of water pollution early warning, cleanup scheduling, and river supervision. Existing technologies typically employ background subtraction, inter-frame subtraction, optical flow analysis, or target detection models to extract motion regions from monitoring videos and identify floating objects and generate motion trajectories based on target appearance or the positional relationship of consecutive frames.

[0003] However, the river surface is constantly changing due to the influence of water flow, ripples, reflections, shadows, rainfall, and local eddies. Traditional background modeling and single-frame detection easily misjudge surface disturbances as floating objects. Floating objects are also characterized by their small scale, irregular shape, partial submersion, and tendency to flip with the water, making it difficult to maintain their appearance consistently. Existing optical flow or trajectory association methods typically do not establish a basic transport state of the water surface corresponding to the spatial location of the river, making it difficult to distinguish between water surface texture movement with the flow and the transport movement of independent objects. Furthermore, trajectory interruptions and duplicate counting are prone to occur when targets are occluded or their shapes change. In addition, existing methods often output results directly based on the confidence level of a single detection, lacking a mechanism to verify whether candidate regions maintain object states and conform to water flow transport patterns using subsequent video. This results in a high false detection rate in complex water surface environments, severe missed detection of weak targets, and insufficient continuity and reliability of detection results. Summary of the Invention

[0004] In view of this, the present application provides a method and system for dynamic detection of floating objects in river monitoring videos, in order to solve the problems of false detection caused by complex water surface disturbances, missed detection of small floating objects, and insufficient reliability of dynamic trajectory recognition in the prior art.

[0005] A first aspect of this application provides a method for dynamic detection of floating objects in river monitoring videos, comprising: acquiring monitoring videos of a target river section; performing image stabilization correction on continuous video frames in the monitoring videos; and determining dynamic water areas based on temporal changes in water surface boundaries; extracting water surface motion information from the dynamic water areas; associating the information with the river's spatial location and video temporal sequence to generate a basic water surface motion field characterizing the water surface transport state and range of change; predicting the water surface evolution state under conditions without floating objects based on the basic water surface motion field; and performing a difference analysis between the water surface evolution state and the actual water surface state of the corresponding video frames to generate... Anomaly regions on the water surface are identified, and these anomaly regions at different times are associated with the basic motion field of the water surface to form candidate trajectories. For each candidate trajectory, assumptions about the water surface disturbance state and the floating object transport state are established. Based on each state assumption, the state changes of the candidate trajectory in subsequent video frames are predicted to obtain the corresponding counterfactual prediction results. Each counterfactual prediction result is matched with the actual observed state in subsequent video frames, and the target attribute of the candidate trajectory is determined according to the matching difference corresponding to each state assumption. When the target attribute represents floating objects, dynamic detection results of floating objects are generated based on the candidate trajectory and the corresponding water surface transport state.

[0006] A second aspect of this application provides a floating object dynamic detection system for river monitoring videos, comprising: an acquisition module for acquiring monitoring videos of a target river section, performing image stabilization correction on continuous video frames in the monitoring video, and determining dynamic water areas based on the temporal changes of the water surface boundaries; an extraction module for extracting water surface motion information from the dynamic water areas, associating it according to the river's spatial location and video temporal sequence, and generating a basic water surface motion field characterizing the water surface transport state and range of change; and an analysis module for predicting the water surface evolution state under conditions without floating objects based on the basic water surface motion field, performing difference analysis between the water surface evolution state and the actual water surface state of the corresponding video frames, and generating... The system identifies anomalous water surface regions and associates these regions with the basic motion field of the water surface at different times to form candidate trajectories. A prediction module is used to establish water surface disturbance state assumptions and floating object transport state assumptions for each candidate trajectory, predicting the state changes of the candidate trajectory in subsequent video frames based on these assumptions to obtain corresponding counterfactual prediction results. A matching module is used to match each counterfactual prediction result with the actual observed state in subsequent video frames, determining the target attribute of the candidate trajectory based on the matching differences corresponding to each state assumption. A generation module is used to generate dynamic detection results of floating objects based on the candidate trajectories and the corresponding water surface transport state when the target attribute represents floating objects.

[0007] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By acquiring monitoring videos of the target river section, image stabilization correction is performed on continuous video frames in the monitoring videos, and dynamic water areas are determined based on the temporal changes of the water surface boundary. Water surface motion information is extracted from the dynamic water areas and correlated according to the spatial location of the river channel and the video temporal sequence to generate a basic motion field representing the water surface transport state and range of change. Based on the basic motion field, the water surface evolution state under the condition of no floating objects is predicted. The difference between the water surface evolution state and the actual water surface state of the corresponding video frame is analyzed to generate water surface anomaly areas. The water surface anomaly areas at different times are correlated according to the basic motion field to form candidate trajectories. For the candidate trajectories, water surface disturbance state assumptions and floating object transport state assumptions are established respectively. Based on each state assumption, the state change of the candidate trajectory in subsequent video frames is predicted to obtain the corresponding counterfactual prediction results. Each counterfactual prediction result is matched with the actual observation state in subsequent video frames, and the target attribute of the candidate trajectory is determined according to the matching difference corresponding to each state assumption. When the target attribute represents floating objects, the floating object dynamic detection result is generated based on the candidate trajectory and the corresponding water surface transport state. This application can reduce the false detection rate of interference from complex water surfaces, improve the detection rate of small floating objects, and enhance the reliability of dynamic trajectory recognition. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating the floating object dynamic detection method for river monitoring videos provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the floating object dynamic detection system for river monitoring video provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] Existing river floating object detection technologies typically utilize background subtraction, inter-frame subtraction, optical flow analysis, or target detection models to extract suspected moving areas from river monitoring videos and identify floating objects based on their appearance features and positional changes across consecutive video frames. Some technologies further employ target tracking algorithms to correlate detection results from different video frames to generate the floating object's trajectory.

[0012] However, the river surface is continuously affected by water flow fluctuations, light reflection, shadow changes, and local disturbances. The water surface background has significant non-rigid dynamic characteristics, making it difficult for existing methods to accurately distinguish between changes in the water surface itself and the movement of floating objects, which easily leads to false detections. At the same time, small floating objects may have low contrast, partial submersion, and continuous shape changes, making it easy to miss them in single-frame detection. Existing trajectory association mainly relies on position or appearance similarity, lacking constraints from water surface transport patterns and subsequent video verification, resulting in insufficient continuity and reliability of trajectory recognition when targets are occluded, adhered, or have changed shapes.

[0013] To address the aforementioned technical issues, this application first performs image stabilization correction on the monitoring video of the target river section and determines the dynamic water area based on the temporal changes of the water surface boundary. Water surface motion information is extracted from the dynamic water area and correlated according to the river channel spatial location and video temporal sequence to construct a basic water surface motion field characterizing the water surface transport state and its range of change. Subsequently, the water surface motion field is used to predict the water surface evolution state under conditions without floating objects. Anomaly regions on the water surface are generated by the difference between the predicted state and the actual water surface state, and these anomaly regions at different times are correlated according to the water surface transport relationship to form candidate trajectories.

[0014] For each candidate trajectory, this application establishes assumptions about the water surface disturbance state and the floating object transport state, respectively, and extrapolates the changes of the candidate trajectory in subsequent video frames under these two assumptions to obtain corresponding counterfactual prediction results. The counterfactual prediction results are then matched with the actual observed states in subsequent video frames, and the target attribute of the candidate trajectory is determined based on the matching difference between the two state assumptions. When the target attribute represents a floating object, the dynamic detection result of the floating object is generated by combining the candidate trajectory and the corresponding water surface transport state.

[0015] Through the above technical solution, this application can eliminate the interference of normal water surface transport changes on target detection by utilizing the basic motion field of the water surface, and verify the authenticity of candidate targets through counterfactual prediction of the dual-state hypothesis and subsequent observation matching, thereby reducing the false detection rate caused by complex water surface interference, improving the detection rate of weak floating objects, and enhancing the continuity and credibility of floating object dynamic trajectory recognition.

[0016] The technical solution of this application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0017] Figure 1This is a flowchart illustrating the floating object dynamic detection method for river monitoring videos provided in this application embodiment. Figure 1 As shown, the method may specifically include: S101, acquire the monitoring video of the target river section, perform image stabilization correction on the continuous video frames in the monitoring video, and determine the dynamic water area based on the temporal changes of the water surface boundary. S102 Extract water surface motion information from dynamic water areas, associate it with the spatial location of the river channel and the video time sequence, and generate a basic water surface motion field that represents the water surface transport status and the range of change. S103, based on the water surface basic motion field, predict the water surface evolution state under the condition of no floating objects, analyze the difference between the water surface evolution state and the actual water surface state of the corresponding video frame, generate water surface abnormal areas, and associate the water surface abnormal areas at different times according to the water surface basic motion field to form candidate trajectories. S104. Establish water surface disturbance state assumptions and floating object transport state assumptions for candidate trajectories respectively. Based on each state assumption, predict the state changes of candidate trajectories in subsequent video frames and obtain the corresponding counterfactual prediction results. S105, match each counterfactual prediction result with the actual observed state in subsequent video frames, and determine the target attribute of the candidate trajectory based on the matching difference corresponding to each state assumption. S106, When the target attribute represents a floating object, generate dynamic detection results of the floating object based on the candidate trajectory and the corresponding water surface transport state.

[0018] In some embodiments, acquiring monitoring video of the target river section, performing image stabilization correction on continuous video frames in the monitoring video, and determining dynamic water areas based on temporal changes in the water surface boundary includes: Extract cross-frame stable static reference features from non-water areas in consecutive video frames, and determine the image transformation relationship based on the positional changes of the static reference features in adjacent video frames; Based on the image transformation relationship, each video frame is registered to a unified spatial reference to obtain a stable video frame; Extract the water surface boundary from the stabilized video frame, and distinguish between the effective boundary caused by water level changes and the instantaneous interference boundary based on the consistency of the water surface boundary changes within a preset time window; The water area corresponding to the effective boundary is merged, and the continuously occluded areas are removed to generate a dynamic water area that updates with the video time sequence.

[0019] Specifically, a fixed camera installed on the bank of the target river section can be used as the video acquisition end, continuously acquiring river monitoring video at a sampling frequency of 25 frames per second, and reading continuous video frames at a basic processing cycle of 1 second. To avoid positional drift of the water surface area caused by slight camera vibration, bracket swaying, and wind load, a non-water area is first divided according to a pre-defined approximate range of the riverbank, and stable static reference features across frames are extracted within the non-water area. Static reference features can be derived from image structures that maintain their position over a long period of time, such as guardrail posts, bridge pier edges, bank wall joints, and the outlines of fixed buildings.

[0020] During extraction, multi-scale feature maps can be built for each video frame. Local feature points are matched between adjacent video frames, and temporary features generated by leaf swaying, pedestrian passing, and vehicle movement are filtered out based on the recurrence frequency, matching distance, and motion consistency over multiple consecutive processing cycles. For example, feature points that achieve stable matching in at least 16 out of 20 consecutive frames and whose matching displacement dispersion is below a set threshold are identified as static reference features. For feature points that are too concentrated, representative features at different locations can be retained according to image partitions to ensure that the image transformation relationship can cover the horizontal and vertical range of the river image.

[0021] Based on the filtered static reference features, a random consistency check method can be used to estimate the image transformation relationship between adjacent video frames. Specifically, based on the coordinate correspondence of the static reference features in the current frame and the reference frame, transformation parameters including translation, rotation, and scale changes can be obtained, and abnormal matching points that do not conform to the main transformation trend can be eliminated. When the fixed camera only experiences slight shaking, affine transformation can be used for correction; when there is a perspective change in the shooting angle, projection transformation can be used to establish the image transformation matrix. Subsequently, the first stable video frame or the key frame that has passed quality screening is used as a unified spatial reference, and coordinate mapping and edge compensation are performed frame by frame on subsequent video frames to obtain the stabilized video frame.

[0022] In some examples, registration quality values ​​can be generated based on reprojection errors and the proportion of valid matching points. When the registration quality value is lower than a set lower limit, the corresponding video frame is marked as a low-confidence frame. If the number of available static reference features for a frame is insufficient, or the transform residual exceeds the limit, the previous valid transform parameters are used and the registration confidence of that frame is reduced to avoid abrupt changes in the water area caused by erroneous transforms.

[0023] After obtaining stable video frames, water surface boundaries can be extracted by combining water color distribution, texture continuity, reflection characteristics, and riverbank structure constraints. First, a baseline of the water body's appearance is generated using historical stable time periods, and candidate water body regions are determined in each stable video frame. Then, candidate water surface boundaries are extracted along the transition points between the candidate water body regions and the shoreline structure. For low-confidence frames, the contribution of the corresponding candidate boundaries to the temporal decision can be reduced to prevent registration errors from being identified as water level changes. Considering that waves, reflections, floating shadows, and short-term occlusions can cause boundary jitter, a 5-second temporal window can be set to analyze the direction, magnitude, and spatial continuity of boundary changes at each location within the window.

[0024] In some examples, when the boundaries of adjacent locations exhibit a consistent direction and gradually varying amplitude of overall movement across multiple consecutive moments, the change is determined to be a valid boundary change caused by water level fluctuations. When the boundary only appears briefly in a local area, the direction of change is discontinuous, or it recovers after a few video frames, the change is determined to be a transient interference boundary. For example, if the water level continues to rise after rainfall, causing the shoreline water surface boundary to shift horizontally towards the shore over 8 seconds, it can be retained as a valid boundary; waves generated by passing ships that only cause local boundary oscillations are not used to update the water area.

[0025] Finally, the water areas enclosed by each effective boundary are merged in chronological order. Newly added water areas that consistently appear across multiple time windows are gradually written into the dynamic water area, while areas that continuously fall outside the effective boundary are gradually removed to prevent boundary jumps caused by a single judgment. During merging, the frequency of water area occurrence and the most recent update time can be recorded for each pixel. When the frequency of water area occurrence reaches the update threshold, it is confirmed to be added; when the duration of continuous absence reaches the exit threshold, it is confirmed to be removed.

[0026] Simultaneously, it can statistically analyze the obstruction areas formed by the back of bridge piers, tree canopies on the shore, fixed pipelines, and long-term mooring facilities. Areas that consistently lack effective water surface observation for a preset duration are marked as persistently obstructed areas and removed from the merged water area. Dynamic water areas can be represented using a binary region mask of the same size as the stabilized video frame, and a corresponding version identifier is output with each processing cycle. The resulting dynamic water area is continuously updated according to the water level and shooting status, and defines the effective processing range for subsequent water surface movement information extraction, reducing interference from moving targets on the shore, fixed obstructions, and image shake on floating object detection.

[0027] In some embodiments, water surface motion information is extracted from dynamic water areas and correlated according to the spatial location of the river channel and the video temporal sequence to generate a basic water surface motion field characterizing the water surface transport state and range of change, including: Filter out persistently abnormal regions from dynamic water areas and extract the local motion vectors of the remaining water surface regions between consecutive video frames; The local motion vectors are mapped to the river coordinate space established based on the river extension direction, and aggregated according to spatial location and temporal window to obtain the basic transport parameters corresponding to each water surface location. Based on the continuity of motion between adjacent water surface locations and the temporal stability of the same water surface location, the basic transport parameters are constrained and corrected, and the corresponding range of motion variation is determined. Based on the corrected basic transport parameters and the range of motion changes, a basic motion field for the water surface is generated and updated with the video sequence.

[0028] Specifically, based on the aforementioned dynamic water area mask, a water surface motion analysis region can be established according to continuous video frames. First, the duration of abnormal responses for each pixel or image patch within the dynamic water area is statistically analyzed. Regions exhibiting persistent high differences, repeated anomalies at fixed occlusion edges, prolonged presence of floating structures, and continuous disturbance of shoreline reflections are identified as persistent abnormal regions. For example, using a 10-second statistical window, if an image patch maintains an abnormal response in more than 80% of video frames, and its center position displacement relative to the stable image reference is less than 2 pixels, this image patch can be marked as a persistent abnormal region and temporarily masked in subsequent water surface motion estimation. For areas covered by confirmed floating object trajectories, they can also be excluded from the basic motion estimation samples during the duration of the corresponding trajectory, preventing independent target motions from being written into the basic water surface transport state.

[0029] Furthermore, after filtering out persistently abnormal regions, the remaining water surface area can be divided into overlapping local image blocks of the same size, and local motion vectors can be extracted between adjacent video frames. Specifically, cross-frame matching can be performed using water surface texture, brightness gradient, and local reflection structure to obtain the displacement direction and amount corresponding to the center position of each image block, which is then converted into local motion velocity by combining the video frame time interval. To reduce erroneous matching caused by ripple flicker and lighting changes, consistency checks can be performed on the forward and reverse matching results, and motion vectors with round-trip position errors exceeding a preset threshold can be removed.

[0030] Furthermore, vector confidence can be generated based on the effective texture ratio within an image patch, the matching residual, and the consistency of neighboring vector directions. For example, if an image patch moves 1.2 pixels downstream in 0.04 seconds, which, combined with the calibration ratio, is converted to 0.18 meters per second, and remains consistent in both forward and backward verifications, then this velocity and its corresponding direction can be considered as an effective local motion vector.

[0031] To ensure that motion results from different shooting angles have a unified spatial meaning, a river coordinate space can be established based on the river centerline or the riverbank boundaries. The vertical coordinates increase along the main extension direction of the river, while the horizontal coordinates represent the position relative to the river centerline. Using camera calibration relationships or riverbank control points, local motion vectors in the image coordinates are converted into vertical and horizontal components in the river coordinate space. Subsequently, effective local motion vectors are aggregated according to the river's vertical position, horizontal partitioning, and time-series windows.

[0032] For example, a longitudinal unit can be set every 2 meters along the river channel, and the river width can be divided into left bank, middle, and right bank areas. Within a 3-second window, the motion vectors within the same spatial unit are weighted and statistically analyzed. The weights can be determined by vector confidence, spatial distance from the unit center, and temporal distance between the acquisition time and the current time. The main transport direction, average transport velocity, lateral offset, and effective sample ratio are calculated from the aggregated results to form the basic transport parameters corresponding to each water surface location.

[0033] Considering that local reflections, eddies, and sample sparsity may cause abrupt changes in parameters of adjacent units, the basic transport parameters can be constrained and corrected based on the continuity of the river channel. Specifically, the main transport direction and velocity difference between adjacent spatial units can be compared. When the difference exceeds the allowable range but the effective sample ratio of the anomalous unit is low, the weighted result of upstream and downstream neighboring units is used to correct the anomalous unit. When there is a long-term stable turning near the river channel or bridge piers, local changes consistent with the river channel structure are retained. At the same time, stability analysis is performed on the basic transport parameters of the same spatial unit within multiple consecutive time windows. If the current parameter undergoes a short-term abrupt change relative to the historical median state and does not continuously reach the set window number, a smooth update method is used to limit the magnitude of the change. If the change persists, the basic state is gradually adjusted so that the motion field can adapt to changes in flow velocity and water level.

[0034] After constraint correction is completed, the range of motion variation can be determined based on the dispersion of the effective motion vectors relative to the correction parameters within the same spatial cell. The range of motion variation can include the allowable velocity range, the directional deflection range, and the lateral drift range. For example, if the corrected velocity of a certain central water surface cell is 0.22 m / s, and most effective samples within a continuous window are between 0.17 m / s and 0.28 m / s, then this range is recorded as the normal velocity variation range; for backflow areas near the shore, a larger directional deflection range can be retained.

[0035] Finally, the calibration basis transport parameters and motion change range of each spatial unit are correlated according to the river channel coordinates and mapped back to the stabilized video frames to form a basic motion field of the water surface that updates sequentially with the video. Each update retains the timestamp, the number of valid samples, and the confidence level for subsequent prediction of water surface evolution without floating objects and anomaly confidence correction, thereby avoiding misidentification of normal water flow fluctuations as floating object motion.

[0036] In some embodiments, the water surface evolution state under no-floating-object conditions is predicted based on the basic motion field of the water surface. The difference between the water surface evolution state and the actual water surface state of the corresponding video frame is analyzed to generate abnormal water surface regions. These abnormal water surface regions are then associated with different time points according to the basic motion field of the water surface to form candidate trajectories, including: Based on the current video frame and the basic motion field of the water surface, the water surface texture is transported in a temporal sequence, and the transport results are compensated according to the temporal changes in the appearance of the water surface to generate the water surface evolution state. The water surface evolution state is compared with the actual water surface state of the corresponding video frame by multi-scale difference calculation, and the difference credibility is corrected according to the motion change range corresponding to the basic motion field of the water surface to generate anomaly response of the water surface. Spatiotemporal aggregation of abnormal water surface responses in consecutive video frames is performed to generate abnormal water surface regions. Based on the basic motion field of the water surface, the transport reach range of each abnormal area of ​​the water surface in subsequent video frames is determined, and a temporal association is established based on the state matching relationship between the abnormal areas of the water surface within the transport reach range to form candidate trajectories.

[0037] Specifically, within the current stabilized video frame, the dynamic water area is used as the processing range. The water surface is divided into regular grids or adaptive texture blocks, and the brightness distribution, texture direction, local reflection intensity, and edge structure are extracted from each grid to form the current water surface state. Then, the main transport direction, transport velocity, and range of motion change corresponding to each grid in the basic motion field of the water surface are read. Based on the video frame time interval, the predicted displacement of each texture block in the next moment is calculated, and the current water surface texture is migrated to the predicted position through interpolation. For texture blocks that cross grid boundaries, segmented transport can be performed according to the transport parameters of adjacent grids; for areas lacking effective motion parameters, the transport states of upstream and downstream neighboring locations can be used to complete the transport. For example, if a texture block in the current frame is located in the middle of the river channel, with a basic transport velocity of 0.24 meters per second and a prediction interval of 0.2 seconds, then the texture block is migrated downstream by 0.048 meters, corresponding to the image distance, to obtain the initial transport result.

[0038] Considering that natural water surfaces are still affected by changes in illumination, ripples, and reflection angles even without floating objects, a water surface appearance change model can be established using multiple stable video frames preceding the current frame. Specifically, the brightness change trend, texture contrast fluctuation, and reflection area migration patterns of the same river location within a continuous time window can be statistically analyzed, and the initial transport results can be compensated based on the overall exposure change at the current moment. When the brightness of the entire dynamic water area increases synchronously, the overall brightness of the transport results can be corrected; when a local reflection band moves slowly along a fixed direction, the appearance of the corresponding area can be adjusted according to the historical rate of change. After texture transport and appearance compensation, the water surface evolution state at the predicted time under conditions without floating objects is generated.

[0039] When performing difference analysis between the water surface evolution state and the actual water surface state in the corresponding video frame, differences can be extracted at the pixel scale, local block scale, and region scale. The pixel scale is used to identify subtle brightness or texture deviations, the local block scale is used to determine structural changes within continuous regions, and the region scale is used to determine whether anomalies form a stable overall target. The brightness residual, texture residual, and motion residual at each scale can be normalized and fused according to the degree of spatial overlap to generate an initial difference response. Subsequently, the normal velocity range and direction deflection range of the basic motion field of the water surface at the corresponding location are read to correct the confidence of the difference. For example, if the actual motion velocity at a certain location deviates from the predicted mean but is still within the normal range of 0.18 m / s to 0.30 m / s, the weight of the corresponding motion residual is reduced; if the motion direction of a local region continuously exceeds the normal deflection range while the texture structure remains intact, the confidence of the anomaly response is increased.

[0040] For anomalous responses in consecutive video frames, an aggregation window of no less than 5 frames can be set, and spatiotemporal aggregation can be performed according to spatial adjacency, response intensity, and duration. Responses that appear only in a single frame, have a small area, and lack subsequent continuation are suppressed; responses that maintain close positional relationships and regional morphology in multiple consecutive frames are merged, and water surface anomaly regions are generated through connected component analysis. Each water surface anomaly region can record its occurrence time, center location, coverage area, average anomaly intensity, and region confidence. For example, reflective flickering that lasts only 2 frames and has no regular positional changes can be filtered out; local anomalies formed by plastic bottles that appear continuously for 12 frames are retained as valid water surface anomaly regions.

[0041] For each abnormal water surface area, the potential spatial range can be calculated by extrapolating from the basic motion field of the water surface corresponding to the center location within a set prediction time. The transport reach range is determined by the predicted center location and the range of motion changes. The main transport velocity determines the center migration distance, while the velocity range, direction deflection range, and lateral drift range determine the range boundary.

[0042] Subsequently, anomaly regions on the water surface that fall within the transport arrival range are searched in the next video frame or multiple subsequent video frames, and the deviation of the region center, area change, morphological similarity, anomaly intensity change, and consistency of movement direction are compared. Temporal associations are established for anomaly regions that meet the state matching conditions. In cases where multiple matching objects exist, the primary associated object is selected according to the comprehensive matching degree, while secondary associations are retained to handle adhesion or separation.

[0043] Furthermore, after continuously completing multi-frame association, water surface anomaly regions with stable temporal relationships are concatenated into candidate trajectories, and the trajectory duration, transport path, state changes, and association reliability are recorded, providing input for subsequent counterfactual prediction under the dual-state assumption. If a candidate trajectory re-enters the corresponding transport arrival range after a short period of occlusion, the association can be restored based on the predicted position and state continuity before and after occlusion, avoiding trajectory interruptions caused by bridge pier shadows or water surface splashes.

[0044] In some embodiments, multi-scale difference calculations are performed between the water surface evolution state and the actual water surface state of the corresponding video frame, and the difference reliability is corrected according to the motion change range corresponding to the basic motion field of the water surface to generate anomaly responses of the water surface, including: According to the preset scale levels, the regional representations corresponding to the water surface evolution state and the actual water surface state are extracted respectively, and the scale difference response is generated based on the degree of state deviation under the same scale level. Cross-scale fusion is performed based on the spatial overlap and temporal continuity relationships between responses at different scales to generate initial anomalous responses; Read the range of motion change of the basic motion field on the water surface at the corresponding water surface position, and determine the response confidence weight based on the degree of deviation of the actual water surface motion relative to the range of motion change. The initial anomalous response is corrected by using the response confidence weight, and the local response lacking temporal continuity is suppressed to generate anomaly response on the water surface.

[0045] Specifically, the water surface evolution state can be aligned with the actual water surface state at the same moment using unified spatial coordinates, and a scale hierarchy consisting of fine, medium, and coarse scales can be established within the dynamic water area. The fine scale can use 8×8 pixel local units or units of corresponding physical size to extract minute brightness changes, local texture direction, and fine-grained motion deviations; the medium scale can use 32×32 pixel regions to characterize the structural continuity between adjacent local units, regional edges, and consistency of local motion; the coarse scale can use 96×96 pixel regions or adaptive regions set according to the river width ratio to determine whether anomalies form an overall morphology distinct from the surrounding water surface. For each scale level, regional brightness distribution, texture statistics, motion direction, and regional morphological representations are extracted from the water surface evolution state and the actual water surface state, and the differences in various states of corresponding regions at the same scale are normalized to a unified range.

[0046] When generating scale difference responses, basic weights can be set according to the importance of different state components to floating object identification, and dynamically adjusted in conjunction with the effective texture ratio of the current region. For example, when the texture information is rich in a fine-scale region, the contribution of texture difference and local motion difference is increased; when the water surface reflection is strong and the texture is unstable, the contribution of brightness difference is reduced, and the contribution of structural edges and motion continuity is increased. The deviations of each state at the same scale can be weighted and aggregated to obtain the scale difference value of the corresponding region, and the scale difference response is formed by mapping according to the region location. If the actual water surface and the predicted water surface only have synchronous changes in overall brightness, but the local texture and motion state are basically the same, the scale difference response remains low; if a stable outline appears in a small area and the motion direction deviates from the surrounding water surface, the fine-scale and meso-scale responses increase synchronously.

[0047] Subsequently, cross-scale fusion is performed on the differential responses at different scales. First, responses at each scale are mapped to a uniform resolution, and then the spatial overlap ratio between high-response regions is compared. When a fine-scale high-response region is covered by a mesoscale high-response region, and the coarse-scale response shows continuous enhancement at the corresponding location, it is considered that multiple scales consistently support the same anomaly, and the fusion result is improved. When only a fine-scale isolated high response appears, while there are no corresponding responses at the mesoscale or coarse scale, the local response is reduced. In addition to spatial overlap, response continuity can be analyzed within a temporal window of 6 to 10 consecutive frames. If a high-response region moves continuously along the water surface transport direction, and the overlap area or predicted arrival range matching degree of adjacent times reaches a set condition, the temporal fusion weight is enhanced; if the response location randomly jumps or disappears within 1 to 2 frames, it is determined to be transient water surface flicker. An initial anomalous response is generated through joint fusion of spatial consistency and temporal continuity.

[0048] For each water surface location in the initial abnormal response, the allowable velocity range, direction deflection range, and lateral drift range of the basic motion field of the water surface are read, and the deviation of the actual water surface motion relative to each range of motion variation is calculated. When the actual motion is within the normal range, the response confidence weight remains at a low or medium level; when the actual motion approaches the range boundary, the weight is gradually increased according to the deviation trend; when the actual motion continuously exceeds multiple ranges of variation, the response confidence weight is increased.

[0049] For example, if the normal range of the basic transport velocity at a certain location is 0.16 m / s to 0.29 m / s, and the directional deflection range is 12 degrees to both sides of the downstream main direction, and the actual regional velocity is 0.23 m / s with a directional deflection of 8 degrees, then the response will not be enhanced due to motion differences. However, if the actual velocity is 0.36 m / s with a 25-degree deflection for 5 consecutive frames, while the regional morphology remains intact, then a higher response confidence weight will be determined. For areas near bridge piers or in the return flow zone, the normal range can be appropriately widened based on the historical motion dispersion to avoid misjudging stable eddies as abnormal.

[0050] Finally, the reliability weights of the responses are combined with the initial anomalous responses on a position-by-position basis. Responses that exceed the normal transport range and have cross-scale support are enhanced, while responses within the normal motion range or generated by only a single scale are attenuated. A secondary screening is then performed using temporal continuity constraints, requiring local responses to appear continuously for at least 3 frames within a preset number of frames and to have spatial adjacency or transport reachability with responses at preceding and following times. Short-term isolated responses caused by raindrop impacts, wave crest reflections, and compressed noise are directly reduced to below the anomalous output threshold; responses that are briefly weakened but reappear at the predicted location are retained based on the continuity of the preceding and following states. After reliability correction and temporal suppression, the output includes the response intensity, spatial location, duration, and reliability level of the water surface anomaly response, providing a stable input for the subsequent generation of water surface anomaly regions.

[0051] In some embodiments, water surface disturbance state assumptions and floating object transport state assumptions are established for candidate trajectories, and the state changes of candidate trajectories in subsequent video frames are predicted based on each state assumption to obtain corresponding counterfactual prediction results, including: Extract the state features of the water surface anomaly area corresponding to the candidate trajectory and the state of the surrounding water surface to generate the initial state of the candidate trajectory; Under the assumption of water surface disturbance state, the surrounding water surface state is used to replace the abnormal water surface area, and the temporal change of the replaced water surface state is deduced according to the basic motion field of the water surface to generate the counterfactual prediction result of water surface disturbance. Under the assumption of floating object transport state, the object state continuity of the abnormal area of ​​the water surface is maintained. By combining the basic motion field of the water surface and the motion state of the candidate trajectory relative to the water surface, the position and shape changes of the abnormal area of ​​the water surface are deduced, and the counterfactual prediction results of floating object transport are generated. The counterfactual prediction results are correlated with the corresponding state assumptions and candidate trajectories.

[0052] Specifically, an independent counterfactual analysis record can be established for each candidate trajectory, and the most recent 8 to 15 consecutive frames of the trajectory are selected as the initial observation window. First, the abnormal water surface regions of the candidate trajectory in each video frame are read, and the center position, coverage area, contour shape, brightness texture, abnormal response intensity, and cross-frame displacement of the region are extracted. At the same time, the basic motion field of the water surface and the actual water surface state within a certain distance outside the abnormal region are read. The outer range can be set to 1.5 to 3 times the equivalent diameter of the abnormal region, avoiding the coverage area of ​​other candidate trajectories. The main transport direction, average velocity, texture change trend, and motion change range of the outer water surface are statistically analyzed. Then, combined with the velocity difference, direction difference, and shape preservation degree of the candidate trajectory relative to the outer water surface, the initial state of the candidate trajectory is generated.

[0053] For example, if a candidate region moves downstream for 10 consecutive frames, with an average area between 180 and 205 pixels and an additional velocity relative to the surrounding water surface of 0.05 meters per second, its position, shape, relative motion, and the state of the surrounding water surface are written into the initial state. If there are missing frames or partial occlusions within the initial observation window, interpolation can be performed based on the transport positions of the preceding and following valid frames to complete the observation. The reliability of the initial state is then reduced according to the completion ratio to avoid overly certain predictions caused by incomplete trajectories.

[0054] Under the assumption of water surface disturbance, candidate anomalies are considered to be caused by ripples, reflections, eddies, or local water surface deformation, and independent objects are not retained. Background samples most similar to the current water surface texture can be selected from upstream, downstream, and both sides of the anomaly region. These background samples are then transported to the anomaly region based on the basic motion field of the water surface to replace the brightness, texture, and motion state of the original anomaly region. During replacement, a gradual blending can be performed according to the boundary distance to avoid abrupt stitching.

[0055] Subsequently, starting with the replaced water surface state, the water surface state for the next 5 to 20 frames is extrapolated frame by frame based on the main transport direction, velocity, and range of motion changes corresponding to each spatial location; simultaneously, brightness and reflection intensity are compensated based on historical changes in water surface appearance. If the anomaly is indeed caused by instantaneous ripples, the replaced area should be transported along with the surrounding water surface and gradually dissipate. The water surface texture, motion state, and expected residual range are recorded at each prediction moment, forming a counterfactual prediction result for water surface disturbance. For stable eddies near the bridge piers, periodic rotation of the replacement area within a defined location is allowed, but without forming a continuous independent boundary, to improve adaptability to special flow structures.

[0056] Under the assumption of floating object transport state, candidate anomalies are treated as objects with continuous boundaries and independent motion states, without completely replacing their regions with the surrounding water surface. Continuous constraints on the object state can be established based on area changes, contour rotation, texture preservation, and occlusion within the initial observation window. For example, the contour of a plastic bottle may periodically shorten when it flips, but the center position and local texture still maintain cross-frame correlation, thus allowing the area to vary within a historical range while preserving the object's identity. Subsequently, the basic motion velocity of the water surface is superimposed with the motion components of the candidate trajectory relative to the water surface to obtain the object's predicted velocity, and its subsequent position is inferred accordingly. The relative motion components can be a weighted average of the most recent frames, gradually adjusted according to the range of motion changes in areas of water flow turning, backflow, or sudden velocity changes.

[0057] For morphological extrapolation, a morphological state sequence can be established based on the historical contours of candidate trajectories to predict the regional state after translation, rotation, scale changes, and local occlusion. If the predicted path passes through the shadow of bridge piers, fixed facilities, or other floating objects, a temporarily invisible, partially visible, or adhered state can be generated, while retaining the trajectory identity. At each prediction time, the predicted center position, possible coverage area, morphological state, anomalous response intensity, and allowable deviation are output, along with the corresponding transport parameters and morphological transfer basis, forming a counterfactual prediction result for floating object transport. For example, if a candidate trajectory is predicted to reach a position 2.4 meters downstream in 1 second, even if it is occluded by reflections in the middle two frames, the corresponding predicted area can be retained for subsequent actual observation matching.

[0058] Finally, a trajectory identifier was assigned to each candidate trajectory. The counterfactual prediction results of surface disturbance were associated with the surface disturbance state hypothesis, and the counterfactual prediction results of floating object transport were associated with the floating object transport state hypothesis. The prediction start time, prediction window, basic motion field version, and initial state version were recorded simultaneously. Both types of results were expressed using a consistent time index and channel coordinates, allowing the actual observed state in subsequent video frames to be compared with the two types of prediction results under the same spatiotemporal reference. This provides a verifiable dual-hypothesis basis for determining the target attributes of candidate trajectories.

[0059] In some embodiments, by combining the basic motion field of the water surface and the motion state of candidate trajectories relative to the water surface, the location and morphological changes of abnormal areas on the water surface are deduced, generating counterfactual prediction results for floating object transport, including: Based on the difference between the cross-frame displacement of the candidate trajectory and the water surface transport state at the corresponding location, the motion state of the water surface anomaly area relative to the water surface is determined. The motion state is superimposed onto the basic motion field on the water surface, and the transport location of the abnormal area on the water surface is deduced according to the predicted time sequence. The morphological evolution constraints are determined based on the state continuity relationship of the water surface anomaly region in the candidate trajectory, and the regional state corresponding to each transport location is updated based on the morphological evolution constraints. The predicted range of regional state is corrected by using the range of motion changes of the water surface basic motion field. The transport position and regional state corresponding to each prediction time are temporally correlated to generate the counterfactual prediction result of floating object transport.

[0060] Specifically, for candidate trajectories assuming a floating object transport state, the coordinates of the center of the abnormal region, boundary contours, coverage area, principal axis direction, and the basic motion field of the water surface at the corresponding location can be read from the most recent 10 to 20 consecutive frames. First, the displacement of the center point of the abnormal region in adjacent video frames is converted into the actual motion velocity in the river channel coordinate space, and the main transport velocity, direction, and lateral drift are read from the basic motion field of the water surface at the same location. The motion component of the abnormal region relative to the water surface is obtained by subtracting the water surface transport vector from the actual motion vector.

[0061] In some examples, to reduce single-frame matching errors, the relative motion components at multiple consecutive time points can be weighted and smoothed. The weights are determined by trajectory association confidence, region visibility ratio, and matching residuals. If the relative motion components maintain consistent direction and stable amplitude across multiple consecutive frames, they are retained as independent motion states; if they deviate only briefly and subsequently return to the water surface motion range, the corresponding components are gradually attenuated. For example, if the actual downstream velocity of a candidate region is 0.31 m / s and the current base transport velocity at the water surface is 0.24 m / s, then its downstream additional velocity relative to the water surface is 0.07 m / s. If the additional velocity remains above 0.05 m / s for six consecutive frames, the motion state can be considered persistent.

[0062] Furthermore, after obtaining the relative water surface motion state, the transport position can be deduced frame by frame according to the predicted time sequence. For each prediction time, the basic motion field of the water surface is first read from the current position, and the smoothed relative motion components are superimposed on the basic transport vector to obtain the predicted motion vector of the abnormal area; then, the center position increment is calculated in combination with the video frame interval, and the result is written into the next prediction position. If the predicted path crosses multiple water surface spatial units, the corresponding basic transport parameters are updated each time a new unit is entered, so that the trajectory can change with the river channel bends, local backflow, and velocity gradient. For areas near bridge piers or riverbank backflow areas, the position update amplitude can be limited according to the local direction change rate to avoid the predicted position crossing inaccessible water areas. With a prediction step size of 0.2 seconds, if the superimposed velocity is 0.30 meters per second, the next moment moves 0.06 meters along the synthesis direction, and the transport position sequence within the next 2 seconds is continuously deduced. If a low-confidence motion unit is entered midway, upstream and downstream unit interpolation is used to complete the prediction, and the position prediction range is expanded synchronously.

[0063] For each predicted location, morphological evolution constraints can be established based on the state continuity relationship between historical anomalous regions of the candidate trajectory. This state continuity relationship can be characterized by the area change rate, contour similarity, principal axis rotation angle, texture preservation degree, and boundary integrity. If the area of ​​the historical region changes by no more than 15% between adjacent frames, the predicted area is limited to the corresponding change range; if the principal axis direction shows a continuous rotation trend, the subsequent attitude is predicted according to the historical rotation angle rate; if the trajectory has experienced partial occlusion, the predicted region is allowed to shrink in a short period, but the center motion and identity continuity are maintained. For soft floating objects, the contour deformation range can be relaxed, but the region's centroid and overall coverage must be continuous; for rigid floating objects, the shape preservation constraint can be increased. Based on these constraints, predicted contours, area ranges, direction ranges, and visible states are generated at each transport location, thereby obtaining the corresponding region state.

[0064] Subsequently, the allowable velocity range, directional deflection range, and lateral drift range of the basic motion field on the water surface are read to correct the range of the predicted location and regional state. Specifically, the motion dispersion of the basic motion field can be propagated to future moments, forming a spatial reach range that gradually expands with the prediction time; at the same time, the uncertain range of the regional boundary is determined by combining morphological evolution constraints. The longer the prediction time, the more the location and morphological ranges can be appropriately expanded according to the cumulative motion fluctuations, but they must not exceed the constraints of the dynamic water area and channel structure. If the water surface velocity range at a certain location is 0.20 m / s to 0.28 m / s, and the directional deflection is allowed to be 10 degrees, then a fan-shaped reach region corresponding to this range is generated within a 1-second prediction time, and the predicted contour range of the abnormal region is superimposed on it.

[0065] Finally, the center position, spatial reach, predicted outline, area range, attitude, visibility status, and prediction confidence at each prediction time are correlated chronologically and recorded along with the candidate trajectory identifier, the basic motion field version, and the relative motion state version. If the predicted path passes through a persistently occluded area, the hidden state and the predicted position of reappearance are retained; if the predicted area may adhere to other candidate trajectories, the merging state and subsequent separation range are recorded. This forms a continuous counterfactual prediction result for floating object transport, providing a complete spatiotemporal prediction basis for subsequent time-by-time matching with the actual observation state.

[0066] In some embodiments, each counterfactual prediction result is matched with the actual observed state in subsequent video frames, and the target attribute of the candidate trajectory is determined based on the matching difference corresponding to each state assumption, including: Extract the actual observation state corresponding to the candidate trajectory from subsequent video frames, and match the state with the counterfactual prediction results of water surface disturbance and floating object transport according to the prediction time sequence. Based on the deviation between the predicted and actual positions at each time point and the degree of preservation between the predicted state and the actual observed state, the temporal matching degree corresponding to each state hypothesis is generated. Based on the range of motion changes of the basic motion field on the water surface and the effectiveness of the actual observation state, the credibility of each time series matching degree is corrected to obtain the hypothesis explanatory power. Compare the explanatory power of the assumptions corresponding to the water surface disturbance state assumption and the floating object transport state assumption, and determine the target attributes of the candidate trajectories based on the difference in explanatory power and time series stability.

[0067] Specifically, after the two types of counterfactual prediction results are generated, actual observation data can be read from subsequent consecutive video frames according to the prediction start time corresponding to the candidate trajectory. First, based on the transport position and spatial arrival range recorded at each prediction time, the corresponding search area is extracted in the dynamic water area, and the surface anomaly response, area center, coverage area, contour morphology, texture state, movement direction, and visibility ratio are obtained from the search area. If there are multiple anomaly areas in the search area, the main observation object can be determined according to the distance from the prediction center, the degree of overlap of the area boundary, and the correlation state of the previous time. If the prediction position falls into the shadow of the bridge pier, fixed obstruction, or strong reflective area, the time is marked as a low effective observation, and the matching entry point for subsequent recurrence is retained. This forms the actual observation state sequence corresponding to the candidate trajectory, and the surface disturbance counterfactual prediction results and floating object transport counterfactual prediction results are aligned according to the prediction time index.

[0068] For assumptions regarding water surface disturbance, the consistency between predicted water surface texture, local motion, and anomaly dissipation trends and actual observed conditions should be compared. If the actual anomaly area gradually weakens and merges into the surrounding water surface during the prediction period, and the actual motion remains within the normal range of variation of the basic motion field of the water surface, then the assumption achieves a high matching result; if the actual area still maintains an independent outline and moves continuously across frames, then the corresponding matching result decreases. For assumptions regarding floating object transport, the distance between the predicted center position and the actual center position, the coverage relationship between the predicted reach area and the actual area, the degree of conformity between the predicted area range and the actual area, and the preservation of shape, attitude, and visibility can be compared. For example, if the prediction result indicates that the target reaches a position 1.9 meters downstream after 0.8 seconds, the actual area center differs from the predicted center by 0.12 meters, and the area is within the predicted range, then both position and state matching values ​​can be high.

[0069] At each prediction time, positional deviations can be converted into positional matching values, and differences between regional states can be converted into state retention values. These are then combined with temporal continuity to generate a temporal matching degree. Positional matching values ​​decrease as the actual center deviates from the prediction center, decreasing significantly when the actual region is outside the predicted spatial reach range. State retention values ​​are determined by area change, contour similarity, texture continuity, motion direction, and anomaly intensity. For the water surface disturbance state hypothesis, state retention primarily reflects whether the actual water surface recovers to the surrounding background; for the floating object transport state hypothesis, state retention primarily reflects the object's identity and whether the regional structure is continuous. Temporal weighted aggregation of matching results from 10 consecutive frames can be performed, giving higher weight to frames closer to the current time and with valid observations, and smoothing abrupt changes in single frames. For different state components within the prediction interval, normalization can be performed first according to the corresponding allowable deviation to avoid direct mixing of positional and morphological dimensions, obtaining the temporal matching degree for each type of hypothesis.

[0070] Subsequently, the reliability of the temporal matching was corrected based on the allowable velocity range, directional deflection range, and lateral drift range recorded at each actual location of the basic motion field on the water surface. When the actual motion deviates from the predicted mean but remains within the normal range of variation, the deviation is not considered as evidence of floating objects; when the actual area continuously exceeds the normal motion range and maintains an independent morphology, the reliability contribution of the floating object transport state hypothesis is increased. Simultaneously, the weights of each time step are adjusted according to the effectiveness of the actual observation state. Observations that are clearly visible, have complete regions, and high registration quality are given higher weights, while observations with severe occlusion, reflective saturation, or missing textures are given lower weights. The corrected matching results at each time step are aggregated to obtain the hypothesis explanatory power of the water surface disturbance state hypothesis and the floating object transport state hypothesis, respectively.

[0071] Finally, the explanatory power of the two hypotheses and their stability of change within a continuous time window were compared. When the explanatory power of the floating object transport state hypothesis was consistently higher than that of the surface disturbance state hypothesis, and the difference in explanatory power continuously reached the judgment threshold, the target attribute of the candidate trajectory was determined to be floating object; when the surface disturbance state hypothesis was dominant and the actual anomaly gradually dissipated, the target attribute was determined to be surface disturbance. If the difference in explanatory power between the two types was small or the anomaly was frequently reversed due to occlusion, it was temporarily marked as a target to be confirmed, and the observation window was extended to continue matching. For example, if the floating object transport state hypothesis maintained a high explanatory power in 10 out of 12 consecutive frames, with an average explanatory power difference of 0.28, the target attribute could be confirmed as floating object; if the anomaly disappeared after only 3 frames, it was determined to be surface disturbance. Through dual hypothesis competition and subsequent observation verification, target determination can be avoided by directly making decisions based solely on the appearance of a single frame or local motion.

[0072] In some embodiments, when the target attribute characterizes a floating object, dynamic detection results of the floating object are generated based on candidate trajectories and corresponding surface transport states, including: Based on the temporal correlation in the candidate trajectories, the trajectory nodes of the floating objects in the states of obstruction, adhesion, or separation are continuously corrected to obtain the effective floating object trajectory. Based on the effective floating object trajectory and the corresponding water surface transport state, the motion characteristics of the floating object relative to the water surface are determined, and the position change and motion state of the floating object are generated. The floating object detection confidence level is generated by combining the target attribute determination results and trajectory continuity of the candidate trajectory. When the confidence level of the floating object detection meets the confirmation criteria, the effective floating object trajectory, motion state, and corresponding spatiotemporal information are correlated to generate a dynamic floating object detection result.

[0073] Specifically, after the target attribute of a candidate trajectory is determined to be a floating object, all trajectory nodes, node times, abnormal area states, and inter-node relationships contained in the candidate trajectory can be read, and a trajectory continuity correction window can be established in chronological order. For continuously visible trajectory nodes, the original relationships can be directly retained; for node loss caused by short-term occlusion, the possible locations during the occlusion period can be predicted based on the position, velocity, direction, and corresponding water surface transport status of the last valid node before occlusion, and the re-emerging abnormal areas can be searched within the transport arrival range after the occlusion ends.

[0074] In some examples, when the reappearance area matches the predicted location, shape range, and direction of movement, the trajectory nodes before and after the occlusion are reconnected, and the predicted node is inserted at the missing moment. For example, if a floating object is missing for four consecutive frames when passing through the shadow of a bridge pier, and the predicted trajectory indicates that the target should reappear within a range of 0.32 meters to 0.46 meters downstream, and the actual abnormal area falls within this range with an area change of no more than 20%, then the trajectory continuity is restored.

[0075] For multiple floating objects that have converged to form a single anomalous region and become stuck together, the predicted reach range, relative water surface motion components, and morphological state of each candidate trajectory before sticking together can be compared to establish a temporary association between multiple input trajectories and the same sticking node. During sticking together, the original trajectory is not immediately terminated; instead, the hidden state and predicted position of each floating object are maintained. When the sticking region separates into multiple anomalous regions in subsequent video frames, the separating nodes are reconnected to their corresponding trajectories based on the distance between each separated region and its predicted position, the continuity of its motion direction, and the similarity of its historical morphology. If a floating object deforms and is divided into multiple anomalous regions, merging is performed based on the total coverage area of ​​each region, its common motion direction, and the complementary relationship of its boundaries to avoid generating multiple trajectories repeatedly. After the above processing, isolated nodes with too short a duration, lack of effective observational support, or inability to be connected through transport laws can be deleted, resulting in valid floating object trajectories.

[0076] Based on the effective floating object trajectory, the center positions of each node can be transformed into the river channel coordinate space, and the actual displacement, actual velocity, and directional changes between adjacent nodes can be calculated. Then, the basic motion field of the water surface corresponding to each node position is read, and the motion characteristics of the floating object relative to the water surface are obtained by subtracting the water surface transport vector from the actual motion vector. Motion characteristics can include relative velocity, lateral drift, directional deviation, and motion stability. If the relative velocity is close to zero and the motion direction is consistent with the main water surface transport direction, it can be determined that the floating object is mainly drifting with the current; if there is a continuous lateral deviation or additional velocity, it can be determined that the floating object is affected by wind, backflow, or local obstruction. For example, if the actual downstream velocity of a floating object is 0.27 m / s, the basic water surface velocity is 0.22 m / s, and the lateral velocity is 0.03 m / s, then the additional downstream velocity of 0.05 m / s and the drift towards the right bank are recorded. Combining the position coordinates, velocity, direction, and visibility at each moment, a sequence of the floating object's position changes and motion states is generated.

[0077] The reliability of floating object detection can be jointly generated by the target attribute determination results and the trajectory continuity. The target attribute determination results can be characterized by the difference in interpretability between the floating object transport state hypothesis and the water surface disturbance state hypothesis; the trajectory continuity can be determined based on the proportion of effective observation nodes, the proportion of occlusion completion, the stability of node association, the degree of morphological preservation, and the consistency of motion. For targets with long continuous observation time, stable interpretability differences, complete trajectory nodes, and motion conforming to transport laws, the detection reliability is increased; for targets with a large number of nodes relying on prediction completion, unclear adhesion and separation relationships, or frequent fluctuations in interpretability, the detection reliability is decreased. Various indicators can be normalized and weighted, and confirmation thresholds and verification thresholds can be set. For example, a detection reliability of 0.80 or higher can be directly confirmed, while those between 0.60 and 0.80 can continue to be observed, and those below 0.60 will not output a formal result.

[0078] When the confidence level of floating object detection meets the confirmation criteria, a unique target identifier is assigned to the valid floating object trajectory. The trajectory's start and end times, river channel location sequence, movement speed, direction of movement, relative water surface movement characteristics, obstruction or adhesion status, and detection confidence level are correlated to generate dynamic floating object detection results. The detection results can simultaneously record the target's first appearance location, current river segment, expected downstream reach range, and most recent update time, and are continuously updated with new video frames. For the same floating object spanning multiple monitoring cycles, the results can be maintained continuously based on the target identifier; for targets confirmed to have left the dynamic water area or not reappearing for a long time, the corresponding trajectory can be terminated and the termination position recorded. The output detection results not only characterize the presence of floating objects but also continuously reflect the transport path and movement status of floating objects in the river channel, providing a basis for subsequent pollution early warning, cleanup scheduling, and target verification.

[0079] The following are system embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the system embodiments of this application, please refer to the method embodiments of this application.

[0080] Figure 2 This is a schematic diagram of the structure of a floating object dynamic detection system for river monitoring video provided in an embodiment of this application. Figure 2 As shown, the system includes: The acquisition module 201 is used to acquire the monitoring video of the target river section, perform image stabilization correction on the continuous video frames in the monitoring video, and determine the dynamic water area based on the temporal changes of the water surface boundary. Extraction module 202 is used to extract water surface motion information from dynamic water areas, associate it according to the spatial location of the river channel and the video time sequence, and generate a basic water surface motion field that represents the water surface transport status and range of change. Analysis module 203 is used to predict the water surface evolution state under the condition of no floating objects based on the basic motion field of the water surface. It performs difference analysis between the water surface evolution state and the actual water surface state of the corresponding video frame, generates water surface abnormal areas, and associates the water surface abnormal areas at different times according to the basic motion field of the water surface to form candidate trajectories. Prediction module 204 is used to establish water surface disturbance state assumptions and floating object transport state assumptions for candidate trajectories respectively, and predict the state changes of candidate trajectories in subsequent video frames based on each state assumption to obtain corresponding counterfactual prediction results. Matching module 205 is used to match each counterfactual prediction result with the actual observed state in subsequent video frames, and determine the target attribute of the candidate trajectory based on the matching difference corresponding to each state hypothesis. The generation module 206 is used to generate dynamic detection results of floating objects based on candidate trajectories and corresponding water surface transport states when the target attributes characterize the floating objects.

[0081] In some embodiments, Figure 2 The acquisition module 201 extracts cross-frame stable static reference features from non-water areas of continuous video frames and determines the image transformation relationship based on the positional changes of the static reference features in adjacent video frames; based on the image transformation relationship, it registers each video frame to a unified spatial reference to obtain a stable video frame; it extracts the water surface boundary from the stable video frame and distinguishes between the effective boundary caused by water level changes and the instantaneous interference boundary based on the consistency of the water surface boundary changes within a preset time window; it fuses the water area corresponding to the effective boundary and removes the continuously occluded area to generate a dynamic water area that updates with the video time sequence.

[0082] In some embodiments, Figure 2The extraction module 202 filters out persistently abnormal regions from the dynamic water area and extracts the local motion vectors of the remaining water surface region between consecutive video frames; it maps the local motion vectors to the river coordinate space established based on the river extension direction, and aggregates them according to spatial location and temporal window to obtain the basic transport parameters corresponding to each water surface position; it constrains and corrects the basic transport parameters based on the motion continuity between adjacent water surface positions and the temporal stability of the same water surface position, and determines the corresponding motion change range; based on the corrected basic transport parameters and motion change range, it generates the basic motion field of the water surface that is updated with the video time sequence.

[0083] In some embodiments, Figure 2 The analysis module 203 performs temporal transport of water surface texture based on the current video frame and the basic motion field of the water surface, and compensates for the transport results according to the temporal changes in the appearance of the water surface to generate the water surface evolution state; it performs multi-scale difference calculation between the water surface evolution state and the actual water surface state of the corresponding video frame, and corrects the difference credibility according to the motion change range corresponding to the basic motion field of the water surface to generate water surface anomaly response; it performs spatiotemporal aggregation of water surface anomaly responses in consecutive video frames to generate water surface anomaly regions; it determines the transport arrival range of each water surface anomaly region in subsequent video frames based on the basic motion field of the water surface, and establishes temporal associations based on the state matching relationship between water surface anomaly regions within the transport arrival range to form candidate trajectories.

[0084] In some embodiments, Figure 2 The prediction module 204 extracts the regional representations corresponding to the water surface evolution state and the actual water surface state according to the preset scale levels, and generates scale difference responses based on the degree of state deviation under the same scale level; it performs cross-scale fusion based on the spatial overlap and temporal continuity relationship between different scale difference responses to generate an initial abnormal response; it reads the range of motion change of the basic motion field of the water surface at the corresponding water surface position, and determines the response credibility weight based on the degree of deviation of the actual water surface motion relative to the range of motion change; it uses the response credibility weight to correct the initial abnormal response and suppress local responses lacking temporal continuity to generate an abnormal water surface response.

[0085] In some embodiments, Figure 2The matching module 205 extracts the state features of the water surface anomaly area corresponding to the candidate trajectory and the state of the surrounding water surface to generate the initial state of the candidate trajectory. Under the assumption of water surface disturbance state, the surrounding water surface state is used to replace the water surface anomaly area, and the temporal changes of the water surface state after replacement are deduced according to the basic motion field of the water surface to generate the counterfactual prediction result of water surface disturbance. Under the assumption of floating object transport state, the object state continuity of the water surface anomaly area is maintained. Combining the basic motion field of the water surface and the motion state of the candidate trajectory relative to the water surface, the position and morphological changes of the water surface anomaly area are deduced to generate the counterfactual prediction result of floating object transport. The counterfactual prediction results are associated with the corresponding state assumption and the candidate trajectory.

[0086] In some embodiments, Figure 2 The matching module 205 determines the motion state of the abnormal water surface region relative to the water surface based on the difference between the cross-frame displacement of the candidate trajectory and the water surface transport state at the corresponding position; it superimposes the motion state onto the basic motion field of the water surface and infers the transport position of the abnormal water surface region according to the prediction time sequence; it determines the morphological evolution constraints based on the state continuity relationship of the abnormal water surface region in the candidate trajectory, and updates the region state corresponding to each transport position based on the morphological evolution constraints; it corrects the prediction range of the region state using the motion change range of the basic motion field of the water surface, and temporally correlates the transport position and region state corresponding to each prediction time to generate the counterfactual prediction result of floating object transport.

[0087] In some embodiments, Figure 2 The generation module 206 extracts the actual observed states corresponding to the candidate trajectories from subsequent video frames and performs state matching with the counterfactual prediction results of water surface disturbance and floating object transport according to the predicted time sequence. Based on the deviation between the predicted position and the actual position at each time and the degree of preservation between the predicted state and the actual observed state, the temporal matching degree corresponding to each state hypothesis is generated. Based on the range of motion changes of the water surface basic motion field and the effectiveness of the actual observed state, the credibility of each temporal matching degree is corrected to obtain the hypothesis explanatory power. The hypothesis explanatory power corresponding to the water surface disturbance state hypothesis and the floating object transport state hypothesis is compared, and the target attributes of the candidate trajectory are determined according to the difference in explanatory power and the time sequence stability.

[0088] In some embodiments, Figure 2The generation module 206 performs continuity correction on the trajectory nodes of floating objects in the states of obstruction, adhesion, or separation based on the temporal correlation in the candidate trajectories to obtain effective floating object trajectories; based on the effective floating object trajectories and the corresponding water surface transport states, it determines the motion characteristics of the floating objects relative to the water surface and generates the position changes and motion states of the floating objects; combining the target attribute judgment results of the candidate trajectories and the continuity of the trajectories, it generates the floating object detection credibility; when the floating object detection credibility meets the confirmation conditions, it associates the effective floating object trajectory, motion state, and corresponding spatiotemporal information to generate the floating object dynamic detection results.

[0089] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0090] Figure 3 This is a schematic diagram of the electronic device 3 provided in an embodiment of this application. Figure 3 As shown, the electronic device 3 of this embodiment includes: a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various system embodiments described above.

[0091] Electronic device 3 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 3 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0092] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0093] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0095] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which may be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0096] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for dynamic detection of floating objects in river monitoring videos, characterized in that, include: Acquire monitoring video of the target river section, perform image stabilization correction on continuous video frames in the monitoring video, and determine the dynamic water area based on the temporal changes of the water surface boundary; Water surface motion information is extracted from the dynamic water area and correlated with the spatial location of the river channel and the video sequence to generate a basic water surface motion field that represents the water surface transport status and range of change. Based on the water surface basic motion field, the water surface evolution state under the condition of no floating objects is predicted. The difference between the water surface evolution state and the actual water surface state of the corresponding video frame is analyzed to generate water surface abnormal regions. The water surface abnormal regions at different times are associated with the water surface basic motion field to form candidate trajectories. For the candidate trajectory, water surface disturbance state assumptions and floating object transport state assumptions are established respectively. Based on each state assumption, the state changes of the candidate trajectory in subsequent video frames are predicted to obtain the corresponding counterfactual prediction results. The counterfactual prediction results are matched with the actual observed states in the subsequent video frames, and the target attributes of the candidate trajectories are determined based on the matching differences corresponding to each state hypothesis. When the target attribute characterizes a floating object, dynamic detection results of the floating object are generated based on the candidate trajectory and the corresponding water surface transport state.

2. The method according to claim 1, characterized in that, The process of acquiring monitoring video of the target river section, performing image stabilization correction on continuous video frames in the monitoring video, and determining dynamic water areas based on the temporal changes of the water surface boundary includes: Extract cross-frame stable static reference features from the non-water areas of the continuous video frames, and determine the image transformation relationship based on the positional changes of the static reference features in adjacent video frames; Based on the aforementioned image transformation relationship, each video frame is registered to a unified spatial reference to obtain a stable video frame. The water surface boundary is extracted from the stabilized video frame, and the effective boundary caused by water level change and the instantaneous interference boundary are distinguished based on the consistency of the water surface boundary change within a preset time window. The water area corresponding to the effective boundary is merged, and the continuously occluded area is removed to generate the dynamic water area that is updated with the video time sequence.

3. The method according to claim 1, characterized in that, The step of extracting water surface movement information from the dynamic water area, associating it according to the spatial location of the river channel and the video time sequence, and generating a basic water surface motion field representing the water surface transport status and range of change includes: From the dynamic water area, continuously abnormal regions are screened out, and the local motion vectors of the remaining water surface region between consecutive video frames are extracted. The local motion vectors are mapped to the river coordinate space established based on the river extension direction, and aggregated according to spatial location and temporal window to obtain the basic transport parameters corresponding to each water surface position. Based on the continuity of motion between adjacent water surface locations and the temporal stability of the same water surface location, the basic transport parameters are constrained and corrected, and the corresponding range of motion variation is determined. Based on the corrected basic transport parameters and the range of motion changes, the basic motion field of the water surface is generated and updated with the video sequence.

4. The method according to claim 1, characterized in that, The method predicts the water surface evolution state under conditions without floating objects based on the basic motion field of the water surface, performs difference analysis between the water surface evolution state and the actual water surface state of the corresponding video frame, generates water surface anomaly regions, and forms candidate trajectories by associating the water surface anomaly regions at different times according to the basic motion field of the water surface, including: Based on the current video frame and the basic motion field of the water surface, the water surface texture is transported in a temporal sequence, and the transport result is compensated according to the temporal changes in the appearance of the water surface to generate the water surface evolution state. The water surface evolution state is compared with the actual water surface state of the corresponding video frame using multi-scale difference calculation, and the difference confidence is corrected according to the motion change range corresponding to the basic motion field of the water surface to generate an abnormal water surface response. Spatiotemporal aggregation is performed on the abnormal water surface responses in consecutive video frames to generate the abnormal water surface region. Based on the water surface basic motion field, the transport arrival range of each of the water surface anomaly regions in subsequent video frames is determined, and a temporal association is established according to the state matching relationship between the water surface anomaly regions within the transport arrival range to form the candidate trajectory.

5. The method according to claim 4, characterized in that, The step of performing multi-scale difference calculations between the water surface evolution state and the actual water surface state in the corresponding video frame, and correcting the difference reliability based on the motion change range corresponding to the basic motion field of the water surface to generate anomaly responses, includes: According to the preset scale levels, the regional representations corresponding to the water surface evolution state and the actual water surface state are extracted respectively, and the scale difference response is generated based on the degree of state deviation under the same scale level. Cross-scale fusion is performed based on the spatial overlap and temporal continuity relationships between responses at different scales to generate initial anomalous responses; Read the range of motion change of the basic motion field of the water surface at the corresponding water surface position, and determine the response confidence weight based on the degree of deviation of the actual water surface motion from the range of motion change; The initial anomalous response is corrected using the response confidence weight, and local responses lacking temporal continuity are suppressed to generate the water surface anomalous response.

6. The method according to claim 1, characterized in that, The process involves establishing water surface disturbance state assumptions and floating object transport state assumptions for the candidate trajectories, predicting the state changes of the candidate trajectories in subsequent video frames based on each state assumption, and obtaining corresponding counterfactual prediction results, including: Extract the state features of the water surface anomaly area corresponding to the candidate trajectory and the state of the surrounding water surface to generate the initial state of the candidate trajectory; Under the assumption of the water surface disturbance state, the surrounding water surface state is used to replace the abnormal water surface area, and the temporal change of the replaced water surface state is deduced according to the basic motion field of the water surface to generate the counterfactual prediction result of the water surface disturbance. Under the assumption of the floating object transport state, the object state continuity of the abnormal water surface area is maintained. Combining the basic motion field of the water surface and the motion state of the candidate trajectory relative to the water surface, the position and morphological changes of the abnormal water surface area are deduced to generate the counterfactual prediction result of the floating object transport. The counterfactual prediction results are correlated with the corresponding state assumptions and candidate trajectories.

7. The method according to claim 6, characterized in that, The method combines the basic motion field of the water surface and the motion state of the candidate trajectory relative to the water surface to deduce the location and morphological changes of the abnormal water surface area, generating counterfactual prediction results for floating object transport, including: The motion state of the abnormal water surface region relative to the water surface is determined based on the difference between the cross-frame displacement of the candidate trajectory and the water surface transport state at the corresponding position. The motion state is superimposed onto the basic motion field of the water surface, and the transport location of the abnormal water surface area is deduced according to the predicted time sequence. The morphological evolution constraints are determined based on the state continuity relationship of the water surface anomaly region in the candidate trajectory, and the region state corresponding to each transport position is updated based on the morphological evolution constraints. The predicted range of the regional state is corrected by using the range of motion changes of the basic motion field on the water surface, and the transport position and regional state corresponding to each prediction time are temporally correlated to generate the counterfactual prediction result of the floating object transport.

8. The method according to claim 1, characterized in that, The step of matching each of the counterfactual prediction results with the actual observed state in the subsequent video frames, and determining the target attribute of the candidate trajectory based on the matching difference corresponding to each state assumption, includes: Extract the actual observation state corresponding to the candidate trajectory from the subsequent video frames, and match the state with the water surface disturbance counterfactual prediction result and the floating object transport counterfactual prediction result according to the prediction time sequence; Based on the deviation between the predicted and actual positions at each time point and the degree of preservation between the predicted state and the actual observed state, the temporal matching degree corresponding to each state hypothesis is generated. Based on the range of motion changes of the basic motion field on the water surface and the validity of the actual observation state, the credibility of each time series matching degree is corrected to obtain the hypothesis explanatory power. The explanatory power of the hypotheses corresponding to the water surface disturbance state hypothesis and the floating object transport state hypothesis is compared, and the target attributes of the candidate trajectories are determined based on the difference in explanatory power and the time series stability.

9. The method according to claim 1, characterized in that, When the target attribute characterizes a floating object, the process of generating a dynamic detection result for the floating object based on the candidate trajectory and the corresponding water surface transport state includes: Based on the temporal correlation in the candidate trajectories, the trajectory nodes of the floating objects in the states of obstruction, adhesion, or separation are continuously corrected to obtain effective floating object trajectories. Based on the effective floating object trajectory and the corresponding water surface transport state, the motion characteristics of the floating object relative to the water surface are determined, and the position change and motion state of the floating object are generated. Based on the target attribute determination results and trajectory continuity of the candidate trajectories, a floating object detection confidence level is generated. When the confidence level of the floating object detection meets the confirmation conditions, the effective floating object trajectory, the motion state, and the corresponding spatiotemporal information are associated to generate the dynamic detection result of the floating object.

10. A dynamic floating object detection system for river monitoring videos, characterized in that, include: The acquisition module is used to acquire monitoring videos of the target river section, perform image stabilization correction on continuous video frames in the monitoring video, and determine the dynamic water area based on the temporal changes of the water surface boundary. The extraction module is used to extract water surface motion information from the dynamic water area, associate it with the spatial location of the river channel and the video time sequence, and generate a basic water surface motion field that represents the water surface transport status and range of change. The analysis module is used to predict the water surface evolution state under the condition of no floating objects based on the water surface basic motion field, perform difference analysis between the water surface evolution state and the actual water surface state of the corresponding video frame, generate water surface abnormal regions, and form candidate trajectories by associating water surface abnormal regions at different times according to the water surface basic motion field. The prediction module is used to establish water surface disturbance state assumptions and floating object transport state assumptions for the candidate trajectories, respectively, and predict the state changes of the candidate trajectories in subsequent video frames based on each state assumption to obtain the corresponding counterfactual prediction results. The matching module is used to match each of the counterfactual prediction results with the actual observed state in the subsequent video frames, and to determine the target attribute of the candidate trajectory based on the matching difference corresponding to each state assumption. The generation module is used to generate dynamic detection results of floating objects based on the candidate trajectories and the corresponding water surface transport states when the target attribute characterizes the floating object.

Images