Monitoring video condensation method based on scene topology decomposition and parallel scheduling
By employing scene topology decomposition and parallel scheduling, channel areas and activity areas in surveillance videos are identified and rearranged in parallel. Target tubes are classified and differentiated algorithms are designed, solving the efficiency and adaptability issues of video condensation in complex scenarios and achieving efficient, collision-free video rearrangement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHIJIAZHUANG TIEDAO UNIV
- Filing Date
- 2026-03-06
- Publication Date
- 2026-05-19
AI Technical Summary
Existing video condensation technologies have high computational complexity in complex scenarios, making it difficult to achieve efficient parallel reordering. Furthermore, homogeneous reordering cannot adapt to the topological structures of different scenarios, resulting in disordered target traversal timing in the channel area.
By decomposing scene topology and parallel scheduling, channel areas and activity areas are identified, and target tubes are classified as channel type, activity type and hybrid type. Differentiated rearrangement algorithms are designed for different categories, and efficient collision-free rearrangement is achieved by using spatiotemporal capacity modeling and conflict graph simulation.
It improves the efficiency and adaptability of video condensation, reduces false collisions and temporal disorder in condensed videos, and ensures the integrity and logical coherence of target trajectories.
Smart Images
Figure CN121789119B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of video compression technology, and specifically to a method for compressing surveillance videos based on scene topology decomposition and parallel scheduling. Background Technology
[0002] In the fields of smart cities, security monitoring, and traffic management, the volume of surveillance video data is growing exponentially, making video summarization technology a core technology for quickly browsing and analyzing surveillance videos. Target optimization and rearrangement are key aspects of video summarization technology, which shortens video length while retaining key information by rearranging the spatiotemporal trajectory of the target.
[0003] In existing video summarization technologies, target tube rearrangement often adopts a homogeneous optimization framework, incorporating all target tubes into the same system for collision processing and trajectory adjustment. This presents the following core technical problems: First, the computational complexity of the rearrangement algorithm is high. When there are many targets in the monitoring scene, the overall rearrangement efficiency drops significantly, making it difficult to meet the needs of real-time video summarization. Second, traditional methods do not consider the differences in the topological structure of the monitoring scene. For example, the trajectory characteristics of corridor and intersection passage-type scenes are completely different from those of square and hall activity-type scenes. Homogeneous rearrangement cannot adapt to the display needs of different scenes and easily leads to the temporal disorder of target crossing in the passage area.
[0004] Some studies have attempted to improve the rearrangement effect by processing different scenarios, but there are still problems such as the reliance on manual annotation for scenario topology recognition, the lack of quantitative standards for target tube classification, and the failure to achieve true parallel rearrangement, making it difficult to achieve efficient optimization and rearrangement of target tubes in complex scenarios. Summary of the Invention
[0005] This invention provides a method for condensing surveillance videos based on scene topology decomposition and parallel scheduling, aiming to solve the problems of low efficiency and poor effect of traditional homogeneous rearrangement, and improve the quality and efficiency of video condensation in complex scenes. The method includes the following steps:
[0006] S1: Input the original surveillance video and obtain the video background image;
[0007] S2: Extract the trajectory information of all targets in the video to form a target tube;
[0008] S3: Project the trajectory points of all target tubes onto a unified reference frame to construct a trajectory point cloud. Perform spatial clustering analysis on the trajectory point cloud to identify the channel area and activity area in the video scene. The unified reference frame is the video background image.
[0009] S4: Based on the relationship between the trajectory of the target tube and the channel area and the active area, the target tube is classified into channel-type target tube, active target tube and hybrid target tube. The hybrid target tube is decoupled on the time axis according to the intersection relationship between the trajectory and the topological region to obtain the channel sub-segment and the active sub-segment.
[0010] S5: Perform differentiated parallel rearrangement operations for target pipes and sub-segments of different trajectory categories: For channel area pipes composed of channel-type target pipes and channel sub-segments, perform a rearrangement algorithm based on spatiotemporal capacity modeling and gap filling scheduling; for active area pipes composed of active target pipes and active sub-segments, perform a rearrangement algorithm based on conflict graph construction and hierarchical force-directed simulation.
[0011] S6: Perform global timeline integration on the channel area pipelines and active area pipelines after parallel rearrangement to complete the overall optimization and rearrangement of the target pipelines;
[0012] S7: The target tube and video background image will be merged after overall optimization and rearrangement to generate a condensed video.
[0013] Preferably, the step of extracting the trajectory information of all targets in the video to form a target tube specifically includes:
[0014] The video frames are processed using object detection algorithms and DeepSORT multi-object tracking algorithms to obtain the detection box sequence of each object in consecutive frames;
[0015] Each target is assigned a unique identifier, and its spatial location, movement speed, movement direction, and interaction behavior category are recorded, which are then integrated to form a target management system that spans the time dimension.
[0016] Preferably, the step of performing spatial clustering analysis on the trajectory point cloud to identify channel areas and activity areas in the video scene specifically includes:
[0017] The starting and ending positions of the target tube are clustered based on the density clustering algorithm, and the dense regions generated by the clustering are identified as channel areas;
[0018] Based on the kernel density estimation algorithm, the spatial distribution of all trajectory points is density estimated, and connected regions with density higher than a preset threshold are identified as active regions. The kernel density estimation formula is as follows:
[0019] ,
[0020] Where f(x) is the kernel density estimate at spatial location x; n is the total number of trajectory points involved in the estimation; h is the kernel function bandwidth, used to control the smoothness of the density estimation; and x is the spatial coordinate of the location to be estimated. Let i be the spatial coordinates of the i-th trajectory point; This is the Gaussian kernel function.
[0021] Preferably, the correlation between the trajectory of the target tube and the channel area and the active area classifies the target tube into channel-type target tubes, active-type target tubes, and hybrid-type target tubes, specifically including:
[0022] If the trajectory point set of the target tube satisfies If so, the target tube is classified as a channel-type target tube;
[0023] If the trajectory point set of the target tube satisfies If so, the target tube is classified as an active target tube;
[0024] If the trajectory point set of the target tube satisfies and If so, the target tube is classified as a hybrid target tube;
[0025] in, It is the set of all trajectory points of the i-th target tube; The set of spatial extents of the passageway area; This refers to the set of spatial boundaries of the activity area.
[0026] Preferably, the decoupling of the hybrid target tube on the time axis according to the intersection relationship between the trajectory and the topological region to obtain the channel segment and the activity segment specifically includes:
[0027] Obtain the complete spatiotemporal trajectory of the hybrid target tube, determine the position where its trajectory first crosses the boundary between the channel area and the activity area, and the corresponding timestamp; using this timestamp as the dividing point, split the original target tube into two independent sub-segments on the timeline, with the part located in the channel area marked as the channel sub-segment and the part located in the activity area marked as the activity sub-segment; the two sub-segments inherit the original target's identity, behavioral attributes, and all recorded information, and record the relationship between them, so as to restore the time sequence after subsequent rearrangement.
[0028] Preferably, for the channel area pipeline consisting of a channel-type target pipe and channel segments, the reordering algorithm based on spatiotemporal capacity modeling and gap filling scheduling is executed, specifically including:
[0029] Construct a spatiotemporal capacity function for the channel region to quantify the spatial carrying capacity of the channel region at different time points;
[0030] Establish a spatiotemporal occupancy model for pipelines in the channel area to characterize their spatial occupancy range and temporal duration within the channel area;
[0031] The objective function is to minimize the travel time.
[0032] ,
[0033] Where J is the optimization target value; M is the total number of pipes in the channel area; The time it takes for the pipe in the k-th channel area to enter the channel area; The time when the pipe in the k-th channel area leaves the channel area;
[0034] Constrained by the absence of collisions and the range of movement speed, a gap filling scheduling algorithm is designed to enable the pipes in the channel area to adaptively adjust their movement speed to fill the space gaps and achieve efficient crossing arrangement.
[0035] Preferably, for the active zone pipeline consisting of active target pipes and active segments, the reordering algorithm based on conflict graph construction and hierarchical force-directed simulation is executed, specifically including:
[0036] Based on the spatiotemporal overlap of the pipelines in the activity area, a target pipeline conflict graph of the activity area is constructed, where nodes are pipelines in the activity area and the weight of the edges is the spatiotemporal overlap between nodes.
[0037] A hierarchical scheduling strategy is adopted to fix the spatiotemporal location of key activity areas with high semantic importance;
[0038] For pipelines in non-critical activity areas, an iterative optimization method based on force-directed simulation is applied, defining the total system energy as:
[0039] ,
[0040] in, N represents the total system energy; N is the total number of pipes in non-critical activity areas. Let be the time label after the pipeline rearrangement in the i-th activity area. Its original time label; Let be the spatiotemporal Euclidean distance between the i-th and j-th activity zone pipes. To prevent extremely small constants with a denominator of zero; For attractiveness weighting coefficient, These are the repulsion force weighting coefficients, used to balance the priority of preserving the original timing sequence and eliminating collisions;
[0041] The optimal allocation of time tags for activity zone pipelines is achieved by minimizing the total system energy; the key activity zone pipelines with high semantic importance are those containing interactive behaviors, which are identified by a behavior recognition model, such as waving, talking, and pointing.
[0042] Preferably, the specific process of global timeline integration is as follows: aligning the channel area pipes and activity area pipes after parallel rearrangement according to the time logic of the original trajectory, restoring the original time causal relationship of the channel sub-segments and activity sub-segments of the hybrid target pipe, and constructing a unified global rearrangement timeline.
[0043] Compared with the prior art, the present invention has at least the following beneficial effects:
[0044] 1. This invention deeply integrates scene topology decomposition and parallel scheduling into the surveillance video condensation process, realizing refined classification and differentiated rearrangement of target tubes from the scene cognition level. It breaks through the technical bottleneck of homogeneous rearrangement in traditional video condensation technology, solves the core problem of difficulty in balancing compression efficiency and information retention in complex scenarios, and improves the overall effect and adaptability of video condensation.
[0045] 2. This invention decouples the time axis of the hybrid target tube while retaining the original identity and behavioral attributes, which not only realizes the targeted rearrangement of target trajectories under different topological regions, but also ensures the integrity and logical coherence of the target trajectory.
[0046] 3. This invention designs differentiated parallel reordering algorithms for target pipes and sub-segments of different trajectory categories. The pipes in the channel area achieve efficient collision-free crossing through spatiotemporal capacity modeling and gap filling scheduling. The pipes in the active area are reordered through conflict graphs and hierarchical force-guided simulation. Parallel processing reduces the computational complexity of the reordering algorithm and improves the processing efficiency of video condensation.
[0047] 4. This invention restores the original time logic and causal relationships of all target tubes through global timeline integration. Combined with the accuracy of differentiated parallel rearrangement, it effectively reduces pseudo-collisions and timing disorder in condensed videos while achieving a high compression ratio, thus improving the visual smoothness of condensed videos. Attached Figure Description
[0048] The features and advantages of this invention will become more apparent by reading the following illustrations for a detailed explanation:
[0049] Figure 1 This is a flowchart of the surveillance video condensation method based on scene topology decomposition and parallel scheduling in an embodiment of the present invention;
[0050] Figure 2 This is a technical roadmap of the surveillance video condensation method based on scene topology decomposition and parallel scheduling in the embodiments of the present invention;
[0051] Figure 3 This is a scene classification result diagram in an embodiment of the present invention;
[0052] Figure 4 This is a schematic diagram of trajectory classification in an embodiment of the present invention. Detailed Implementation
[0053] To enable those skilled in the art to better understand the present invention, the following description is provided in conjunction with specific embodiments.
[0054] like Figure 1The flowchart shown is a process for a surveillance video condensation method based on scene topology decomposition and parallel scheduling provided by the present invention, including:
[0055] S1: Input the original surveillance video and obtain the video background image;
[0056] S2: Extract the trajectory information of all targets in the video to form a target tube;
[0057] S3: Project the target tube trajectory points onto a unified reference frame to construct a trajectory point cloud, perform spatial clustering analysis on the trajectory point cloud, and identify the channel area and activity area;
[0058] S4: Classify target tubes according to the relationship between the trajectory and the channel area and the activity area, and decouple the hybrid target tubes on the time axis to obtain the channel sub-segment and the activity sub-segment;
[0059] S5: Perform differentiated parallel rearrangement for target pipes and segments of different trajectory categories. The channel area pipes correspond to the spatiotemporal capacity modeling and gap filling scheduling rearrangement algorithm, and the active area pipes correspond to the conflict graph construction and hierarchical force-guided simulation rearrangement algorithm.
[0060] S6: Perform global timeline integration on the channel area pipelines and active area pipelines after parallel rearrangement to complete the overall optimization and rearrangement of the target pipelines;
[0061] S7: The optimized and rearranged target tube is merged with the video background image to generate a condensed video.
[0062] The overall technical route of the method of the present invention is as follows: Figure 2 As shown.
[0063] Example 1:
[0064] This invention provides a preferred embodiment for executing S1, which involves inputting the original surveillance video, obtaining a video background image, and using a Gaussian mixture model to model the background of the input video. Specifically, the first N frames of the video are randomly sampled, the background model is initialized using the MOG2 algorithm, and in subsequent processing, frames with no target motion or very few targets are periodically selected to perform median fusion updates on the background to obtain a clean video background image.
[0065] Example 2:
[0066] This invention provides a preferred embodiment for executing S2, extracting trajectory information of all targets in the video to form a target tube. YOLOv8 is used as the target detector to perform pedestrian detection on each frame of the video. Simultaneously, the DeepSORT multi-target tracking algorithm is employed, utilizing its cascaded matching mechanism and joint measurement of appearance and motion features to assign a globally unique ID to each detected target. For each successfully tracked target i, its detection bounding box information in all frames is recorded: top-left corner coordinates, width w, and height h, forming the target tube. Simultaneously record the target's spatial position, movement speed v, and movement direction. It also obtains the target's interactive behavior categories, such as "waving", "talking", and "pointing", through a behavior recognition model based on CNN+GRU.
[0067] Example 3:
[0068] The present invention provides a preferred embodiment for executing S3, performing spatial clustering analysis on the trajectory point cloud to identify channel regions and activity regions.
[0069] S31: Put all target tubes The trajectory points are obtained by projecting the center point (x, y) at the bottom of the detection box onto a unified video frame coordinate system, thus forming a two-dimensional "trajectory point cloud" set. .
[0070] S32: Perform channel area identification: Filter out the starting points of all target tubes from the trajectory point cloud. and the end point This forms an endpoint set. The DBSCAN density clustering algorithm is then applied to this endpoint set, using Euclidean distance as the similarity metric. By setting the neighborhood radius Eps and the minimum number of points MinPts, the algorithm clusters dense endpoint regions into clusters. These clusters represent locations where the target frequently appears or disappears, i.e., scene entrances and exits. The main passageway regions connecting these entrances and exits are identified as passageways. Using a binary mask express, This indicates that the pixel belongs to the channel region; otherwise, it is 0.
[0071] S33: Perform active region identification: Apply a kernel density estimation algorithm to the entire trajectory point cloud. The calculation formula is as follows:
[0072] ,
[0073] Where f(x) is the kernel density estimate at spatial location x; n is the total number of trajectory points involved in the estimation; h is the kernel function bandwidth, used to control the smoothness of the density estimation; and x is the spatial coordinate of the location to be estimated. Let i be the spatial coordinates of the i-th trajectory point; The kernel function is a Gaussian kernel, and the probability density function is taken as a standard normal distribution. The density distribution map of the entire scene is calculated using this formula, and a density threshold is set. The threshold value is adaptively determined based on the overall distribution density of trajectory points in the scene. Specifically, the kernel density estimates of all trajectory points are sorted in ascending order, and the median value is taken as the threshold T. density In this embodiment, T is determined based on experiments. density A value of 0.55 is used to classify regions with density higher than this threshold as areas of high target activity. The largest connected component in these high-density regions is extracted as the active region. Using a binary mask Indicates. For example... Figure 3 The image shown is a scene classification result diagram in an embodiment of the present invention.
[0074] Example 4:
[0075] This invention provides a preferred embodiment for executing S4, which first classifies the target tubes according to the geometric inclusion relationship between the trajectory point set of each target tube and the identified channel area and active area. Let the set of all trajectory points of the i-th target tube be denoted as... The spatial range of the passage area is set as follows The spatial range of the activity area is set as follows The classification logic is as follows:
[0076] like If so, the target tube is classified as a channel-type target tube;
[0077] like If so, the target tube is classified as an active target tube;
[0078] like and If so, the target tube is classified as a hybrid target tube.
[0079] like Figure 4 The diagram shown is a schematic diagram of trajectory classification based on scene decomposition in an embodiment of the present invention.
[0080] For hybrid target tubes, whose trajectories span both the channel and activity zones, two possible scenarios exist: either the target tube passes through the channel zone first and then enters the activity zone, or vice versa. Temporal decoupling is necessary. Specifically, the complete spatiotemporal trajectory of the hybrid target tube is obtained, determining its initial spatial location and corresponding timestamp as it enters another zone. Using this timestamp as a dividing point, the original target tube is split into two independent segments on the timeline: one in the first zone and one in the second zone. These two segments inherit the original target's identity, behavioral attributes, and all recorded information, and their interrelationships are recorded for subsequent temporal recovery after rearrangement. For ease of distinction, the segment in the channel zone is called the channel segment, and the segment in the activity zone is called the activity segment.
[0081] Example 5:
[0082] The present invention provides a preferred embodiment for executing S5.
[0083] For pipelines in the passage area, a rearrangement algorithm based on spatiotemporal capacity modeling and gap-filling scheduling is executed. The specific process is as follows:
[0084] S51: Spatiotemporal capacity modeling: Construct a spatiotemporal capacity function C(x, y, t) for the channel region, representing the spatial occupancy status at time t and position (x, y). Initially, C(x, y, t) = 1, indicating that the spatiotemporal unit is in an idle and occupant state; when a channel-type target tube occupies the spatiotemporal unit, it is marked as 0, indicating that it has been occupied.
[0085] S52: Spatiotemporal Occupancy Model: For the kth channel region pipe to be inserted Construct its spatiotemporal occupancy model , : This indicates that at time t and position (x, y), target k exists and occupies that point. This indicates that target k does not exist at this time and place. It accurately represents the spatial occupancy range and time duration of the pipeline within the channel area.
[0086] S53: Optimization Objective: The core optimization objective is to minimize the total travel time of all passageways. The objective function is:
[0087] ,
[0088] Where J is the optimization target value; M is the total number of pipes in the channel area; The time it takes for the pipe in the k-th channel area to enter the channel area; The time it takes for the pipe in the k-th channel area to leave the channel area.
[0089] S54: Constraints:
[0090] (1) No collision constraints: , C(x, y, t) is the spatiotemporal occupancy indicator function for the k-th channel area pipe. A value of 1 indicates that the pipe is occupied at time t and position (x, y), and 0 otherwise. C(x, y, t) is the capacity function of the spatiotemporal point. The initial value is 1, and it is set to 0 after being occupied to ensure that there is no spatiotemporal collision.
[0091] (2) Movement speed range constraint: the movement speed of the target in the channel area It needs to be kept within a reasonable range, that is ,in This represents the path length of the pipeline within the passageway. , These are the minimum and maximum values of the motion speed, respectively, to avoid visual artifacts caused by sudden speed changes.
[0092] S55: Gap Filling Scheduling: A greedy strategy is adopted to search for the earliest insertable time slot for each channel area pipeline in the original time sequence. During the search process, the target's movement speed in the channel area is allowed to be adaptively adjusted within the above-mentioned speed constraints to adapt to the time and space slots in the channel area. Under the premise of satisfying all constraints, the total time span of the channel area is compressed to maximize the efficiency and orderly crossing arrangement of pipelines in the channel area.
[0093] For the active area pipelines, a rearrangement algorithm based on conflict graph construction and hierarchical force-directed simulation is executed. The specific process is as follows:
[0094] S56: Conflict Graph Construction: Extracting Active Regions All active target pipes and active sub-segments of hybrid target pipes within the set A are considered. The spatiotemporal overlap of pipes i and j in any two active regions within set A is calculated. :
[0095] ,
[0096] in, The average intersection-over-union (IoU) ratio of the two target tube detection boxes in the co-occurring frame. This represents the proportion of the overlap between their time intervals. Using elements in A as nodes, and... Edges are constructed based on conditions where the edge weight is greater than a preset threshold θ (value 0.1), and the weight of the edge is... Thus, the pipeline conflict G(V,E,W) in the active area is obtained.
[0097] S57: Layered Scheduling: Mark the activity area pipeline containing interactive behaviors as "key nodes with high semantic importance". Prioritize fixing these key nodes and their first-order neighbor nodes in the conflict graph as the first layer; the remaining nodes are set as the second layer, and force-directed optimization is performed subsequently.
[0098] S58: Force-Directed Simulation: Applying a force-directed simulation iterative optimization method to non-critical nodes (second layer). Defining the system's total energy function. for:
[0099] ,
[0100] in, N represents the total system energy; N is the total number of pipes in non-critical activity areas. Let be the time label after the pipeline rearrangement in the i-th activity area. Its original time label; Let be the spatiotemporal Euclidean distance between the i-th and j-th activity zone pipes. To prevent extremely small constants with a denominator of zero; For attractiveness weighting coefficient, The two coefficients are repulsion force weighting coefficients, which are used to balance the priority of preserving the original time sequence and eliminating collisions. In this embodiment, α=0.6 and β=0.4 are set experimentally to ensure that the original time sequence characteristics of the target are preserved as much as possible while prioritizing the elimination of target collisions.
[0101] S59: By minimizing Iterative updates By converging the system energy, a new set of optimized time labels can be obtained, which avoids collisions while preserving the original distribution and interaction of targets within the activity area to the greatest extent.
[0102] Example 6:
[0103] This invention provides a preferred embodiment that executes step S6, performing global timeline integration on all target tubes and sub-segments after parallel rearrangement to complete the overall optimization and rearrangement of the target tubes. All independent units after parallel rearrangement are integrated onto a unified timeline according to their respective new time labels. Specifically, for hybrid target tubes, based on the temporal relationship between their channel sub-segments and active sub-segments in the original video, their temporal order after rearrangement is checked and adjusted to restore the original logical causal relationship: if the original order is passing through the channel area first and then entering the active area, then the end time of the channel sub-segment should be ensured to be earlier than the start time of the active sub-segment; if the original order is passing through the active area first and then entering the channel area, then the end time of the active sub-segment should be ensured to be earlier than the start time of the channel sub-segment. If the order is reversed, it is corrected by shifting the time position of the next sub-segment until the correct temporal constraints are met. Thus, the overall optimization and rearrangement of all target tubes is completed.
[0104] Example 7:
[0105] In a preferred embodiment of the present invention, S7 is performed to seamlessly paste the obtained optimized and rearranged target tubes onto the video background image according to their new time positions using image processing technology based on Poisson fusion, and output the final condensed video.
[0106] Experiment and Results Analysis:
[0107] To verify the effectiveness of the above embodiments, this invention conducted extensive experiments. Eight representative videos, containing various scenes, were selected from the VISOR, CAVIAR, and self-built datasets. The effectiveness of the condensed videos was primarily measured by frame compression ratio (FR) and collision rate (OR). The FR formula is as follows:
[0108] ,
[0109] in, To condense the video length, The FR value represents the original video length. The smaller the FR value, the higher the compression level of the condensed video compared to the original video.
[0110] The OR formula is as follows:
[0111] ,
[0112] Where w and h represent the length and width of the video background, respectively. Indicates the number of video frames. The OR value represents the number of pixels in the i-th frame of the video that collide with the target. The smaller the OR value, the fewer the target collisions and the better the visual effect.
[0113] The proposed method was compared with the particle swarm optimization method MM, the spatiotemporal offset optimization method ST, and the multi-objective balance optimization method MBO. As shown in Table 1, the proposed method has the lowest average frame compression ratio (FR) and collision rate (OR) in the experimental video.
[0114] Table 1 Comparison of Experimental Results
[0115]
[0116] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention. The various preferred features described above can be used in any combination without conflict.
Claims
1. A method for condensing surveillance video based on scene topology decomposition and parallel scheduling, characterized in that, Includes the following steps: S1: Input the original surveillance video and obtain the video background image; S2: Extract the trajectory information of all targets in the video to form a target tube, specifically including: The video frames are processed using object detection algorithms and DeepSORT multi-object tracking algorithms to obtain the detection box sequence of each object in consecutive frames; Each target is assigned a unique identifier, and its spatial location, movement speed, movement direction, and interaction behavior category are recorded, which are then integrated to form a target management system that spans the time dimension. S3: Project the trajectory points of all target tubes onto a unified reference frame to construct a trajectory point cloud. Perform spatial clustering analysis on the trajectory point cloud to identify the channel areas and activity areas in the video scene. The unified reference frame is the video background image. The spatial clustering analysis of the trajectory point cloud to identify the channel areas and activity areas in the video scene specifically includes: The starting and ending positions of the target tube are clustered based on the density clustering algorithm, and the dense regions generated by the clustering are identified as channel areas; Based on the kernel density estimation algorithm, the spatial distribution of all trajectory points is density estimated, and connected regions with density higher than a preset threshold are identified as active regions. The kernel density estimation formula is as follows: ， Where f(x) is the kernel density estimate at spatial location x; n is the total number of trajectory points involved in the estimation; h is the kernel function bandwidth, used to control the smoothness of the density estimation; and x is the spatial coordinate of the location to be estimated. Let i be the spatial coordinates of the i-th trajectory point; The Gaussian kernel function; S4: Based on the relationship between the trajectory of the target tube and the channel area and the active area, the target tube is classified into channel-type target tube, active target tube and hybrid target tube. The hybrid target tube is decoupled on the time axis according to the intersection relationship between the trajectory and the topological region to obtain the channel sub-segment and the active sub-segment. S5: Perform differentiated parallel rearrangement operations for target pipes and sub-segments of different trajectory categories: For channel area pipes composed of channel-type target pipes and channel sub-segments, perform a rearrangement algorithm based on spatiotemporal capacity modeling and gap filling scheduling; for active area pipes composed of active target pipes and active sub-segments, perform a rearrangement algorithm based on conflict graph construction and hierarchical force-directed simulation. S6: Perform global timeline integration on the channel area pipelines and active area pipelines after parallel rearrangement to complete the overall optimization and rearrangement of the target pipelines; S7: The target tube and video background image will be merged after overall optimization and rearrangement to generate a condensed video.
2. The surveillance video compression method based on scene topology decomposition and parallel scheduling as described in claim 1, characterized in that, The correlation between the trajectory of the target tube and the channel area and the active area classifies the target tubes into channel-type target tubes, active target tubes, and hybrid target tubes, specifically including: If the trajectory point set of the target tube satisfies If so, the target tube is classified as a channel-type target tube; If the trajectory point set of the target tube satisfies If so, the target tube is classified as an active target tube; If the trajectory point set of the target tube satisfies and If so, the target tube is classified as a hybrid target tube; Among them, P i S is the set of all trajectory points of the i-th target tube; ch S is the set of spatial extents of the passageway area; ac This refers to the set of spatial boundaries of the activity area.
3. The surveillance video compression method based on scene topology decomposition and parallel scheduling as described in claim 1, characterized in that, The process of decoupling the hybrid target tube along the time axis based on the intersection relationship between the trajectory and the topological region to obtain the channel segment and the active segment specifically includes: Obtain the complete spatiotemporal trajectory of the hybrid target tube, determine the position where its trajectory first crosses the boundary between the channel area and the activity area, and the corresponding timestamp; using this timestamp as the dividing point, split the original target tube into two independent sub-segments on the timeline, with the part located in the channel area marked as the channel sub-segment and the part located in the activity area marked as the activity sub-segment; the two sub-segments inherit the original target's identity, behavioral attributes, and all recorded information, and record the relationship between them, so as to restore the time sequence after subsequent rearrangement.
4. The surveillance video compression method based on scene topology decomposition and parallel scheduling as described in claim 1, characterized in that, For the channel area pipeline consisting of a channel-type target pipe and channel segments, a rearrangement algorithm based on spatiotemporal capacity modeling and gap-filling scheduling is executed, specifically including: Construct a spatiotemporal capacity function for the channel region to quantify the spatial carrying capacity of the channel region at different time points; Establish a spatiotemporal occupancy model for pipelines in the channel area to characterize their spatial occupancy range and temporal duration within the channel area; The objective function is to minimize the travel time. ， Where J is the optimization target value; M is the total number of pipes in the channel area; The time it takes for the pipe in the k-th channel area to enter the channel area; The time when the pipe in the k-th channel area leaves the channel area; Constrained by the absence of collisions and the range of movement speed, a gap filling scheduling algorithm is designed to enable the pipes in the channel area to adaptively adjust their movement speed to fill the space gaps and achieve efficient crossing arrangement.
5. The surveillance video condensation method based on scene topology decomposition and parallel scheduling as described in claim 1, characterized in that, For the active zone pipeline consisting of active target pipes and active segments, a rearrangement algorithm based on conflict graph construction and hierarchical force-directed simulation is executed, specifically including: Based on the spatiotemporal overlap of the pipelines in the activity area, a target pipeline conflict graph of the activity area is constructed, where nodes are pipelines in the activity area and the weight of the edges is the spatiotemporal overlap between nodes. A hierarchical scheduling strategy is adopted to fix the spatiotemporal location of key activity areas with high semantic importance; For pipelines in non-critical activity areas, an iterative optimization method based on force-directed simulation is applied, defining the total system energy as: ， in, N represents the total system energy; N is the total number of pipes in non-critical activity areas. Let be the time label after the pipeline rearrangement in the i-th activity area. Its original time label; Let be the spatiotemporal Euclidean distance between the i-th and j-th activity zone pipes. To prevent extremely small constants with a denominator of zero; For attractiveness weighting coefficient, These are the repulsion force weighting coefficients, used to balance the priority of preserving the original timing sequence and eliminating collisions; The optimal allocation of time tags for activity zone pipelines is achieved by minimizing the total system energy; the key activity zone pipelines with high semantic importance are those containing interactive behaviors, which are identified by a behavior recognition model, such as waving, talking, and pointing.
6. The surveillance video compression method based on scene topology decomposition and parallel scheduling as described in claim 1, characterized in that, The specific process of global timeline integration is as follows: the channel area pipelines and activity area pipelines after parallel rearrangement are aligned according to the time logic of the original trajectory; the original time causal relationship of the channel sub-segments and activity sub-segments of the hybrid target pipe is restored; and a unified global rearrangement timeline is constructed.