Distributed real-time target detection tracking method and system based on dynamic perception and adaptive regulation and control

Abstract

The invention discloses a distributed real-time target detection tracking method and system based on dynamic perception and adaptive regulation and control, and belongs to the technical field of intelligent monitoring and computer vision. The method comprises the following steps: firstly, calculating the weight of each camera in real time by using the time, position and map information of an event and camera network topology, generating an optimal search path by predicting a target motion track, and decomposing a global tracking task into multiple levels of subtasks; then, according to the task complexity and the node resource state, adopting a hybrid scheduling strategy to realize short-term container elastic expansion and contraction and long-term resource prediction, and ensuring efficient and stable resource scheduling; then, through multi-modal feature fusion and spatio-temporal joint modeling, cross-camera video data alignment, target association and continuous tracking are realized; and finally, performing multi-source data fusion and risk reasoning by using the large model, and outputting graded early warning and three-dimensional visual decision information. According to the method, the tracking accuracy can be improved, and resource utilization optimization and real-time response are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent monitoring and computer vision, and particularly relates to a distributed real-time target detection and tracking method and system based on dynamic perception and adaptive regulation, which is suitable for mobile target continuous tracking and risk early warning in the fields of public security and traffic management. BACKGROUND

[0002] The existing monitoring system has the following technical defects: (1) The camera selection strategy is static: the traditional scheme selects the camera based on fixed rules (such as distance priority), which cannot dynamically adjust the search path according to the target motion state, resulting in a target loss rate of up to 30%-40%. (2) The allocation of computing resources is rigid: the centralized processing architecture cannot adapt to sudden task loads, and the GPU resource utilization rate fluctuates by more than 60%, causing a response delay of more than 500ms. (3) The cross-lens correlation accuracy is low: the existing method relies on single appearance feature matching, and the ID switching rate is more than 15% in the case of view angle change and occlusion. (4) The risk early warning lacks spatio-temporal reasoning: the traditional alarm system is based on single-frame detection results, and lacks deep analysis of multi-target behavior patterns and spatio-temporal correlation, with a false alarm rate of more than 25%. SUMMARY

[0003] In view of the above technical problems, the application proposes a distributed real-time target detection and tracking scheme based on dynamic perception and adaptive regulation.

[0004] The first aspect of the application proposes a distributed real-time target detection and tracking method based on dynamic perception and adaptive regulation, in which a master node, a message middleware and a plurality of distributed computing nodes are configured, and the master node is further configured with a dynamic task analysis and path planning module, an adaptive resource optimization deployment module, a cross-lens collaboration module and a data fusion and decision module. The method specifically includes:

[0005] The dynamic task analysis and path planning module is called to use the time, location, map information and camera network topology of the event to calculate the weight of each camera in real time, and to generate an optimal search path by predicting the target motion trajectory, so as to decompose the global tracking task into multiple sub-tasks;

[0006] The adaptive resource optimization deployment module is called to use a hybrid scheduling strategy to realize container elastic expansion and contraction and resource prediction according to the task complexity and node resource state, and to perform resource scheduling;

[0007] The cross-lens collaboration module is called to perform cross-camera video data alignment, target association and continuous tracking through multi-modal feature fusion and spatio-temporal joint modeling;

[0008] Call data fusion and decision module, use large model to carry out multi-source data fusion and risk analysis, output graded early warning and three-dimensional visual decision information;

[0009] Asynchronous processing and high-throughput transmission of data flow are performed via message middleware, realizing real-time of task scheduling and expansibility of system;

[0010] A plurality of distributed computing nodes are used to execute tasks issued by the master node, including target detection, feature extraction, and video stream processing; the plurality of distributed computing nodes are dynamically expanded or reduced to adapt to different computing requirements.

[0011] According to the method of the first aspect of the application, for the dynamic task analysis and path planning module:

[0012] An event trigger signal containing a timestamp t, geographic coordinates (lat, lng), and a target feature vector is received at the input interface thereof, and topological data of a digital map and a camera state matrix are synchronously accessed; in the topological data, nodes represent camera positions, and edges represent road connectivity; the camera state matrix data includes resolution, frame rate, and fault state;

[0013] Dynamic correlation graph modeling is performed, and a camera node weight function is defined:

[0014]

[0015] wherein, d i is the Euclidean distance between the target and the camera i, θ i is the angle between the camera optical axis and the target motion direction, s i ∈{0,0.5,1} represents the device state, 0 represents normal, 0.5 represents degradation, and 1 represents failure, and the coefficients α, β, and γ obtain optimal values through historical data training;

[0016] An optimal path is generated; wherein: an improved Dijkstra algorithm is used to preferentially select a path with the largest total weight, and a camera search sequence {Cam_1, Cam_2,..., Cam_n} is output, wherein Cam_n represents the nth camera; when it is detected that the target motion speed v is greater than a speed threshold v_threshold, a prediction mode is activated: an extended Kalman filter is used to predict the target position distribution in the future Δt time, and a candidate camera set C={Cam_k|P(k)>0.7} is generated; the analysis task is pre-assigned to the edge computing nodes corresponding to the cameras in the set C, and the target feature template is loaded in advance.

[0017] According to the method of the first aspect of the application, for the adaptive resource optimization deployment module:

[0018] Resource modeling is performed, and a task computing complexity quantification model is defined:

[0019] Q task =N obj *R res *(1+δ motion )*C_alg

[0020] wherein, N obj is the number of targets to be detected in the video, R res is the video resolution, δ motion is the motion blur compensation coefficient, and C_alg represents the algorithm type weight;

[0021] Configure a hybrid scheduling strategy; including:

[0022] Short-term elastic scheduling: build a deep Q network, the state space includes the CPU utilization of each node, GPU memory occupation, network bandwidth, the action space is the expansion and contraction operation of container instances, and the reward function is:

[0023] R = -0.7 * max (0, T latency -T SLA ) -0.2

[0024] *∑ (Utilization i ) + 0.1 * Energy_saving

[0025] wherein, T latency is the task delay, T SLA is the service level agreement threshold, energy_saving is the energy saving score, and Utilization i represents the resource utilization comprehensive score of the i-th computing node;

[0026] Utilization i = 0.4 * current CPU i usage rate +

[0027] 0.5 * current GPU i usage rate + 0.1 * current network delay

[0028] Long-term prediction scheduling: adopt Prophet time series model, input historical resource usage data, output predicted GPU memory demand curve, and pre-start standby computing nodes;

[0029] Perform energy consumption optimization, design a three-state node management strategy; including:

[0030] Active mode: the task queue length Q is not lower than the upper limit value Q_high;

[0031] Sleep mode: the task queue length Q is lower than the lower limit value Q_low, and the GPU acceleration unit is turned off;

[0032] Energy saving mode: continuous idle time T_idle> 300 seconds, switch to low power consumption state.

[0033] According to the method of the first aspect of the application, for the cross-lens coordination module:

[0034] Perform spatial alignment and feature enhancement: calculate the perspective transformation matrix H = K[R|t] based on the camera calibration parameters, map each camera coordinate system to the WGS84 geographic coordinate system, and the camera calibration parameters include the intrinsic matrix K, the extrinsic rotation matrix R, and the translation vector t;

[0035] Perform spatio-temporal correlation analysis and define a joint probability model for cross-camera target correlation:

[0036]

[0037] Where σ is the Sigmoid function, λ1, λ2, λ3 are empirical coefficients, Δd is the geographic distance difference, v max is the maximum moving speed of the target type; when occlusion occurs, start the memory-enhanced graph neural network MA-GNN, build the spatio-temporal graph G_t, where the nodes represent target instances and the edge weights are spatio-temporal correlation probabilities, and the node state update formula is:

[0038] h v(t+1) = GRU(h v(t) ,∑_{u∈N(v)}) W.ReLU([h v (t)||e_{uv}])

[0039] Where e_uv is the edge feature and W is a trainable parameter matrix.

[0040] According to the method of the first aspect of the application, for the data fusion and decision module:

[0041] Perform multi-modal data encoding using a hierarchical Transformer architecture:

[0042] The bottom layer is a spatio-temporal encoder that processes target trajectory point sequences using relative position encoding:

[0043]

[0044] The middle layer is an appearance feature fusion layer that aligns the 512-dimensional vector output by ViT-Track.

[0045] The top layer is a knowledge reasoning layer that constructs an entity-relation graph.

[0046] Perform threat assessment and visualization; where the risk propagation model is represented as:

[0047] Risk(u)=

[0048] α*Risk self(u) +β*∑_{v∈N(u)}Risk(v)*sim(u,v)

[0049] wherein sim(u,v) represents an entity type compatibility visualization interface, a three-dimensional geographic information system with superimposed real-time tracking trajectory, risk heat map, automatically generated event report, including key frame screenshot, space-time trajectory playback, associated target list.

[0050] The second aspect of the application proposes a distributed real-time target detection and tracking system based on dynamic perception and adaptive regulation, characterized in that the system comprises a master node, a message middleware and a plurality of distributed computing nodes; the master node is configured with a dynamic task analysis and path planning module, an adaptive resource optimization deployment module, a cross-lens collaboration module and a data fusion and decision module; wherein:

[0051] The dynamic task analysis and path planning module is configured to use the time, location, map information and camera network topology of the event to calculate the weight of each camera in real time, and generate an optimal search path by predicting the target motion trajectory, and decompose the global tracking task into multiple sub-tasks;

[0052] The adaptive resource optimization deployment module is configured to use a hybrid scheduling strategy to realize container elastic expansion and contraction and resource prediction according to the task complexity and node resource state, and perform resource scheduling;

[0053] The cross-lens collaboration module is configured to perform cross-camera video data alignment, target association and continuous tracking through multi-modal feature fusion and spatio-temporal joint modeling;

[0054] The data fusion and decision module is configured to use a large model to perform multi-source data fusion and risk analysis, and output hierarchical early warning and three-dimensional visualization decision information;

[0055] The message middleware is configured to perform asynchronous processing and high-throughput transmission of data flow, realize the real-time of task scheduling and the expansibility of the system;

[0056] The plurality of distributed computing nodes are configured to execute the tasks issued by the master node, including target detection, feature extraction and video stream processing; and dynamically expand or reduce the plurality of distributed computing nodes to adapt to different computing demands.

[0057] The third aspect of the application discloses an electronic device. The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the first aspect of the application, a distributed real-time target detection and tracking method based on dynamic perception and adaptive regulation.

[0058] The fourth aspect of the present application discloses a computer readable storage medium. The computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the first aspect of the present application, that is, a distributed real-time target detection and tracking method based on dynamic perception and adaptive regulation.

[0059] In summary, the present application proposes a four-stage linkage dynamic perception and adaptive regulation architecture. The core innovation is that through the spatio-temporal joint modeling, elastic resource scheduling, multi-modal feature fusion and closed-loop optimization mechanism of knowledge reasoning, the success rate of target tracking is improved to 98.5%, the resource utilization rate is stabilized at 85%±5%, and the ID switching rate is less than 2%. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the description of the specific embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0061] Figure 1 FIG. 1 is a schematic diagram of a distributed real-time target detection and tracking system based on dynamic perception and adaptive regulation according to an embodiment of the present application.

[0062] Figure 2 FIG. 3 is a schematic diagram of a dynamic task analysis and path planning module according to an embodiment of the present application.

[0063] Figure 3 FIG. 5 is a schematic diagram of an adaptive resource optimization and deployment module according to an embodiment of the present application.

[0064] Figure 4 FIG. 7 is a schematic diagram of a cross-lens collaborative tracking module according to an embodiment of the present application.

[0065] Figure 5 FIG. 9 is a schematic diagram of a large model data fusion and decision module according to an embodiment of the present application. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0067] The core idea of the present application is to build a multi-level intelligent monitoring system to realize the closed-loop process of "dynamic perception → elastic scheduling → collaborative analysis → intelligent decision-making". Specifically, first, the weight of each camera is calculated in real time by using the time, location, map information and camera network topology of the event, and the optimal search path is generated by predicting the target motion trajectory, so as to decompose the global tracking task into multiple sub-tasks; then, according to the task complexity and node resource state, a hybrid scheduling strategy is adopted to realize short-term container elastic expansion and contraction and long-term resource prediction, so as to ensure efficient and stable resource scheduling; then, through multi-modal feature fusion and spatio-temporal joint modeling, video data alignment, target association and continuous tracking across cameras are realized; finally, multi-source data fusion and risk reasoning are carried out by using a large model, and hierarchical early warning and three-dimensional visual decision information are output. The present application effectively overcomes the shortcomings of traditional systems in dynamic tracking, resource allocation, cross-camera cooperation and early warning decision-making, realizes the improvement of tracking accuracy, optimization of resource utilization and real-time response.

[0068] As shown in Figure 1 in the distributed real-time target detection and tracking system based on dynamic perception and adaptive regulation:

[0069] 1. Master node (central control unit), responsible for task scheduling and global management, including four key modules:

[0070] (1) Dynamic task analysis and path planning module: according to the target detection result, map information and camera distribution, the optimal search path and task decomposition strategy are calculated.

[0071] (2) Adaptive resource optimization deployment module: according to the calculation load (CPU / GPU / bandwidth, etc.), the task allocation is dynamically adjusted, and the resource utilization rate is optimized.

[0072] (3) Cross-camera cooperation module: realizes target matching, tracking and data sharing among multiple cameras, and ensures continuous tracking of the target.

[0073] (4) Data fusion and decision-making module: analyzes and fuses multi-source data to improve the intelligent decision-making ability of the system.

[0074] 2. Message middleware (Apache Kafka), as the core communication hub, realizes asynchronous processing and high-throughput transmission of data flow, guarantees the real-time performance of task scheduling and the scalability of the system.

[0075] 3. Distributed computing nodes (node 1 to node n), responsible for specific analysis tasks such as target detection, feature extraction, video stream processing, etc. These nodes can be dynamically expanded or reduced to adapt to different computing demands, improving the scalability and stability of the system.

[0076] The above system is used to realize:

[0077] (1) Distributed collaborative computing: efficient task distribution through Kafka, improving system parallel computing capability.

[0078] (2) Dynamic task allocation: master node intelligently analyzes task priority and reasonably schedules computing resources.

[0079] (3) Efficient target tracking: through cross-lens collaborative module, realize continuous tracking of multiple cameras, reduce loss rate.

[0080] (4) Adaptive resource management: combining load balancing and resource prediction, optimizing the utilization of computing resources.

[0081] The above system realizes efficient, scalable real-time target detection and tracking system through the distributed mode of master node-message middleware-computing node, combined with dynamic task allocation, resource optimization, cross-lens tracking and data fusion.

[0082] First embodiment

[0083] Take the train station scene as an example to illustrate the system running process:

[0084] (1) Initial event: the east side camera on the platform detects a left-over package (time t0, coordinate p0).

[0085] (2) Dynamic task analysis module: calculate the weight of cameras within 200 meters, generate search sequence [C1 (facing exit), C2 (waiting hall entrance), C3 (ticket office)] to predict the possible direction of suspicious person to exit, activate C1 and C4 (security check) analysis tasks in advance.

[0086] (3) Resource optimization module: detect that the GPU node load reaches 85%, trigger DQN scheduling strategy, add 2 GPU container instances, migrate C3 analysis task to idle CPU node

[0087] (4) Cross-lens collaborative module: C1 finds a matching target (similarity 0.92), but the target enters the blind area MA-GNN correlates the suspicious target captured by C4 according to the motion trend, ID switching rate 1.3%.

[0088] (5) Large model decision module: analyze the target trajectory to form a gathering pattern (linger in the security check for more than 3 minutes), generate a red warning, and the visual interface shows that the risk hot area spreads to the west exit.

[0089] For dynamic task analysis and path planning module, as shown in Figure 2 :

[0090] Input interface: receive event trigger signal containing timestamp t, geographic coordinates (lat, lng), target feature vector, synchronously access topological data of digital map (nodes represent camera positions, edges represent road connectivity) and camera state matrix (resolution, frame rate, fault status).

[0091] Dynamic association graph modeling: define camera node weight function,

[0092]

[0093] where d i is the Euclidean distance between the target and camera i, θ i is the angle between the camera optical axis and the target motion direction, s i ∈{0,0.5,1} represents the device state (normal / degraded / fault). Coefficients α, β, γ are obtained by training historical data to obtain optimal values (typical values α=0.6, β=0.3, γ=0.1).

[0094] Optimal path generation: use improved Dijkstra algorithm, preferentially select path with maximum weight sum, output camera search sequence {Cam_1, Cam_2,..., Cam_n}. When detecting target motion speed v>v_threshold, activate prediction mode: use extended Kalman filter to predict target position distribution in future Δt time, generate candidate camera set C={Cam_k|P(k)>0.7}. Pre-allocate analysis tasks to edge computing nodes corresponding to cameras in C, and load target feature template in advance.

[0095] For adaptive resource optimization deployment module, as shown in Figure 3 :

[0096] Resource modeling: define task computing complexity quantification model as follows,

[0097] Q task = N obj *R res *(1+δ motion )*C_alg

[0098] where N obj is the number of targets to be detected in the video, R res is the video resolution (e.g. 1920x1080), δ motion is the motion blur compensation coefficient (0.1-0.3), and C_alg represents the algorithm type weight (YOLOv8:1.2, DeepSORT:1.5).

[0099] Hybrid scheduling strategy:

[0100] Short-term elastic scheduling: build a deep Q network (DQN), the state space includes the CPU utilization of each node, GPU memory occupation, network bandwidth, and the action space is the container instance scaling operation (+1 / -1 / 0). The reward function is designed as:

[0101] R = -0.7 * max (0, T latency -T SLA )-0.2

[0102] *∑(Utilization i )+0.1*Energy_saving

[0103] where T latency is the task delay, T SLA is the service level agreement threshold (such as 200ms), and energy_saving is the energy saving score. Utilization i represents the resource utilization comprehensive score of the i-th computing node, as follows:

[0104] Utilization i = 0.4 * current CPU i usage rate + 0.5 * current GPU i usage rate + 0.1 * current network delay

[0105] Long-term prediction scheduling: adopt Prophet time series model, input historical resource usage data (sampling interval 5 minutes), output future 1 hour GPU memory demand curve, pre-start standby computing node.

[0106] Energy optimization: design a three-state node management strategy, as follows:

[0107] Active mode: task queue length Q ≥ Q_high (such as 10 tasks).

[0108] Sleep mode: turn off GPU acceleration unit when Q < Q_low (such as 2 tasks).

[0109] Energy saving mode: switch to low power state when T_idle > 300 seconds.

[0110] For cross-lens cooperative tracking module, as shown in Figure 4 :

[0111] Spatial alignment and feature enhancement: based on camera calibration parameters (intrinsic matrix K, extrinsic rotation matrix R and translation vector t), calculate perspective transformation matrix H = K[R|t], map each camera coordinate system to WGS84 geographic coordinate system

[0112] Space-time correlation reasoning: define joint probability model for cross-camera target correlation as follows

[0113]

[0114] where σ is the Sigmoid function, λ1=0.6, λ2=0.3, λ3=0.1 are empirical coefficients, Δd is the geographical distance difference, v_max is the maximum moving speed of target type (pedestrian 5m / s, vehicle 30m / s). When occlusion occurs, start memory augmented graph neural network (MA-GNN), build space-time graph G_t, node represents target instance, edge weight is space-time correlation probability node state update formula:

[0115] h v(t+1) = GRU(h v(t) ,∑_{u∈N(v)})W.ReLU([h v (t)||e_{uv}])

[0116] where e_uv is the edge feature (time difference, spatial distance), W is the trainable parameter matrix.

[0117] For large model data fusion and decision module, as shown in Figure 5

[0118] Multi-modal data encoding: use hierarchical Transformer architecture, as follows:

[0119] Bottom layer, space-time encoder, process target trajectory point sequence {(x_i, y_i, t_i)}, use relative position encoding:

[0120]

[0121] Middle layer: appearance feature fusion layer, align 512-dimensional vector output by ViT-Track.

[0122] Top layer: knowledge reasoning layer, build entity-relation graph (such as "Person A-Enter-Region B")

[0123] Threat assessment and visualization: risk propagation model, define threat value propagation formula:

[0124] Risk(u)=α*Risk self(u) +β*∑_{v∈N(u)}Risk(v)*sim(u,v)

[0125] where sim(u,v) is based on entity type compatibility (person-person: 0.8, person-car: 0.3.

[0126] ​Visualization interface, overlaying real-time tracking trajectory (color-coded ID), risk heat map (red-high risk, yellow-warning) in 3D GIS, automatically generating event reports, including key frame screenshots, spatiotemporal trajectory playback, and associated target list.

[0127] In some embodiments, for a multi-level linkage dynamic perception system architecture: a closed-loop architecture of dynamic task planning → elastic resource scheduling → cross-lens collaboration → large model decision-making is created, data / control flow decoupling between modules is achieved through message middleware, and support for thousand-node expansion is provided.

[0128] Each module realizes asynchronous communication through Kafka / RabbitMQ to ensure system stability under high concurrency, and the processing flow includes spatiotemporal joint modeling, multi-modal feature fusion, and progressive optimization of knowledge reasoning.

[0129] In some embodiments, for dynamic resource scheduling and energy efficiency optimization mechanism: a hybrid scheduling strategy of online reinforcement learning + offline time series prediction is proposed, resource utilization rate fluctuation is reduced from 25% to 5%, node three-state management (active / sleep / power saving) is realized with energy consumption perception, and energy efficiency ratio is improved by 27.3%.

[0130] The DQN reward function integrates delay penalty, load balancing, and energy saving factors (weights 0.7:0.2:0.1), the Prophet model predicts GPU memory requirements, the error rate is <8%, and the dynamic neural network slicing technology (DySlice) automatically adjusts the model depth according to GPU memory.

[0131] In some embodiments, for cross-lens collaborative tracking technology: a ViT-Track network is designed to integrate RGB and optical flow features, ID switching rate is reduced from 15% to 2%, a memory-enhanced GNN is proposed to solve the occlusion problem, and target loss rate is reduced by 10 times.

[0132] Through spatial alignment, a perspective transformation matrix H = K[R|t] is constructed based on calibration parameters, and a joint probability type is used:

[0133]

[0134] MA-GNN node state update formula:

[0135] h v(t+1) = GRU(h v(t) ,∑_{u∈N(v)})W.ReLU([h v (t)||e_{uv}])

[0136] In some embodiments, for large model driven intelligent decision and risk early warning: the false positive rate is reduced from 25% to 6.7%, the threat diffusion model combines the random walk algorithm, and the early warning response speed is improved by 40%.

[0137] By static (type / color) + dynamic (speed / acceleration), the relationship edge weight update formula is determined: Risk(u)=α*Risk self(u) +β*∑_{v∈N(u)}Risk(v)*sim(u,v). Visual interface, three-dimensional GIS superimposed real-time trajectory and risk hot area (RGB color temperature coding).

[0138] In summary, the application proposes a four-stage linkage dynamic perception and adaptive regulation architecture, the core innovation of which is to realize the target tracking success rate of 98.5%, the resource utilization rate of 85%±5%, and the ID switching rate of less than 2% through the closed-loop optimization mechanism of spatio-temporal joint modeling, elastic resource scheduling, multi-modal feature fusion and knowledge reasoning.

[0139] Please note that the technical features of the above embodiments can be combined arbitrarily, and in order to make the description simple, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the description. The above embodiments only express several implementation manners of the application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the application, some modifications and improvements can be made, which are all within the protection scope of the application. Therefore, the protection scope of the patent of the application should be subject to the appended claims.

Claims

1. A distributed real-time target detection and tracking method based on dynamic perception and adaptive control, characterized in that: In the method, a master node, a message middleware, and multiple distributed computing nodes are configured, and the master node is further configured with a dynamic task analysis and path planning module, an adaptive resource optimization and deployment module, a cross-lens collaboration module, and a data fusion and decision module; The method specifically includes: The dynamic task analysis and path planning module is called to use the time, location, map information, and camera network topology of the event to calculate the weight of each camera in real time. It also generates the optimal search path by predicting the target's motion trajectory, and decomposes the global tracking task into multiple subtasks. Call the adaptive resource optimization deployment module, adopt a hybrid scheduling strategy to achieve container elastic scaling and resource prediction based on task complexity and node resource status, and perform resource scheduling; The cross-camera collaboration module is called to perform cross-camera video data alignment, target association, and continuous tracking through multimodal feature fusion and spatiotemporal joint modeling. Call the data fusion and decision-making module, use the large model to perform multi-source data fusion and risk analysis, and output graded warning and three-dimensional visual decision information; Asynchronous processing and high-throughput transmission of data streams are performed through message middleware, achieving real-time task scheduling and system scalability; Utilize multiple distributed computing nodes to execute tasks assigned by the master node, including target detection, feature extraction, and video stream processing; dynamically expand or reduce multiple distributed computing nodes to adapt to different computing needs.

2. A distributed real-time target detection and tracking method based on dynamic perception and adaptive control according to claim 1, characterized in that: For the dynamic task analysis and path planning module: The input interface receives event trigger signals containing timestamp t, geographic coordinates (lat, lng), and target feature vectors, and synchronously accesses the topological data of the digital map and the camera status matrix; in the topological data, nodes represent camera locations and edges represent road connectivity; Camera status matrix data includes resolution, frame rate, and fault status; Perform dynamic association graph modeling and define the camera node weight function: Among them, d i is the Euclidean distance between the target and camera i, θ i is the angle between the camera optical axis and the target motion direction, s i ∈{0,0.5,1} represents the device status, 0 represents normal, 0.5 represents degradation, and 1 represents failure. The coefficients α, β, and γ are trained with historical data to obtain their optimal values. Generate the optimal path; wherein: use the improved Dijkstra algorithm, give priority to the path with the largest total weight, output the camera search sequence {Cam_1,Cam_2,...,Cam_n}, Cam_n represents the nth camera; when it is detected that the target motion speed v is greater than the speed threshold v_threshold, activate the prediction mode: use the extended Kalman filter to predict the target position distribution within the next Δt time, and generate a candidate camera set C = {Cam_k|P(k)>0.7}; pre-assign analysis tasks to the edge computing nodes corresponding to the cameras in set C, and load the target feature template in advance.

3. The distributed real-time target detection and tracking method based on dynamic perception and adaptive control according to claim 2 is characterized in that: For the adaptive resource optimization deployment module: Execute resource modeling and define a quantitative model for task computational complexity: Q task =N obj *R res *(1+d motion )*C_alg Among them, N obj is the number of targets to be detected in the video, R res is the video resolution, δm otion is the motion blur compensation coefficient, C_alg represents the algorithm type weight; Configure hybrid scheduling policies; including: Short-term elastic scheduling: Build a deep Q network. The state space includes the CPU utilization, GPU memory usage, and network bandwidth of each node. The action space is the scaling of container instances. The reward function is: R=-0.7*max(0,T latency -T SLA )-0.2*∑(Utilization i )+0.1*Energy_saving Among them, T latency is the task delay, T SLA is the service level agreement threshold, energy_saving is the energy saving score, and Utilization i represents the comprehensive score of resource utilization of the i-th computing node; Utilization i =0.4*Current CPU i Utilization Rate+ 0.5*current GPU i Usage rate + 0.1 * current network delay Long-term predictive scheduling: Using the Prophet time series model, historical resource usage data is input, the predicted GPU memory demand curve is output, and backup computing nodes are pre-started; Perform energy consumption optimization and design three-state node management strategies; including: Active mode: The task queue length Q is not less than the upper limit Q_high; Sleep mode: The task queue length Q is lower than the lower limit Q_low, and the GPU acceleration unit is turned off; Energy-saving mode: If the idle time T_idle is greater than 300 seconds, the system switches to low-power mode.

4. The distributed real-time target detection and tracking method based on dynamic perception and adaptive control according to claim 3 is characterized in that: For the cross-lens collaboration module: Perform spatial alignment and feature enhancement: Calculate the perspective transformation matrix H = K[R|t] based on the camera calibration parameters, and map each camera coordinate system to the WGS84 geographic coordinate system. The camera calibration parameters include the intrinsic parameter matrix K, the extrinsic parameter rotation matrix R, and the translation vector t. Perform spatiotemporal correlation analysis and define a joint probability model for cross-camera object correlation: Among them, σ is the Sigmoid function, λ1, λ2, λ3 are empirical coefficients, Δd is the geographical distance difference, v max is the maximum moving speed of the target type; when occlusion occurs, the memory-augmented graph neural network MA-GNN is started to construct a spatiotemporal graph G_t, where the nodes represent target instances, the edge weights are the spatiotemporal association probabilities, and the node state update formula is: h v(t+1) =GRU(h v(t) ,∑_{u∈N(v)})W.ReLU([h v (t)||e_{uv}]) Among them, e_uv is the edge feature and W is the trainable parameter matrix.

5. The distributed real-time target detection and tracking method based on dynamic perception and adaptive control according to claim 4 is characterized in that: For the data fusion and decision module: Perform multimodal data encoding using a layered Transformer architecture: The bottom layer is a spatiotemporal encoder that processes target trajectory point sequences and uses relative position encoding: The middle layer is the appearance feature fusion layer, which aligns the 512-dimensional vector output by ViT-Track. The top layer is the knowledge reasoning layer, which builds the entity-relationship graph; Perform threat assessment and visualization; the risk propagation model is characterized as follows: Risk(s)=α*Risk self (u)+β*∑_{v∈N(u)}Risk(v)*sim(u,v) Among them, sim(u,v) represents the entity type compatibility visualization interface, real-time tracking trajectories and risk heat maps are superimposed in the three-dimensional geographic information system, and event reports are automatically generated, including key frame screenshots, spatiotemporal trajectory playback, and associated target lists.

6. A distributed real-time target detection and tracking system based on dynamic perception and adaptive control, characterized in that: The system includes: a master node, a message middleware, and multiple distributed computing nodes; the master node is equipped with a dynamic task analysis and path planning module, an adaptive resource optimization and deployment module, a cross-lens collaboration module, and a data fusion and decision module; wherein: The dynamic task analysis and path planning module is configured to use the time, location, map information, and camera network topology of the event to calculate the weight of each camera in real time, and generate the optimal search path by predicting the target motion trajectory, thereby decomposing the global tracking task into multiple subtasks. The adaptive resource optimization and deployment module is configured to: Based on task complexity and node resource status, adopt a hybrid scheduling strategy to achieve container elastic scaling and resource prediction, and perform resource scheduling; The cross-camera collaboration module is configured to: perform cross-camera video data alignment, target association, and continuous tracking through multimodal feature fusion and spatiotemporal joint modeling; The data fusion and decision-making module is configured to: utilize large models to perform multi-source data fusion and risk analysis, and output graded warnings and three-dimensional visual decision-making information; The message middleware is configured to: perform asynchronous processing and high-throughput transmission of data streams, achieve real-time task scheduling and system scalability; Multiple distributed computing nodes are configured to: execute tasks assigned by the master node, including target detection, feature extraction, and video stream processing; dynamically expand or reduce multiple distributed computing nodes to adapt to different computing needs.

7. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it implements the distributed real-time target detection and tracking method based on dynamic perception and adaptive control as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements the distributed real-time target detection and tracking method based on dynamic perception and adaptive control as described in any one of claims 1 to 5.

Images