Target dynamic detection and tracking system and method
By improving the YOLO network model and optimizing edge computing, and combining dynamic convolution and Kalman filtering mechanisms, the problems of detection accuracy and real-time performance in dynamic scenes are solved, achieving high-precision detection and stable trajectory prediction, which is suitable for security and intelligent transportation scenarios.
Patent Information
- Application Number
- CN202511417943.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-30
- Publication Date
- 2026-01-13
AI Technical Summary
Existing target detection methods lack accuracy and real-time performance in dynamic scenarios and lack trajectory prediction capabilities, failing to meet the needs of continuous monitoring and risk warning for complex moving targets.
An improved YOLO network model is adopted, which combines dynamic convolutional structure, temporal sequence analysis and Kalman filtering mechanism. High-precision detection and stable trajectory prediction are achieved through edge computing optimization, and early warning information is output.
It improved detection accuracy by about 5%, reduced trajectory prediction error by about 20%, and reduced inference latency by about 35%, enabling real-time processing in complex dynamic environments.
Smart Images
Figure CN121330263A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and artificial intelligence, specifically to a target dynamic detection and tracking system and method. Background Technology
[0002] Existing target detection technologies are commonly used in security monitoring, traffic analysis, and industrial inspection, among which real-time detection algorithms based on the YOLO (You Only Look Once) series are widely adopted due to their end-to-end detection capabilities. However, traditional YOLO suffers from problems such as decreased detection accuracy, missed detections, or false detections when dealing with scenarios involving rapidly changing target shapes, complex motion states, and significant occlusion interference. Furthermore, some methods lack the ability to predict and analyze the dynamic trajectory of targets, failing to meet the needs of continuous monitoring and risk warning for moving targets.
[0003] Existing target detection methods suffer from insufficient accuracy and real-time performance in dynamic scenarios, as well as a lack of trajectory prediction, which are technical problems that need to be addressed. Summary of the Invention
[0004] The technical objective of this invention is to address the above-mentioned shortcomings by providing a target dynamic detection and tracking system and method to solve the technical problems of insufficient accuracy and real-time performance of existing target detection methods in dynamic scenarios, as well as the lack of trajectory prediction.
[0005] In a first aspect, the present invention provides a target dynamic detection and tracking system, comprising a target detection module, a time sequence analysis module, an edge computing optimization module, and a trajectory prediction and early warning module;
[0006] The target detection module is equipped with an improved YOLO network model that incorporates a dynamic convolutional structure. This dynamic convolutional structure is embedded in the feature extraction layer of the YOLO network model and can adaptively generate weight coefficients based on the features of the input video frame. It can also perform weighted combination of multiple basic convolutional kernels to achieve high-precision detection of rapidly changing and complex-shaped targets and output target state features.
[0007] The time series analysis module is used to perform state modeling and motion trend prediction based on the target state characteristics, and output the target state sequence;
[0008] The trajectory prediction and early warning module is used to predict the trajectory based on the target state sequence, combined with optical flow features and Kalman filter update mechanism, and output early warning information when an abnormal trajectory pattern is detected.
[0009] The edge computing optimization module is used to perform pruning and quantization on the improved YOLO network model, and to perform joint optimization in conjunction with the computational requirements of trajectory prediction, so as to improve the real-time inference efficiency of edge devices.
[0010] As a preferred approach, dynamic convolutional structures generate weight coefficients through an attention mechanism, and then combine multiple basic convolutional kernels in a weighted manner to improve detection robustness in scenarios where the target shape changes rapidly.
[0011] Preferably, the state update of the time series analysis module satisfies the following relationship:
[0012] X t =X t-1 +V t-1 ·Δt,
[0013] Among them, X t Let X be the target state at time t. t-1 V represents the target state from the previous moment. t-1 Let Δt be the instantaneous velocity of the target, and Δt be the time interval.
[0014] Preferably, the trajectory prediction and the state prediction of the early warning module satisfy the Kalman filter update formula:
[0015] X t =A·X t-1 +B·U t +W t ,
[0016] P t =A·P t-1 ·A T +Q,
[0017] Among them, X t Let A be the state transition matrix and B be the control matrix, representing the predicted state. U t To control the input, P t Let W be the covariance matrix, Q be the process noise, and W be the process noise. t This is the process noise term.
[0018] Secondly, the present invention provides a target dynamic detection and tracking method, which realizes target dynamic detection and tracking through a target dynamic detection and tracking system as described in any one of the first aspects, comprising the following steps:
[0019] S100 Edge Computing Optimization: The improved YOLO network model is pruned and quantized, and jointly optimized based on the computational dependency characteristics of object detection and trajectory prediction to achieve efficient inference on edge devices.
[0020] S200 Target Detection: Based on the dynamic convolutional structure in the improved YOLO network model, it performs feature extraction and target detection on the input video frame and outputs target state features;
[0021] S300 Time Series Analysis: Based on the target state characteristics, state modeling and motion trend prediction are performed, and the target state sequence is output.
[0022] S400 Trajectory Prediction and Early Warning: Combining optical flow characteristics and Kalman filtering mechanism, the trajectory of the target state sequence is predicted, and early warning information is output when abnormal motion patterns are detected.
[0023] Preferably, the dynamic convolutional structure is introduced into the feature extraction layer of YOLO, which improves the detection accuracy in occlusion and fast motion scenarios by adaptively adjusting the convolutional kernel parameters.
[0024] Preferably, when modeling the target state features in consecutive video frames, the state update satisfies X. t =X t-1 +V t-1 The relationship between Δt and Δt.
[0025] Preferably, in the trajectory prediction and early warning steps, the state prediction satisfies the Kalman filter update formula:
[0026] X t =A·X t-1 +B·U t +W t ,
[0027] P t =A·P t-1 ·A T +Q,
[0028] Among them, X t Let A be the state transition matrix and B be the control matrix, representing the predicted state. U t To control the input, P t Let W be the covariance matrix, Q be the process noise, and W be the process noise. t This is the process noise term.
[0029] The target dynamic detection and tracking method and system of the present invention have the following advantages:
[0030] 1. High-precision detection: The dynamic convolutional structure enhances YOLO's detection capabilities in complex environments, and performs better on rapidly changing and partially occluded targets;
[0031] 2. Stable trajectory prediction: The Kalman filtering mechanism, which combines optical flow characteristics, makes trajectory prediction more stable in nonlinear motion scenarios;
[0032] 3. Real-time early warning: The system can output risk warnings in real time when abnormal trajectories are detected, which is suitable for security-critical scenarios such as security and intelligent transportation;
[0033] 4. Edge optimization: A joint optimization strategy of pruning and quantization is proposed, which not only reduces the number of model parameters, but also improves the inference efficiency in the trajectory prediction process.
[0034] 5. Experimental verification: Experiments in complex dynamic environments show that the detection accuracy of this method is improved by about 5% compared with the traditional YOLOv5, the trajectory prediction error is reduced by about 20%, the inference latency is reduced by about 35%, and it can achieve real-time processing of 4K video at 30 frames / second. Attached Figure Description
[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention, 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0036] The invention will be further described below with reference to the accompanying drawings.
[0037] Figure 1 This is a block diagram of a target dynamic detection and tracking system according to Example 1;
[0038] Figure 2 This is a flowchart of a target dynamic detection and tracking method in Example 2. Detailed Implementation
[0039] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments are not intended to limit the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.
[0040] This invention provides a target dynamic detection and tracking system and method to solve the technical problems of insufficient accuracy and real-time performance of existing target detection methods in dynamic scenarios and the lack of trajectory prediction.
[0041] Example 1:
[0042] The present invention provides a target dynamic detection and tracking system, comprising a target detection module, a time sequence analysis module, an edge computing optimization module, and a trajectory prediction and early warning module.
[0043] The target detection module is equipped with an improved YOLO network model that incorporates a dynamic convolutional structure. This dynamic convolutional structure is embedded in the feature extraction layer of the YOLO network model and can adaptively generate weight coefficients based on the features of the input video frames. It can also perform weighted combination of multiple basic convolutional kernels to achieve high-precision detection of rapidly changing and complex-shaped targets and output target state features.
[0044] In practice, dynamic convolutional structures generate weight coefficients through an attention mechanism and combine multiple basic convolutional kernels in a weighted manner, thereby improving detection robustness in scenarios where the target shape changes rapidly.
[0045] The time series analysis module is used to perform state modeling and motion trend prediction based on the target state characteristics, and output the target state sequence.
[0046] As a specific implementation of the time series analysis module, the state update of the time series analysis module satisfies the following relationship:
[0047] X t =X t-1 +V t-1 ·Δt,
[0048] Among them, X t Let X be the target state at time t. t-1 V represents the target state from the previous moment. t-1 Let Δt be the instantaneous velocity of the target, and Δt be the time interval.
[0049] The trajectory prediction and early warning module is used to predict the trajectory based on the target state sequence, combined with optical flow features and Kalman filter update mechanism, and output early warning information when abnormal trajectory patterns are detected.
[0050] As a specific implementation of the trajectory prediction and early warning module, the state prediction satisfies the Kalman filter update formula:
[0051] X t =A·X t-1 +B·U t +W t ,
[0052] P t =A·P t-1 ·A T +Q,
[0053] Among them, X t Let A be the state transition matrix and B be the control matrix, representing the predicted state. U t To control the input, P t Let W be the covariance matrix, Q be the process noise, and W be the process noise. t This is the process noise term.
[0054] The edge computing optimization module is used to perform pruning and quantization on the improved YOLO network model, and to perform joint optimization in conjunction with the computational requirements of trajectory prediction, so as to improve the real-time inference efficiency of edge devices.
[0055] This module compresses the network structure through methods such as channel pruning and model quantization, reducing computational resource consumption and ensuring real-time inference of high-resolution video on edge devices.
[0056] This embodiment employs an improved YOLO network, integrating a dynamic convolutional structure and a time-series analysis mechanism, optimized for multi-target detection and dynamic tracking tasks in complex scenarios. Dynamic convolution automatically adjusts the convolutional kernel based on input features to adapt to changes in target shape. Combined with time-series modeling, it achieves continuous tracking of target states and prediction of motion trajectories, improving detection accuracy and real-time performance. The device supports real-time processing of high-resolution video and can be widely applied in scenarios such as security monitoring, intelligent transportation, and drone inspection.
[0057] Example 2:
[0058] This invention provides a target dynamic detection and tracking method, comprising four steps: edge computing optimization, target detection, time series analysis, and trajectory prediction and early warning.
[0059] S100 Edge Computing Optimization: The improved YOLO network model is pruned and quantized, and jointly optimized based on the computational dependency characteristics of object detection and trajectory prediction to achieve efficient inference on edge devices.
[0060] S200 Object Detection: Based on the dynamic convolutional structure in the improved YOLO network model, it performs feature extraction and object detection on the input video frames and outputs object state features.
[0061] In this step, a dynamic convolutional structure is introduced into the feature extraction layer of YOLO. By adaptively adjusting the convolutional kernel parameters, the detection accuracy in occlusion and fast motion scenarios is improved.
[0062] S300 time series analysis: Based on the target state characteristics, state modeling and motion trend prediction are performed, and the target state sequence is output.
[0063] When modeling the target state features in consecutive video frames, the state update satisfies X. t =X t-1 +V t-1 The relationship between Δt and Δt.
[0064] S400 Trajectory Prediction and Early Warning: Combining optical flow characteristics and Kalman filtering mechanism, the trajectory of the target state sequence is predicted, and early warning information is output when abnormal motion patterns are detected.
[0065] In the trajectory prediction and early warning steps of this embodiment, the state prediction satisfies the Kalman filter update formula:
[0066] X t =A·Xt-1 +B·U t +W t ,
[0067] P t =A·P t-1 ·A T +Q,
[0068] Among them, X t Let A be the state transition matrix and B be the control matrix, representing the predicted state. U t To control the input, P t Let W be the covariance matrix, Q be the process noise, and W be the process noise. t This is the process noise term.
[0069] The method in this embodiment is based on the system disclosed in Embodiment 1.
[0070] The target dynamic detection and tracking system and method provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A target dynamic detection and tracking system, characterized in that, It includes a target detection module, a time series analysis module, an edge computing optimization module, and a trajectory prediction and early warning module; The target detection module is equipped with an improved YOLO network model that incorporates a dynamic convolutional structure. This dynamic convolutional structure is embedded in the feature extraction layer of the YOLO network model and can adaptively generate weight coefficients based on the features of the input video frame. It can also perform weighted combination of multiple basic convolutional kernels to achieve high-precision detection of rapidly changing and complex-shaped targets and output target state features. The time series analysis module is used to perform state modeling and motion trend prediction based on the target state characteristics, and output the target state sequence; The trajectory prediction and early warning module is used to predict the trajectory based on the target state sequence, combined with optical flow features and Kalman filter update mechanism, and output early warning information when an abnormal trajectory pattern is detected. The edge computing optimization module is used to perform pruning and quantization on the improved YOLO network model, and to perform joint optimization in conjunction with the computational requirements of trajectory prediction, so as to improve the real-time inference efficiency of edge devices.
2. The target dynamic detection and tracking system according to claim 1, characterized in that, Dynamic convolutional structures generate weight coefficients through an attention mechanism, and combine multiple basic convolutional kernels in a weighted manner, thereby improving the detection robustness in scenarios where the target shape changes rapidly.
3. The target dynamic detection and tracking system according to claim 1, characterized in that, The state update of the time series analysis module satisfies the following relationship: X t =X t-1 +V t-1 ·Δt, Among them, X t Let X be the target state at time t. t-1 V represents the target state from the previous moment. t-1 Let Δt be the instantaneous velocity of the target, and Δt be the time interval.
4. The target dynamic detection and tracking system according to claim 1, characterized in that, The state prediction of the trajectory prediction and early warning module satisfies the Kalman filter update formula: X t =A·X t-1 +B·U t +W t , P t =A·P t-1 ·A T +Q, Among them, X t Let A be the state transition matrix and B be the control matrix, representing the predicted state. U t To control the input, P t Let W be the covariance matrix, Q be the process noise, and W be the process noise. t This is the process noise term.
5. A method for dynamic target detection and tracking, characterized in that, The target dynamic detection and tracking system as described in any one of claims 1-4 achieves target dynamic detection and tracking, including the following steps: S100 Edge Computing Optimization: The improved YOLO network model is pruned and quantized, and jointly optimized based on the computational dependency characteristics of object detection and trajectory prediction to achieve efficient inference on edge devices. S200 Target Detection: Based on the dynamic convolutional structure in the improved YOLO network model, it performs feature extraction and target detection on the input video frame and outputs target state features; S300 Time Series Analysis: Based on the target state characteristics, state modeling and motion trend prediction are performed, and the target state sequence is output. S400 Trajectory Prediction and Early Warning: Combining optical flow characteristics and Kalman filtering mechanism, the trajectory of the target state sequence is predicted, and early warning information is output when abnormal motion patterns are detected.
6. The target dynamic detection and tracking method according to claim 5, characterized in that, The dynamic convolutional structure is introduced into the feature extraction layer of YOLO, which improves the detection accuracy in occlusion and fast motion scenarios by adaptively adjusting the convolutional kernel parameters.
7. The target dynamic detection and tracking method according to claim 5, characterized in that, When modeling the target state features in consecutive video frames, the state update satisfies X. t =X t-1 +V t-1 The relationship between Δt and Δt.
8. The target dynamic detection and tracking method according to claim 5, characterized in that, In the trajectory prediction and early warning steps, the state prediction satisfies the Kalman filter update formula: X t =A·X t-1 +B·U t +W t , P t =A·P t-1 ·A T +Q, Among them, X t Let A be the state transition matrix and B be the control matrix, representing the predicted state. U t To control the input, P t Let W be the covariance matrix, Q be the process noise, and W be the process noise. t This is the process noise term.