Target tracking method and device based on ring network fusion
By combining ordinary differential equation networks and ring networks, the real-time performance and accuracy bottlenecks of UAV target tracking in complex dynamic scenarios in existing technologies have been solved, achieving efficient and accurate target tracking results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2026-04-02
- Publication Date
- 2026-07-14
AI Technical Summary
In complex and dynamic scenarios, existing signal tracking and visual obstacle avoidance methods struggle to achieve efficient and accurate target tracking during high-speed movement of UAVs. Especially under resource-constrained conditions, the information fusion efficiency is insufficient, leading to path planning delays and inaccurate target position predictions.
We employ an ordinary differential equation-based network to extract features and sparsify visual images and target tracking signals. Combined with ring network information fusion, we obtain target fusion features and determine the tracking strategy through multiple iterations.
It significantly improves the speed and success rate of target tracking, providing an efficient and accurate technical solution for intelligent target tracking of drones in high-speed motion scenarios.
Smart Images
Figure CN122388901A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target tracking technology, and in particular to a target tracking method and apparatus based on ring network fusion. Background Technology
[0002] In complex forest environments, it is necessary to combine the positioning capabilities of remote target signals (such as very high frequency (VHF) signals transmitted back by the target) or visual obstacle avoidance technology to achieve accurate target tracking.
[0003] For existing signal tracking and visual obstacle avoidance methods, in complex environments such as forests, the computational workload for signal localization increases significantly due to multipath effects and signal attenuation, while also requiring real-time processing of large amounts of data from visual sensors. This leads to path planning delays and slow information processing under resource-constrained conditions (such as embedded computing platforms or low-power devices). To ensure the real-time performance and reliability of the overall system, the tracking speed needs to be reduced, making it difficult to achieve efficient target tracking.
[0004] Existing technologies typically employ multi-method integration techniques that combine signal tracking with visual obstacle avoidance to achieve target tracking. However, in scenarios involving high-speed movement of UAVs, the rapid changes in target position can lead to delays in signal strength tracking methods, while visual feature methods are susceptible to dynamic blurring or occlusion, resulting in inaccurate predictions of target position. Although multi-method integration techniques improve the adaptability of tracking systems, bottlenecks remain in balancing real-time performance and accuracy. Particularly in complex dynamic scenarios, insufficient information fusion efficiency can further reduce the success rate of tracking. Summary of the Invention
[0005] This invention provides a target tracking method and apparatus based on ring network fusion, which addresses the bottleneck in the trade-off between real-time performance and accuracy in existing technologies. In complex dynamic scenarios, insufficient information fusion efficiency may further reduce the success rate of tracking. The technical solution of this invention extracts features from visual images and target tracking signals separately through ordinary differential equation networks and combines them with ring network information fusion, which significantly improves the speed and success rate of target tracking, providing an efficient and accurate technical solution for intelligent target tracking of UAVs in high-speed motion scenarios.
[0006] This invention provides a target tracking method based on ring network fusion, comprising the following steps.
[0007] Feature extraction of visual images is performed based on ordinary differential equation networks to obtain the first visual image features; The target tracking signal is sparsified based on the ordinary differential equation network to obtain the sparse features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals. The first visual image features and the sparsed features of the tracking signal are mapped to different neurons of a ring network to obtain the input features; The input features are iterated multiple times to obtain target fusion features, and a target tracking strategy is determined based on the target fusion features.
[0008] According to the present invention, a target tracking method based on ring network fusion is provided, wherein the feature extraction of visual images based on ordinary differential equation networks to obtain first visual image features includes: The visual image is subjected to dimensionality reduction processing to obtain the dimensionality reduction features corresponding to the visual image; The dimensionality reduction features are input into the ordinary differential equation network to obtain the first visual image features output by the ordinary differential equation network.
[0009] According to the present invention, a target tracking method based on ring network fusion is provided, wherein the ordinary differential equation network includes a first layer ordinary differential equation network and a second layer ordinary differential equation network. The step of inputting the dimensionality-reduced features into the ordinary differential equation network to obtain the first visual image features output by the ordinary differential equation network includes: The dimensionality-reduced features are input into the first-layer ordinary differential equation network to obtain the initial visual image features output by the first-layer ordinary differential equation network; the first-layer ordinary differential equation network is used to perform global feature extraction on the dimensionality-reduced features. The initial visual image features are input into the second-layer ordinary differential equation network to obtain the first visual image features output by the second-layer ordinary differential equation network; the second-layer ordinary differential equation network is used to perform target enhancement on the initial visual image features.
[0010] According to the target tracking method based on ring network fusion provided by the present invention, the step of inputting the dimensionality-reduced features into the first layer ordinary differential equation network to obtain the initial visual image features output by the first layer ordinary differential equation network includes: in, express Initial visual image features at time 10:00 express Dimensionality reduction characteristics at any given moment express The amount of visual enhancement at any given moment, the stated The amount of visual enhancement at any given moment is based on First visual image features at any moment and The sparsity characteristics of the tracking signal at any given time are determined; This represents the global feature extraction function corresponding to the global feature extraction. This represents the first learning parameter.
[0011] According to the target tracking method based on ring network fusion provided by the present invention, the target tracking signal is sparsified based on the ordinary differential equation network to obtain sparse features of the tracking signal, including: Obtain the multidimensional signal features corresponding to the target tracking signal; the multidimensional signal features include signal strength, signal direction information, and signal frequency domain features. The multidimensional signal features are encoded into a time series format to obtain the multidimensional signal time series; The multidimensional signal time series is input into the ordinary differential equation network to obtain the sparsification features of the tracking signal output by the ordinary differential equation network; the ordinary differential equation network is also used to perform sparsification processing on the multidimensional signal time series.
[0012] According to the present invention, a target tracking method based on ring network fusion is provided, wherein the multidimensional signal time series includes multidimensional signal component features; The step of inputting the multidimensional signal time series into the ordinary differential equation network to obtain the sparsity features of the tracking signal output by the ordinary differential equation network includes: For each of the signal component features, the signal component features are input into the ordinary differential equation network to obtain the sparse component features output by the ordinary differential equation network; The sparsity features of the tracking signal are determined based on all the sparsity component features.
[0013] According to the target tracking method based on ring network fusion provided by the present invention, the step of inputting the signal component features into the ordinary differential equation network to obtain the sparse component features output by the ordinary differential equation network includes: in, express The first moment The sparsified component features corresponding to the 3D signal component features express The first moment 3D signal component characteristics, This represents the time-series feature extraction function. Indicates the second learning parameter. This represents the third learning parameter. This represents the sparsification function corresponding to the sparsification process.
[0014] According to the present invention, a target tracking method based on ring network fusion is provided, wherein mapping the first visual image features and the sparsed features of the tracking signal to different neurons of the ring network to obtain input features includes: in, Indicates input features, This represents the visually learnable projection matrix corresponding to the ring network. express The first visual image features at any given moment This indicates that the signal corresponding to the ring network is a learnable projection matrix. express The sparsity characteristics of the tracking signal at any given time.
[0015] According to the target tracking method based on ring network fusion provided by the present invention, the step of performing multiple rounds of iteration on the input features to obtain target fusion features includes: For the current iteration time, the fusion feature corresponding to the current iteration time is determined based on the input features, the connection matrix corresponding to the ring network, and the fusion feature corresponding to the previous iteration time. The target fusion feature is determined based on the fusion features corresponding to all iteration times.
[0016] According to the target tracking method based on ring network fusion provided by the present invention, the connection matrix corresponding to the ring network is determined based on the following method: Determine the connection weights between each pair of neurons in the ring network; The connection matrix is determined based on all the connection weights.
[0017] According to the present invention, a target tracking method based on ring network fusion is provided, wherein determining the target tracking strategy based on the target fusion features includes: The target fusion features are projected onto a two-dimensional target space to obtain the roll angle and yaw angle; The target tracking strategy is determined based on the roll angle and the yaw angle; the target tracking strategy is used to control the UAV to track the target unit.
[0018] The present invention also provides a target tracking device based on ring network fusion, comprising the following modules: The feature extraction module is used to extract features from visual images based on ordinary differential equation networks to obtain the first visual image features; The sparsification module is used to perform sparsification processing on the target tracking signal based on the ordinary differential equation network to obtain the sparsified features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals. The mapping module is used to map the first visual image features and the sparsed features of the tracking signal to different neurons of the ring network to obtain input features; The tracking module is used to perform multiple iterations on the input features to obtain target fusion features, and to determine a target tracking strategy based on the target fusion features.
[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the target tracking method based on ring network fusion as described above.
[0020] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the target tracking method based on ring network fusion as described above.
[0021] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the target tracking method based on ring network fusion as described above.
[0022] This invention provides a target tracking method and apparatus based on ring network fusion. It extracts features from visual images using an ordinary differential equation network (ODE) to obtain first visual image features; it then performs sparsification processing on the target tracking signal using the same ODE network to obtain sparse tracking signal features. The ODE network is trained based on historical visual images and historical tracking signals. The first visual image features and the sparse tracking signal features are mapped to different neurons in the ring network to obtain input features. Multiple iterations are performed on the input features to obtain target fusion features, and a target tracking strategy is determined based on these fusion features. This invention's technical solution significantly improves the speed and success rate of target tracking by extracting features from visual images and target tracking signals separately using an ODE network and combining this with ring network information fusion. This provides an efficient and accurate technical solution for intelligent target tracking of UAVs in high-speed motion scenarios. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0024] Figure 1 This is a flowchart illustrating the target tracking method based on ring network fusion provided by the present invention.
[0025] Figure 2 This is a schematic diagram of the target tracking device based on ring network fusion provided by the present invention.
[0026] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0028] To address the aforementioned problems in the prior art, this invention provides a target tracking method based on ring network fusion, which can be applied to unmanned aerial vehicles (UAVs). Figure 1 This is a flowchart illustrating the target tracking method based on ring network fusion provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps 110 to 140.
[0029] Step 110: Extract features from the visual image based on the ordinary differential equation network to obtain the first visual image features.
[0030] Specifically, features can be extracted from the visual image based on an Ordinary Differential Equations (ODE) network to obtain the first visual image features. The visual image can be a visual image acquired by a visual sensor on the UAV, used for obstacle avoidance and target tracking assistance.
[0031] In one embodiment, the step of extracting features from a visual image based on an ordinary differential equation network to obtain first visual image features includes: The visual image is subjected to dimensionality reduction processing to obtain the dimensionality reduction features corresponding to the visual image; The dimensionality reduction features are input into the ordinary differential equation network to obtain the first visual image features output by the ordinary differential equation network.
[0032] Specifically, dimensionality reduction of visual images can be performed using convolutional pooling modules. For example, a 4x4 convolutional kernel can be used to segment the visual image into feature units, reducing the resolution to 1 / 4 of the original while increasing the number of feature channels to retain key feature information, thus obtaining the dimensionality-reduced features of the visual image. Furthermore, the dimensionality-reduced features can be input into an ODE network to obtain the first visual image features output by the ODE network.
[0033] In the above embodiments, dimensionality reduction of the visual image facilitates feature extraction by the subsequent ODE network.
[0034] In one embodiment, the ordinary differential equation network includes a first-layer ordinary differential equation network and a second-layer ordinary differential equation network; The step of inputting the dimensionality-reduced features into the ordinary differential equation network to obtain the first visual image features output by the ordinary differential equation network includes: The dimensionality-reduced features are input into the first-layer ordinary differential equation network to obtain the initial visual image features output by the first-layer ordinary differential equation network; the first-layer ordinary differential equation network is used to perform global feature extraction on the dimensionality-reduced features. The initial visual image features are input into the second-layer ordinary differential equation network to obtain the first visual image features output by the second-layer ordinary differential equation network; the second-layer ordinary differential equation network is used to perform target enhancement on the initial visual image features.
[0035] Specifically, the ODE network consists of a first-layer ODE network and a second-layer ODE network, which can dynamically transform the dimensionality-reduced features. The dimensionality-reduced features can be input into the first-layer ordinary differential equation network to obtain the initial visual image features output by the first-layer ordinary differential equation network. The first layer of the ordinary differential equation network is used to extract global features from the dimensionality-reduced features.
[0036] Furthermore, the initial visual image features can be input into the second-layer ordinary differential equation network (ODE network) to obtain the first visual image features output by the second-layer ODE network. The second-layer ODE network is used to enhance the target features of the initial visual image to highlight the importance of the target region. The first visual image features can be represented by the following formula: in, This indicates the enhancement module corresponding to the target enhancement. This represents the sparsification regularization term. Used to suppress background noise and All of these are learning parameters.
[0037] In the above embodiments, the dynamic dimensionality reduction and transformation of image features through the ODE network significantly reduces redundant computation.
[0038] In one embodiment, inputting the dimensionality-reduced features into the first-layer ordinary differential equation network to obtain the initial visual image features output by the first-layer ordinary differential equation network includes: in, express Initial visual image features at time 10:00 express Dimensionality reduction characteristics at any given moment express The amount of visual enhancement at any given moment, the stated The amount of visual enhancement at any given moment is based on First visual image features at any moment and The sparsity characteristics of the tracking signal at any given time are determined; This represents the global feature extraction function corresponding to the global feature extraction. This represents the first learning parameter.
[0039] Specifically, in this embodiment, the formula used to solve for the initial visual image features is essentially a continuous-time dynamic system model. express The amount of visual enhancement at any given moment The amount of visual enhancement at any given moment is based on First visual image features at any moment and The sparsity characteristics of the tracking signal at each time step are determined, and can be specifically expressed by the following formula: in, This represents the enhancement factor, used to adjust the feedback strength. express The tracking signal exhibits sparsity at each time step. It's easy to understand that the visual enhancement at the initial time step is zero.
[0040] In the above embodiments, a visual enhancement amount is introduced to feed back the signal direction information to the visual feature layer, thereby enhancing the salience of the target area.
[0041] Step 120: Based on the ordinary differential equation network, the target tracking signal is sparsified to obtain the sparse features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals.
[0042] Specifically, the target tracking signal can be sparsified based on the ordinary differential equation network to obtain the sparsified features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals.
[0043] In one embodiment, the sparsification of the target tracking signal based on the ordinary differential equation network to obtain sparsed features of the tracking signal includes: Obtain the multidimensional signal features corresponding to the target tracking signal; the multidimensional signal features include signal strength, signal direction information, and signal frequency domain features. The multidimensional signal features are encoded into a time series format to obtain the multidimensional signal time series; The multidimensional signal time series is input into the ordinary differential equation network to obtain the sparsification features of the tracking signal output by the ordinary differential equation network; the ordinary differential equation network is also used to perform sparsification processing on the multidimensional signal time series.
[0044] Specifically, target tracking signals can be acquired, and multi-dimensional signal features can be determined based on the target tracking signals. These multi-dimensional signal features include signal strength, signal direction information, and signal frequency domain features.
[0045] Furthermore, the multidimensional signal features can be encoded into a time series format to obtain a multidimensional signal time series. ,in, This represents the characteristics of the first-dimensional signal component. This represents the characteristics of the second-dimensional signal component. Indicates the first 3D signal component characteristics, i.e. It indicates the position of a multidimensional signal time series.
[0046] After obtaining the multidimensional signal time series, the multidimensional signal time series can be input into an ordinary differential equation network. The ordinary differential equation network can also perform sparsification processing on the multidimensional signal time series to obtain the sparsified characteristics of the tracking signal output by the ordinary differential equation network.
[0047] In the above embodiments, the robustness of tracking signal processing is enhanced by sparse modeling and temporal dynamic modeling.
[0048] In one embodiment, the multidimensional signal time series includes multidimensional signal component features; The step of inputting the multidimensional signal time series into the ordinary differential equation network to obtain the sparsity features of the tracking signal output by the ordinary differential equation network includes: For each of the signal component features, the signal component features are input into the ordinary differential equation network to obtain the sparse component features output by the ordinary differential equation network; The sparsity features of the tracking signal are determined based on all the sparsity component features.
[0049] Specifically, regarding the characteristics of signal components , to characterize signal components Inputting an ordinary differential equation network yields the sparsified component features of the network's output. .
[0050] Furthermore, the sparsity features of the tracking signal can be determined based on the features of all sparsity components. .
[0051] In the above embodiments, compared with existing methods, this process uses sparse modeling and time dynamic modeling to efficiently process and suppress noise in the tracking signal.
[0052] In one embodiment, the step of inputting the signal component features into the ordinary differential equation network to obtain the sparse component features output by the ordinary differential equation network includes: in, express The first moment The sparsified component features corresponding to the 3D signal component features express The first moment 3D signal component characteristics, This represents the time-series feature extraction function. Indicates the second learning parameter. This represents the third learning parameter. This represents the sparsification function corresponding to the sparsification process.
[0053] In the above embodiments, by simulating the sparsity process using an ODE network, multipath interference can be eliminated.
[0054] Step 130: Map the first visual image features and the sparse features of the tracking signal to different neurons of the ring network to obtain the input features.
[0055] Specifically, a ring neural network centered on the signal direction can be pre-constructed, and the features of the first visual image and the sparsified features of the tracking signal can be mapped to different neurons in the ring network to obtain the input features. The neurons in the ring network are arranged in a closed loop to represent different signal directions.
[0056] In one embodiment, mapping the first visual image features and the sparsed features of the tracking signal to different neurons of a ring network to obtain input features includes: in, Indicates input features, This represents the visually learnable projection matrix corresponding to the ring network. express The first visual image features at any given moment This indicates that the signal corresponding to the ring network is a learnable projection matrix. express The sparsity characteristics of the tracking signal at any given time.
[0057] In the above embodiments, visual information and signal information are dynamically fused through a ring attractor network to achieve efficient integration of multimodal features.
[0058] Step 140: Perform multiple iterations on the input features to obtain target fusion features, and determine the target tracking strategy based on the target fusion features.
[0059] Specifically, the input features can be iterated through network dynamics to obtain target fusion features, and the target tracking strategy can be determined based on the target fusion features.
[0060] In one embodiment, performing multiple iterations on the input features to obtain the target fused features includes: For the current iteration time, the fusion feature corresponding to the current iteration time is determined based on the input features, the connection matrix corresponding to the ring network, and the fusion feature corresponding to the previous iteration time. The target fusion feature is determined based on the fusion features corresponding to all iteration times.
[0061] Specifically, for the current iteration time, the fusion feature corresponding to the current iteration time can be determined based on the input features, the connection matrix corresponding to the ring network, and the fusion feature corresponding to the previous iteration time. For example, the current iteration time is... At time, the fused feature corresponding to the current iteration time. It can be expressed by the following formula: in, This represents the connection matrix corresponding to the ring network.
[0062] Furthermore, the target fusion feature can be determined based on the fusion features corresponding to all iteration times. It should be noted that the initial time... It is 0.
[0063] In the above embodiments, multimodal feature fusion based on ring attractor networks enables faster and more accurate target tracking.
[0064] In one embodiment, the connection matrix corresponding to the ring network is determined based on the following method: Determine the connection weights between each pair of neurons in the ring network; The connection matrix is determined based on all the connection weights.
[0065] Specifically, the connection weights between neurons can be defined using the Gaussian kernel function. The connection weights between any two neurons in a circular network can be expressed by the following formula: in, Both represent the indices of neurons. This represents the natural exponential function. This represents the diffusion parameter, which is used to control the range of diffusion in the connection.
[0066] Furthermore, the connection matrix can be determined based on all connection weights.
[0067] In the above embodiments, the connection weights between neurons are determined by the Gaussian kernel function, and then the connection matrix is determined, which lays the foundation for the determination of fusion features.
[0068] In one embodiment, determining the target tracking strategy based on the target fusion features includes: The target fusion features are projected onto a two-dimensional target space to obtain the roll angle and yaw angle; The target tracking strategy is determined based on the roll angle and the yaw angle; the target tracking strategy is used to control the UAV to track the target unit.
[0069] Specifically, the target fusion features can be projected onto a two-dimensional target space using a dimensionality reduction module to obtain the roll angle and yaw angle. This process can be represented by the following formula: in, Indicates the roll angle. Indicates the yaw angle. It can represent a fully connected layer of a ring network. This indicates the target fusion feature.
[0070] Furthermore, a target tracking strategy can be determined based on the roll angle and yaw angle; the target tracking strategy is used to control the UAV to track the target unit, that is, the roll angle and yaw angle of the UAV can be controlled based on the target tracking strategy to control the direction of the UAV.
[0071] In the above embodiments, the target tracking strategy is determined by roll angle and yaw angle to control the UAV, which enables precise directional control of the UAV and achieves precise obstacle avoidance and target tracking.
[0072] This invention provides a target tracking method based on ring network fusion. It extracts features from visual images using an ordinary differential equation network (ODE) to obtain first visual image features; it then performs sparsification processing on the target tracking signal using the same ODE network to obtain sparse tracking signal features. The ODE network is trained based on historical visual images and historical tracking signals. The first visual image features and the sparse tracking signal features are mapped to different neurons in the ring network to obtain input features. Multiple iterations are performed on the input features to obtain target fusion features, and a target tracking strategy is determined based on these fusion features. This invention significantly improves the speed and success rate of target tracking by extracting features from visual images and target tracking signals separately using an ODE network and combining this with ring network information fusion. This provides an efficient and accurate technical solution for intelligent target tracking of UAVs in high-speed motion scenarios.
[0073] The target tracking device based on ring network fusion provided by the present invention is described below. The target tracking device based on ring network fusion described below and the target tracking method based on ring network fusion described above can be referred to in correspondence.
[0074] Figure 2 This is a schematic diagram of the target tracking device based on ring network fusion provided by the present invention, as shown below. Figure 2 As shown, the target tracking device 200 based on ring network fusion includes the following modules: Feature extraction module 210 is used to extract features from visual images based on ordinary differential equation networks to obtain first visual image features; The sparsity module 220 is used to perform sparsification processing on the target tracking signal based on the ordinary differential equation network to obtain the sparsified features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals. Mapping module 230 is used to map the first visual image features and the sparse features of the tracking signal to different neurons of the ring network to obtain input features; The tracking module 240 is used to perform multiple iterations on the input features to obtain target fusion features, and to determine a target tracking strategy based on the target fusion features.
[0075] In one embodiment, the feature extraction module 210 is specifically used for: The visual image is subjected to dimensionality reduction processing to obtain the dimensionality reduction features corresponding to the visual image; The dimensionality reduction features are input into the ordinary differential equation network to obtain the first visual image features output by the ordinary differential equation network.
[0076] In one embodiment, the ordinary differential equation network includes a first-layer ordinary differential equation network and a second-layer ordinary differential equation network; the feature extraction module 210 is further configured to: The dimensionality-reduced features are input into the first-layer ordinary differential equation network to obtain the initial visual image features output by the first-layer ordinary differential equation network; the first-layer ordinary differential equation network is used to perform global feature extraction on the dimensionality-reduced features. The initial visual image features are input into the second-layer ordinary differential equation network to obtain the first visual image features output by the second-layer ordinary differential equation network; the second-layer ordinary differential equation network is used to perform target enhancement on the initial visual image features.
[0077] In one embodiment, the feature extraction module 210 is further configured to: in, express Initial visual image features at time 10:00 express Dimensionality reduction characteristics at any given moment express The amount of visual enhancement at any given moment, the stated The amount of visual enhancement at any given moment is based on First visual image features at any moment and The sparsity characteristics of the tracking signal at any given time are determined; This represents the global feature extraction function corresponding to the global feature extraction. This represents the first learning parameter.
[0078] In one embodiment, the sparsification module 220 is specifically used for: Obtain the multidimensional signal features corresponding to the target tracking signal; the multidimensional signal features include signal strength, signal direction information, and signal frequency domain features. The multidimensional signal features are encoded into a time series format to obtain the multidimensional signal time series; The multidimensional signal time series is input into the ordinary differential equation network to obtain the sparsification features of the tracking signal output by the ordinary differential equation network; the ordinary differential equation network is also used to perform sparsification processing on the multidimensional signal time series.
[0079] In one embodiment, the multidimensional signal time series includes multidimensional signal component features; the sparsity module 220 is further configured to: For each of the signal component features, the signal component features are input into the ordinary differential equation network to obtain the sparse component features output by the ordinary differential equation network; The sparsity features of the tracking signal are determined based on all the sparsity component features.
[0080] In one embodiment, the sparsification module 220 is further configured to: in, express The first moment The sparsified component features corresponding to the 3D signal component features express The first moment 3D signal component characteristics, This represents the time-series feature extraction function. Indicates the second learning parameter. This represents the third learning parameter. This represents the sparsification function corresponding to the sparsification process.
[0081] In one embodiment, the mapping module 230 is specifically used for: in, Indicates input features, This represents the visually learnable projection matrix corresponding to the ring network. express The first visual image features at any given moment This indicates that the signal corresponding to the ring network is a learnable projection matrix. express The sparsity characteristics of the tracking signal at any given time.
[0082] In one embodiment, the tracking module 240 is specifically used for: For the current iteration time, the fusion feature corresponding to the current iteration time is determined based on the input features, the connection matrix corresponding to the ring network, and the fusion feature corresponding to the previous iteration time. The target fusion feature is determined based on the fusion features corresponding to all iteration times.
[0083] In one embodiment, the target tracking device based on ring network fusion further includes a matrix determination module, which is specifically used for: Determine the connection weights between each pair of neurons in the ring network; The connection matrix is determined based on all the connection weights.
[0084] In one embodiment, the tracking module 240 is further configured to: The target fusion features are projected onto a two-dimensional target space to obtain the roll angle and yaw angle; The target tracking strategy is determined based on the roll angle and the yaw angle; the target tracking strategy is used to control the UAV to track the target unit.
[0085] The target tracking device based on ring network fusion provided by this invention extracts features from visual images using an ordinary differential equation network (ODE) to obtain first visual image features; it then performs sparsification processing on the target tracking signal using the ODE network to obtain sparse tracking signal features. The ODE network is trained based on historical visual images and historical tracking signals. The first visual image features and the sparse tracking signal features are mapped to different neurons in the ring network to obtain input features. Multiple iterations are performed on the input features to obtain target fusion features, and a target tracking strategy is determined based on these fusion features. This invention significantly improves the speed and success rate of target tracking by extracting features from visual images and target tracking signals separately using an ODE network and combining this with ring network information fusion, providing an efficient and accurate technical solution for intelligent target tracking of UAVs in high-speed motion scenarios.
[0086] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 3 As shown, the electronic device may include: a processor 310, a communications interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communications interface 320, and the memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a target tracking method based on ring network fusion, the method including: Feature extraction of visual images is performed based on ordinary differential equation networks to obtain the first visual image features; The target tracking signal is sparsified based on the ordinary differential equation network to obtain the sparse features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals. The first visual image features and the sparsed features of the tracking signal are mapped to different neurons of a ring network to obtain the input features; The input features are iterated multiple times to obtain target fusion features, and a target tracking strategy is determined based on the target fusion features.
[0087] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0088] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the target tracking method based on ring network fusion provided by the above methods, the method comprising: Feature extraction of visual images is performed based on ordinary differential equation networks to obtain the first visual image features; The target tracking signal is sparsified based on the ordinary differential equation network to obtain the sparse features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals. The first visual image features and the sparsed features of the tracking signal are mapped to different neurons of a ring network to obtain the input features; The input features are iterated multiple times to obtain target fusion features, and a target tracking strategy is determined based on the target fusion features.
[0089] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the target tracking method based on ring network fusion provided by the above methods, the method comprising: Feature extraction of visual images is performed based on ordinary differential equation networks to obtain the first visual image features; The target tracking signal is sparsified based on the ordinary differential equation network to obtain the sparse features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals. The first visual image features and the sparsed features of the tracking signal are mapped to different neurons of a ring network to obtain the input features; The input features are iterated multiple times to obtain target fusion features, and a target tracking strategy is determined based on the target fusion features.
[0090] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0091] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A target tracking method based on ring network fusion, characterized in that, include: Feature extraction of visual images is performed based on ordinary differential equation networks to obtain the first visual image features; The target tracking signal is sparsified based on the ordinary differential equation network to obtain the sparse features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals. The first visual image features and the sparsed features of the tracking signal are mapped to different neurons of a ring network to obtain the input features; The input features are iterated multiple times to obtain target fusion features, and a target tracking strategy is determined based on the target fusion features.
2. The target tracking method based on ring network fusion according to claim 1, characterized in that, The feature extraction of the visual image based on the ordinary differential equation network to obtain the first visual image features includes: The visual image is subjected to dimensionality reduction processing to obtain the dimensionality reduction features corresponding to the visual image; The dimensionality reduction features are input into the ordinary differential equation network to obtain the first visual image features output by the ordinary differential equation network.
3. The target tracking method based on ring network fusion according to claim 2, characterized in that, The ordinary differential equation network includes a first-layer ordinary differential equation network and a second-layer ordinary differential equation network. The step of inputting the dimensionality-reduced features into the ordinary differential equation network to obtain the first visual image features output by the ordinary differential equation network includes: The dimensionality-reduced features are input into the first-layer ordinary differential equation network to obtain the initial visual image features output by the first-layer ordinary differential equation network; the first-layer ordinary differential equation network is used to perform global feature extraction on the dimensionality-reduced features. The initial visual image features are input into the second-layer ordinary differential equation network to obtain the first visual image features output by the second-layer ordinary differential equation network; the second-layer ordinary differential equation network is used to perform target enhancement on the initial visual image features.
4. The target tracking method based on ring network fusion according to claim 3, characterized in that, The step of inputting the dimensionality-reduced features into the first-layer ordinary differential equation network to obtain the initial visual image features output by the first-layer ordinary differential equation network includes: in, express Initial visual image features at time 10:00 express Dimensionality reduction characteristics at any given moment express The amount of visual enhancement at any given moment, the stated The amount of visual enhancement at any given moment is based on First visual image features at any moment and The sparsity characteristics of the tracking signal at any given time are determined; This represents the global feature extraction function corresponding to the global feature extraction. This represents the first learning parameter.
5. The target tracking method based on ring network fusion according to claim 1, characterized in that, The process of sparsifying the target tracking signal based on the ordinary differential equation network to obtain the sparsified features of the tracking signal includes: Obtain the multidimensional signal features corresponding to the target tracking signal; the multidimensional signal features include signal strength, signal direction information, and signal frequency domain features. The multidimensional signal features are encoded into a time series format to obtain the multidimensional signal time series; The multidimensional signal time series is input into the ordinary differential equation network to obtain the sparsification features of the tracking signal output by the ordinary differential equation network; the ordinary differential equation network is also used to perform sparsification processing on the multidimensional signal time series.
6. The target tracking method based on ring network fusion according to claim 5, characterized in that, The multidimensional signal time series includes multidimensional signal component features; The step of inputting the multidimensional signal time series into the ordinary differential equation network to obtain the sparsity features of the tracking signal output by the ordinary differential equation network includes: For each of the signal component features, the signal component features are input into the ordinary differential equation network to obtain the sparse component features output by the ordinary differential equation network; The sparsity features of the tracking signal are determined based on all the sparsity component features.
7. The target tracking method based on ring network fusion according to claim 6, characterized in that, The step of inputting the signal component features into the ordinary differential equation network to obtain the sparse component features output by the ordinary differential equation network includes: in, express The first moment The sparsified component features corresponding to the 3D signal component features express The first moment 3D signal component characteristics, This represents the time-series feature extraction function. Indicates the second learning parameter. This represents the third learning parameter. This represents the sparsification function corresponding to the sparsification process.
8. The target tracking method based on ring network fusion according to claim 1, characterized in that, The step of mapping the first visual image features and the sparsed features of the tracking signal to different neurons of a ring network to obtain input features includes: in, Indicates input features, This represents the visually learnable projection matrix corresponding to the ring network. express The first visual image features at any given moment This indicates that the signal corresponding to the ring network is a learnable projection matrix. express The sparsity characteristics of the tracking signal at any given time.
9. The target tracking method based on ring network fusion according to any one of claims 1 to 8, characterized in that, The process of performing multiple iterations on the input features to obtain the target fusion features includes: For the current iteration time, the fusion feature corresponding to the current iteration time is determined based on the input features, the connection matrix corresponding to the ring network, and the fusion feature corresponding to the previous iteration time. The target fusion feature is determined based on the fusion features corresponding to all iteration times.
10. The target tracking method based on ring network fusion according to claim 9, characterized in that, The connection matrix corresponding to the ring network is determined based on the following method: Determine the connection weights between each pair of neurons in the ring network; The connection matrix is determined based on all the connection weights.
11. The target tracking method based on ring network fusion according to claim 10, characterized in that, The step of determining the target tracking strategy based on the target fusion features includes: The target fusion features are projected onto a two-dimensional target space to obtain the roll angle and yaw angle; The target tracking strategy is determined based on the roll angle and the yaw angle; the target tracking strategy is used to control the UAV to track the target unit.
12. A target tracking device based on ring network fusion, characterized in that, include: The feature extraction module is used to extract features from visual images based on ordinary differential equation networks to obtain the first visual image features; The sparsification module is used to perform sparsification processing on the target tracking signal based on the ordinary differential equation network to obtain the sparsified features of the tracking signal; the ordinary differential equation network is trained based on historical visual images and historical tracking signals. The mapping module is used to map the first visual image features and the sparsed features of the tracking signal to different neurons of the ring network to obtain input features; The tracking module is used to perform multiple iterations on the input features to obtain target fusion features, and to determine a target tracking strategy based on the target fusion features.