A target tracking method, system, device and medium for an Internet of Things resource-constrained terminal device

Abstract

The application provides a target tracking method, system, device and medium for an Internet of Things resource-limited terminal device, which comprises the following steps: training a translation filter and a scale filter based on an adaptive feature dimension and an adaptive time scale interval to obtain a scale filter based on an adaptive time scale interval and a translation filter with an adaptive dimension feature; predicting a target position through the translation filter with the adaptive dimension feature and a histogram model; and then calculating a final target size by using the scale filter with the adaptive time scale interval. The application reduces the calculation amount and further improves the tracking efficiency.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) terminal devices, and specifically relates to a target tracking method, system, device, and medium for IoT resource-constrained terminal devices. Background Technology

[0002] Sixth-generation (6G) wireless communication technology has driven the rapid development of the Internet of Things (IoT), enabling the widespread deployment of various terminal devices in applications such as wireless multimedia sensor networks. However, due to limited computing resources, most terminal devices struggle to process massive amounts of IoT video data for visual target tracking.

[0003] Visual target tracking has wide applications in IoT fields such as intelligent transportation, wildlife monitoring, and military surveillance, but it faces challenges in scenarios involving partial occlusion, motion blur, varying lighting, cluttered backgrounds, and scale changes. Currently, the most popular visual tracking methods are mainly divided into trackers based on Discriminative Correlation Filters (DCF) and trackers based on deep learning. While deep learning-based methods offer superior performance, they rely on expensive GPUs, making them impractical for most terminal devices equipped only with general-purpose CPUs. Furthermore, the stringent real-time requirements of some IoT applications further complicate tracking.

[0004] DCF trackers are suitable for resource-constrained terminal devices due to their high computational efficiency. However, most existing DCF trackers use multiple features and spatiotemporal scales to estimate the target state. The fixed feature dimensions and dense scale intervals result in high computational costs and are not optimal when dealing with large-scale changes in the target. It is difficult to achieve a good balance between effectiveness and efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a target tracking method for resource-constrained IoT terminal devices, in order to solve the problems of high computational cost and low efficiency caused by fixed feature dimensions and dense scale intervals in existing DCF trackers on resource-constrained terminal devices, and to achieve efficient and accurate target tracking.

[0006] The technical solution of the present invention is as follows:

[0007] A target tracking method for resource-constrained IoT terminal devices, the method comprising:

[0008] The translation filter and the scaling filter are trained based on adaptive feature dimension and adaptive time scale interval to obtain a scaling filter based on adaptive time scale interval and a translation filter with adaptive dimension feature.

[0009] The target position is predicted by a translation filter with adaptive dimensionality features and a histogram model; then the final target size is calculated using a scaling filter with adaptive time scale intervals.

[0010] Furthermore, training the translation filter based on adaptive feature dimensions specifically includes:

[0011] Principal component analysis is used for feature dimensionality reduction. By minimizing the reconstruction error, a projection matrix is ​​constructed, and the features are projected onto a low-dimensional subspace. The features are then reconstructed through Fourier transform and compressed to train a translation filter.

[0012] Furthermore, the expression for the adaptive time scale interval is:

[0013]

[0014] in, Indicates the frequency of the switching scale filter. This represents the original time-domain scale interval.

[0015] Furthermore, the target location is predicted using a translation filter and histogram model with adaptive dimensional features, specifically as follows:

[0016] A histogram model is constructed based on the color map, and the pixel-by-pixel score map of the target template is calculated. The score map is then fused with the Hanning window through adaptive mapping fusion parameters to generate adaptive mapping features.

[0017] Furthermore, the adaptive mapping feature formula is as follows:

[0018] ψ'=ψ⊙{χ(H⊙W)+(1-χ)W},

[0019] Where χ is the adaptive mapping fusion parameter, ψ' represents the dimensionality-reduced adaptive mapping feature, H is the adaptive mapping obtained by calculating the pixel-by-pixel score of the target template using a histogram model based on , W is the Hanning window, and ψ is the adaptive dimensionality feature.

[0020] A target tracking system for resource-constrained IoT terminal devices, the system comprising:

[0021] The filter training module is configured to train translation and scaling filters based on adaptive feature dimensions and adaptive time scale intervals to obtain a scaling filter based on adaptive time scale intervals and a translation filter with adaptive dimensional features.

[0022] The training and target prediction module is configured to predict the target position using a translation filter with adaptive dimensionality features and a histogram model; then, a scaling filter with adaptive time scale intervals is used to calculate the final target size.

[0023] An electronic device includes a memory and a processor, the memory storing a computer program and the processor running the computer program to cause the electronic device to perform the target tracking method described above.

[0024] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the target tracking method described above.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] Improved tracking efficiency: The dimensionality reduction of the feature response through adaptive mapping greatly reduces the feature dimension, thereby reducing the amount of computation; the adaptive time scale interval estimation reduces the frequency of scale estimation, further improving tracking efficiency.

[0027] Ensuring tracking accuracy: The adaptive mapping feature response cleverly combines dimensionality reduction and histogram score map to generate the final response, which significantly enhances the robustness of state estimation; the adaptive time scale interval estimation can effectively adapt to changes in target scale while reducing the amount of computation, thus strongly guaranteeing tracking accuracy. Attached Figure Description

[0028] The accompanying drawings illustrate various embodiments generally by way of example rather than limitation, and are used, together with the specification and claims, to explain embodiments of the invention. Where appropriate, the same reference numerals are used in all drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the apparatus or method.

[0029] Figure 1 A schematic diagram of the method flow of the present invention is shown. Detailed Implementation

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] This invention provides a target tracking method for resource-constrained IoT terminal devices, comprising:

[0032] The translation filter and the scaling filter are trained based on adaptive feature dimension and adaptive time scale interval to obtain a scaling filter based on adaptive time scale interval and a translation filter with adaptive dimension feature.

[0033] The target position is predicted by a translation filter with adaptive dimensionality features and a histogram model; then the final target size is calculated using a scaling filter with adaptive time scale intervals.

[0034] This invention introduces a response interference suppression (RISTrack) mechanism to improve temporal consistency and suppress background interference. The modified objective is defined as an energy function:

[0035]

[0036] Where * represents correlation calculations and ⊙ represents element-wise multiplication. This is to minimize the difference between the predicted response map and the Gaussian-shaped label y. c The objective function, centered on the target, is learned through supervised correlation filtering to determine the difference. The predicted response map is obtained by extracting the c-th channel of the target appearance features from the current frame t. With the filter g to be trained c This is obtained through relevant calculations. Here, ω represents the penalty regularization weight. Furthermore, to effectively suppress background interference, a new part is introduced into the objective function. Where R... t-1 This represents the response graph of the previous frame. It is a binary mask centered at the position (i,j) preceding the target. P is the background enhancement feature extracted from frame t, and P is the pixel-wise background weight map (where the weight value of the background region is higher than that of the target region).

[0037] By removing the subscript (·) c The solution to (1) is decoupled into C independent subproblems. To facilitate efficient optimization, the objective function is reconstructed in the frequency domain by introducing an auxiliary variable u and equality constraints. The resulting frequency domain formula is as follows:

[0038]

[0039] Where ^ denotes the Discrete Fourier Transform, and ˉ denotes the complex conjugate. Equality constraints Ensure that the filter learned in the frequency domain is consistent with its corresponding term in the spatial domain. To solve this constrained problem, the augmented Lagrangian method is used, yielding the following objective function:

[0040]

[0041] In this approach, a penalty term and a dual variable λ are introduced to enforce soft constraints. In the frequency domain, μ acts as a Lagrange multiplier, representing a penalty parameter for balancing constraint strength and optimization stability. Subsequently, the ADMM algorithm is used to iteratively solve the three subproblems. First, the solution is obtained at iteration i. The closed-form solution is as follows:

[0042]

[0043] in, Representing training features With tags The cross-correlation, and and These are the autocorrelation values ​​of the target feature and the background enhancement feature, respectively. A penalty term μ is used to ensure convergence and numerical stability. Next, the auxiliary variable u is updated as follows:

[0044]

[0045] in This represents the inverse discrete Fourier transform. In the i-th iteration, Update as follows:

[0046]

[0047] Here, the penalty parameter μ i By multiplying the previous value by β and using the maximum value μ max Update by applying constraints, i.e., μ i =min{μ max ,βμ i-1}, where μ max An upper limit was set, and β controlled the growth rate. Subsequently, to account for the changing appearance of the target over time, the model template was updated online using an exponential moving average.

[0048]

[0049] Here, η∈[0,1] is the learning rate, used to control the weight allocation between new observations and historical information. This update mechanism allows the model to retain historical templates. While receiving information, gradually adapt to the new appearance features.

[0050] Finally, to ensure that only the area near the expected target center receives high attention, a response-assisted strategy is used to stabilize the response map. The formula is expressed as follows:

[0051]

[0052] Here, R t It is caused by the previous peak (i) c ,j c Spatial weighted mask centered on ) Modulated response diagram. The weight matrix is ​​defined as follows: Where δ<0 controls the steepness of the decay.

[0053] Thanks to spatiotemporal regularization, adaptive appearance modeling, and spatial response smoothing, the learned filters become more robust and exhibit good generalization ability. Meanwhile, many effective acceleration techniques are used to reduce computational cost and achieve real-time tracking when solving for the response interference suppression correlation filter. However, since the RISTrack framework uses multiple features and spatiotemporal scale spaces to address target state estimation, direct operations with fixed feature dimensions and dense scale intervals can lead to high computational costs and are not optimal when handling large-scale target variations.

[0054] This invention proposes an adaptive feature dimension and scale interval method based on histogram score maps to reduce the frequency of feature dimension and scale estimation when updating target location and scale.

[0055] Adaptive mapping feature dimension

[0056] Multi-feature ensemble is often used to improve tracking performance. However, this simple operation of concatenating features is not suitable for improving tracking accuracy and may even lead to a decrease in tracking performance. To avoid this performance degradation in the multi-feature ensemble process, an adaptive mapping of feature responses based on dimensionality reduction and histogram-based score maps is proposed to improve tracking performance. Specifically,

[0057] First, optimal dimension selection is adaptively performed on the ensemble features to maintain the real-time performance of RISTrack. Then, an adaptive color map based on histogram-based per-pixel scores is used to generate the final response from the selected dimensional features to enhance the robustness of state estimation. We apply three commonly used features to the adaptively mapped feature response: Histogram of Oriented Gradients (HOG), Color Name (CN), and Color Map (CM), where CM is applied to the HOG+CN features. Therefore, translation filters and histogram models are trained on the response maps for different features, where... denoted as HOG+CN based filter, and h represents the histogram model based on CM.

[0058] Dimensionality reduction techniques are based on standard principal component analysis (PCA) applied in the adaptive mapping feature response method. The update model template for HOG+CN features is defined as u. (t) =(1-η)u (t-1) +ηψ (t) This is used to adaptively reduce feature dimensionality. Based on the linear property of the Fourier transform... Formula (7) can be reconstructed as Finally, the updated model template u (t) Used to construct the projection matrix P (t) (i.e., low-dimensional subspace). Features can be represented by a size of... The projection matrix P (t)Projected onto this subspace, where Let P be the dimension of the compressed feature. (t) By minimizing the reconstruction update model template u (t) The objective function is obtained as follows:

[0059]

[0060] Where m represents the traversal model template u (t) A tuple of indices for all elements in u. Then, by... (t) The autocorrelation matrix is ​​decomposed into eigenvalues ​​to obtain the solution.

[0061]

[0062] P (t) The row is set to C (t) Having the largest eigenvalue Each feature vector. The target block is obtained through the compressed model template. Updated to:

[0063]

[0064] P (t) Linear operations are defined as element-wise matrix multiplication (P... (t) i (t) (m)=P (t) u (t) (m), i.e., the feature vector Projected onto P (t) The line space. During the detection phase, the model template... The response map is calculated as follows: for the compressed template The filter based on HOG+CN features is applied, and its specific form is similar to that of formula (8):

[0065]

[0066] The Hanning window W is used to avoid boundary effects. Color information is adaptively mapped and applied to adaptive dimensional features in a simple way to generate the final accurate response: ψ'=ψ⊙{χ(H⊙W)+(1-χ)W}, where χ is the adaptive mapping fusion parameter, ψ' represents the dimensionality-reduced adaptive mapping feature, H is the adaptive mapping obtained by calculating the pixel-by-pixel score of the target template through a CM-based histogram model h, and ψ is the adaptive dimensional feature.

[0067] Adaptive timescale interval estimation

[0068] This method uses an adaptive time scale interval. (in Indicates the frequency of the switching scale filter. The original temporal scale interval is used to select video frames for which the target size needs to be estimated. This invention consistently uses the original fixed scale step size in the spatial domain to ensure tracking stability. A scale filter is then applied to these selected video frames with adaptive temporal scale intervals to adapt to the target size. Even if the selected video frames with adaptive scale intervals in the temporal domain may not contain a suitable scale, given the generally small and smooth target scale variation across all video frames, it is likely that a suitable scale can be determined through numerous adaptive scale intervals of subsequently selected video frames.

[0069] By combining response interference suppression correlation filtering, adaptive mapping of feature responses based on dimensionality reduction and histogram score maps, and adaptive time-scale interval estimation, an ARIST (Adaptive Response Interference Suppression Tracker) is proposed. Its flowchart is shown below. Figure 1 As shown. The entire tracking process is divided into two parts: filter training and target prediction. In the training and update phases, the translation filter based on adaptive dimensional features is trained by suppressing response interference, similar to the RISTrack tracker. In addition, a scaling filter based on adaptive time scale intervals and a histogram model based on color maps are prepared in parallel. In the target prediction phase, the target position is first predicted by formula (12) using the translation filter with adaptive dimensional features and the histogram model with adaptive mapping. Then, the final target size is calculated using the scaling filter with adaptive time scale intervals.

[0070] This invention proposes an adaptive mapping feature and scaling scheme based on a response interference suppression correlation filter to address the issues of fixed feature dimensions and sparse spatiotemporal scale sampling. The paper presents an adaptive mapping feature response and adaptive time scale estimation based on histogram score graph dimensionality reduction, employing features with fewer dimensions and sparse time scale intervals to further improve the effectiveness and efficiency of target state estimation. Extensive testing on the UAV123@10fps, DTB70, UAV112, and UAVDT datasets demonstrates that the proposed tracker offers higher computational efficiency compared to baseline trackers using fixed features and sparse scale intervals, and outperforms state-of-the-art trackers in real-time operation. This makes it suitable for resource-constrained terminal devices in various IoT applications.

[0071] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A target tracking method for resource-constrained terminal devices in the Internet of Things, characterized in that, The method includes: A response interference suppression mechanism is introduced to suppress background interference by constructing an energy function; The translation filter and the scaling filter are trained based on adaptive feature dimension and adaptive time scale interval to obtain a scaling filter based on adaptive time scale interval and a translation filter with adaptive dimension feature. The target position is predicted using a translation filter with adaptive dimensionality features and a histogram model; then the final target size is calculated using a scaling filter with adaptive time scale intervals. The energy function is: in, For the filter to be trained, The total number of channels is a feature. Indicates the first The filter corresponding to each feature channel For Gaussian labels, For the target appearance features, To penalize regularization weights, The background suppression coefficient, This represents the response graph of the previous frame. It is a binary mask centered on the position preceding the target. It is a background enhancement feature. It is a pixel-by-pixel background weight map.

2. The target tracking method for resource-constrained terminal devices in the Internet of Things according to claim 1, characterized in that, Training a translation filter based on adaptive feature dimensions specifically includes: Principal component analysis is used for feature dimensionality reduction. By minimizing the reconstruction error, a projection matrix is ​​constructed, and the features are projected onto a low-dimensional subspace. The features are then reconstructed through Fourier transform and compressed to train a translation filter.

3. The target tracking method for resource-constrained terminal devices in the Internet of Things according to claim 1, characterized in that, The expression for the adaptive time scale interval is: in, Indicates the frequency of the switching scale filter. This represents the original time-domain scale interval.

4. The target tracking method for resource-constrained terminal devices in the Internet of Things according to claim 1, characterized in that, Specifically, the target location is predicted using a translation filter and histogram model with adaptive dimensionality features: A histogram model is constructed based on the color map, and the pixel-by-pixel score map of the target template is calculated. The score map is then fused with the Hanning window through adaptive mapping fusion parameters to generate adaptive mapping features.

5. The target tracking method for resource-constrained terminal devices in the Internet of Things according to claim 4, characterized in that, The adaptive mapping feature formula is: ， in, For adaptive mapping fusion parameters, This represents the adaptive mapping features after dimensionality reduction. The adaptive mapping is obtained by calculating the pixel-by-pixel score of the target template using a histogram-based model. For Hanning window, For adaptive dimensional features.

6. A target tracking system for resource-constrained terminal devices in the Internet of Things, characterized in that, The system includes: The filter training module is configured to train translation and scaling filters based on adaptive feature dimensions and adaptive time scale intervals to obtain a scaling filter based on adaptive time scale intervals and a translation filter with adaptive dimensional features. The training and target prediction module is configured to predict the target position using a translation filter with adaptive dimensionality features and a histogram model; then, a scaling filter with adaptive time scale intervals is used to calculate the final target size.

7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the target tracking method according to claims 1-5.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the target tracking method as described in claims 1-5.

Images