Single-target real-time tracking method, system, device and medium under edge device

Abstract

The present application relates to a single-target real-time tracking method, system, device and medium under an edge device, belonging to the technical field of deep learning and image processing, comprising the following steps: constructing an AVTrack single-target lightweight tracking model, performing image pre-processing of an IVE-NPU zero-copy pipeline, performing a cooperative process of detection-tracking-state estimation, and adopting a two-stage post-processing optimization process; the beneficial effects of the present application are: using the Hanning window to suppress the confidence of the tracking box, reducing the influence of target selection instability caused by confidence fluctuation, using EMA exponential weighted average to smooth the tracking box, reducing the jitter of the tracking box, so that the time consumption of the model on the edge device of the standard platform of the present application is reduced from the original 65ms to 28ms, the inference speed reaches 35FPS, meeting the requirements of real-time tracking.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and image processing technology, and specifically relates to a method, system, device and medium for real-time single-target tracking in edge devices, which is applicable to embedded devices with limited computing power and memory. Background Technology

[0002] The significant breakthroughs achieved by the attention mechanism (Transformer) in speech and text translation tasks have provided a theoretical basis for its application in image processing, enabling single object tracking (SOT) to transition from a convolutional neural network-based architecture to a Transformer-based neural network architecture. However, due to the high computational complexity and memory consumption of Transformers (e.g., SimTrack, OSTrack), although representing high accuracy, its large computational complexity and memory consumption make it difficult to run in real time at edge computing.

[0003] Furthermore, when tracking small targets, the Transformer may overemphasize the background area, causing the target to be lost. Especially on edge devices, in existing technologies, the computational complexity of the Transformer's self-attention mechanism is still O(n²) when the input sequence length is short. This results in a single-frame end-to-end latency of more than 30ms under 10 TOPS computing power on the edge side, which is difficult to meet the 30FPS real-time requirement. Summary of the Invention

[0004] This invention provides a method, system, device, and medium for real-time single-target tracking on edge devices, which solves the technical problem of poor real-time target tracking in the prior art. It reduces the model time on edge devices from 65ms to 28ms and achieves an inference speed of 35FPS, thus meeting the requirements for real-time tracking.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0006] A method for real-time single-target tracking on edge devices is provided, including the following steps:

[0007] Construct a lightweight single-target tracking model for AVTrack: Replace the two 16×16 convolutions on the input side of the original model with a stacked structure of four 3×3 convolutions, and add a residual structure branch with 1×1 convolution channel fusion. Retain the receptive field equivalent to a single-level 16×16 convolution. Perform INT8 precision quantization on the model to reduce inference time on edge devices.

[0008] Image preprocessing using the IVE-NPU zero-copy pipeline: The tasks of cropping and scaling templates and search regions are migrated to the IVE on the edge device. By using the zero-copy scheduling method triggered by the IVE and NPU sharing the DDR memory physical address and DMA synchronization, the CPU transfer overhead is eliminated.

[0009] The collaborative process of detection-tracking-state estimation is implemented: the YOLOv8S target detection model is introduced as the relocalization module, and a dual threshold triggering condition is set. When the tracking confidence is lower than the set threshold or the IoU between the target box of the current frame and the previous frame is lower than the set threshold, relocalization is triggered and the template update is paused. Combined with Kalman filtering, the observation noise covariance matrix is ​​dynamically adjusted, and the smoothed target state is output.

[0010] A two-stage post-processing optimization is adopted: the confidence of the tracking box is weighted and suppressed by a Hanning window of size M×N to reduce scoring jitter, and the tracking box coordinates are smoothed by EMA exponential weighted averaging to reduce jitter between adjacent frames.

[0011] Optionally, the parameters of the four 3×3 convolutions are calculated using C. 3×3 =4×128×128×3², and undergoes 4 nonlinear activation function calculations to enhance the extraction of small target features.

[0012] Optionally, the execution steps of the IVE-NPU zero-copy pipeline include:

[0013] IVE Hardware Adaptation: Utilizes the IVE's built-in image processing unit and hardware acceleration instructions to process template and search region cropping and scaling in parallel.

[0014] Memory sharing mechanism: IVE initializes and allocates a physical memory region that overlaps with the NPU, and stores the processed data directly in the shared region;

[0015] Zero-copy pipeline scheduling: IVE and NPU are synchronously triggered via DMA. IVE dynamically calculates coordinates and writes them to shared memory, and NPU reads the processing results.

[0016] Optionally, the dual threshold triggering condition satisfies the following formula:

[0017] ;

[0018] in, This indicates that a relocation has been triggered; This represents the confidence score output for the current tracking frame; This indicates the currently set threshold. This represents the area of ​​the target bounding box in the previous frame; This indicates the area of ​​the target bounding box in the previous frame; It is the intersection-union ratio function; Represents the intersection-union ratio function The set threshold;

[0019] when When the value is 1, the system initiates the relocation process, calling the target detection model to re-detect the target within the search area expanded based on the previous frame's position;

[0020] when When the value is 0, the system does not trigger.

[0021] Optionally, optimizations to the YOLOv8S object detection model include: reducing the input resolution from 640×640 to 256×256 to achieve fast object detection in the current frame's search region; removing the neck structure and retaining only the backbone and head to improve model inference speed.

[0022] Optionally, the state vector of the Kalman filter is ,in, The coordinates of the target center are For velocity components, The target width and height are defined; the diagonal elements of the observation noise covariance matrix Q are dynamically adjusted with the tracking confidence level conf. When the confidence level conf is lower than the threshold, Q is increased to reduce the impact of erroneous observations.

[0023] Optionally, the Hanning window suppresses the tracking box confidence by recording the confidence and target box of the previous frame, and then using a two-dimensional Hanning window for weighted suppression when calculating the confidence of the current frame, and using the Hanning window for local sliding smoothing.

[0024] Additionally, a real-time single-target tracking system for edge devices is provided to execute the real-time single-target tracking method for edge devices described in any one of the above embodiments, including:

[0025] The model inference module is configured to run the AVTrack single-target lightweight tracking model and the optimized YOLOv8S target detection model, and perform INT8 precision quantization inference.

[0026] The preprocessing module is configured as an IVE integration, enabling template and search area cropping, scaling, and memory sharing through a zero-copy pipeline;

[0027] The collaborative control module is configured to execute collaborative logic for detection-tracking-state estimation, including relocation trigger judgment, template update control, and Kalman filter state estimation.

[0028] The post-processing module is configured to perform Hanning window confidence suppression and EMA exponential weighted average smoothing.

[0029] An edge computing device is provided to perform the single-target real-time tracking method under any of the above-described edge devices, comprising:

[0030] At least one NPU computing unit is used to perform deep learning model inference;

[0031] The memory is used to store the target template, candidate box data, and image processing results in shared memory.

[0032] The CPU processor is used to trigger the re-detection thread, control template updates, and coordinate the scheduling of various modules.

[0033] The intelligent video engine performs hardware-accelerated processing for image cropping and scaling, and shares DDR memory with the NPU computing unit.

[0034] A computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the single-target real-time tracking method for edge devices described in any of the above embodiments.

[0035] The beneficial effects of this invention are:

[0036] 1. This invention utilizes an improved AVTrack lightweight network model for single-target tracking, replacing the two 16x16 convolutions in the input with four 3x3 convolutions stacked with residual structures, reducing parameter computation. During model conversion, INT8 precision quantization is used. Compared to lightweight model architectures (NanoTrack, lightTrack), this method reduces model complexity while maintaining accuracy. Without calculating pre- and post-processing, the inference time of the single model at the edge is reduced from 36ms to 22ms. Compared to the original model, the improved model's tracking success rate decreases by only 0.4% (test platform: HiSilicon hi3403, NPU: 5.6TOPS, SVP_NPU: 4.8TOPS, improved AVTrack INT8 quantized model, batch=1, input resolutions of 128x128 and 256x256).

[0037] 2. This invention establishes an IVE-NPU zero-copy pipeline to eliminate preprocessing bottlenecks. Image preprocessing is ported to the device's Intelligent Video Engine (IVE), including the cropping and scaling of templates and search areas. After memory is created on the IVE, it shares the same physical address with the NPU during inference, no longer occupying the CPU. This patented standard platform reduces the preprocessing time from the original 26ms to 4ms, improving the overall inference speed of the algorithm.

[0038] 3. This invention incorporates a YOLOv8s target detection model into the template update and target tracking strategies to achieve "detection-tracking-state estimation" collaboration, solving the tracking failure problem caused by target occlusion and loss. When the tracking confidence is less than a set threshold, the IoU between the target box in the current frame and the target box in the previous frame is calculated, and an IoU threshold is set. If the IoU is lower than the set threshold, the target detection model is used for relocalization, and the template update is stopped to prevent incorrect template updates after relocalization failure. Kalman filtering is also introduced to record the target tracking state.

[0039] 4. This invention uses the Hanning window to suppress the confidence of the tracking box, reducing the impact of confidence fluctuations on target selection instability; the EMA exponential weighted average is used to smooth the tracking box, reducing tracking box jitter; through the above methods, the method proposed in this invention reduces the model's time on the edge device of this patent standard platform from the original 65ms to 28ms, and the inference speed reaches 35FPS, meeting the requirements of real-time tracking. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present 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 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.

[0041] Figure 1 This is a schematic diagram of the original AVTrack single-target lightweight tracking model of this invention;

[0042] Figure 2 This is a schematic diagram of the improved AVTrack single-target lightweight tracking model of the present invention;

[0043] Figure 3 This is a schematic diagram of the single-target real-time tracking system structure of the present invention;

[0044] Figure 4 This is a schematic diagram of the edge computing device structure of the present invention;

[0045] Figure 5 This is a schematic diagram of the connection structure between the computer-readable storage medium and the processor according to the present invention. Detailed Implementation

[0046] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments.

[0047] This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0048] Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this disclosure. The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0049] Example 1;

[0050] like Figure 2 As shown, this embodiment provides a real-time single-target tracking method for edge devices, including the following steps:

[0051] Step S1: Construct the AVTrack single-target lightweight tracking model: Replace the two 16×16 convolutions on the input side of the original model with a stacked structure of four 3×3 convolutions, and add a residual structure branch with 1×1 convolution channel fusion. Retain the receptive field equivalent to a single-level 16×16 convolution. Perform INT8 precision quantization on the model to reduce the inference time of edge devices.

[0052] Step S2: Image preprocessing of the IVE-NPU zero-copy pipeline: The cropping and scaling tasks of the template and search area are migrated to the IVE on the edge device. By using the zero-copy scheduling method triggered by the IVE and NPU sharing the DDR memory physical address and DMA synchronization, the CPU transfer overhead is eliminated.

[0053] Step S3: Execute the collaborative process of detection-tracking-state estimation: Introduce the YOLOv8S target detection model as the relocalization module, set dual threshold trigger conditions, and trigger relocalization and pause template update when the tracking confidence is lower than the set threshold or the IoU of the target box between the current frame and the previous frame is lower than the set threshold. Combine Kalman filtering to dynamically adjust the observation noise covariance matrix and output the smoothed target state.

[0054] Step S4: Employ a two-stage post-processing optimization: Use a Hanning window of size M×N to weight and suppress the confidence of the tracking box, reducing scoring jitter, and use EMA exponential weighted averaging to smooth the tracking box coordinates, reducing jitter between adjacent frames.

[0055] Specifically, in step S1, the improved AVTrack single-target lightweight tracking model replaces the two 16x16 convolutions in the input with four stacked 3x3 convolutions, reducing the number of parameters. Simultaneously, a new network residual structure branch is added, using 1x1 convolutions for channel fusion. The network structure is as follows: Figure 2 As shown, Figure 1 The original AVTrack structure, Figure 2 For the improved AVTrack structure, the output size of each layer is... The calculation formula is as follows:

[0056] ;

[0057] in, Indicates zero padding. Indicates the kernel size. Indicates the step size. Indicates the expansion parameter. For the first Layer output size, For the first Output dimensions of the layer.

[0058] By using the above method, the number of parameters required for computation is reduced from the original 16x16 convolution parameters: C 16x16 =128x128x16 2 Reduced to 4 3x3 convolution parameters: C 3x3 =4 x 128 x 128 x 3 2 The number of parameters reduced is:

[0059] ;

[0060] like Figure 2 As shown, the improved model ensures that four 3x3 receptive fields are equivalent to a single-level 16x16 receptive field, preserving global contextual information features. At the same time, it enhances the feature extraction capability of small targets by undergoing four non-linear activation functions. In addition, the small convolutional kernel is more hardware-friendly and is the dedicated minimum convolutional unit for mobile NPUs, reducing the inference time of a single model on edge devices from the original 36ms to 22ms.

[0061] like Figure 1 As shown, the preprocessing (template / search region pruning and scaling) of the target tracking kernel detection algorithm of traditional heavy models (e.g., OSTrack, SimTrack) relies on the CPU, which is computationally intensive and consumes CPU resources, resulting in an increase in overall inference latency.

[0062] Since the real-time requirements of edge devices cannot be met, in order to solve the preprocessing bottleneck, in step S2, this method migrates the image preprocessing task to the device's Intelligent Video Engine (IVE) for processing, and formulates an "IVE-NPU zero-copy pipeline". The specific working steps are as follows:

[0063] Step a: IVE hardware adaptation: Using the built-in image processing unit of IVE, the coordinate information of the image to be cropped is passed in. Relying on the cropping interface, bilinear interpolation and other hardware acceleration instructions provided by IVE, it supports the simultaneous parallel processing of template and search area image cropping and scaling, reducing CPU usage.

[0064] Step b: Memory sharing mechanism: During the IVE initialization phase, a physical address region overlapping with the NPU memory space is allocated to it. The two share the same DDR memory. After the IVE completes image scaling and cropping, it directly stores the data in the shared region without going through the CPU, thus reducing CPU usage.

[0065] Step c: Zero-copy pipeline scheduling: During model inference, IVE and NPU are synchronously triggered through direct memory access (DMA). IVE dynamically calculates the template and search area coordinates based on the target position information coordinates of the current frame, and writes the image after cropping and scaling to shared memory. When NPU reads the shared memory data, it directly obtains the result processed by IVE, completely eliminating the intermediate overhead of CPU participation in preprocessing.

[0066] Through the above methods, on HiSilicon's 3403 device, the preprocessing time was reduced from 26ms, which was traditionally CPU-driven, to 4ms, and the overall inference speed was increased to 38FPS, leaving room for subsequent post-processing algorithms.

[0067] In step S3, a "detection-tracking-state estimation collaborative" method was developed. To improve tracking stability, this method introduces the YOLOv8S target detection model as a relocalization module and combines it with Kalman filtering. The specific working steps are as follows:

[0068] Step 1: Set relocation trigger conditions: When the template region is searched, if the initial target initialization fails, target detection is used for relocation. When searching the region, a dual threshold judgment criterion is defined: relocation is triggered when the confidence score of the current frame output by the tracking model is less than the set threshold, or when the intersection-union ratio of the current frame target box and the previous frame target box is less than the set threshold. , The calculation formula is as follows:

[0069] ;

[0070] in, This represents the confidence score output for the current tracking frame. This indicates the currently set threshold. This represents the area of ​​the target bounding box in the previous frame. This represents the area of ​​the target bounding box in the previous frame. It is the intersection-union ratio function; Represents the intersection-union ratio function The set threshold;

[0071] when When the value is 1, the system initiates the relocalization process, calling the target detection model to perform target redetection within the search area expanded based on the previous frame's position. Simultaneously, as a protective measure, online template updates are paused at this stage to prevent incorrect target appearance features from being introduced into the template due to tracking errors or relocalization failures, thus causing continued tracking failures. When the value is 0, the system does not trigger.

[0072] Step II: Object Detection Relocalization: Relocalization is performed using the YOLOV8S object detection model. Specifically, the image input is reduced from 640x640 to 256x256 to achieve fast object detection in the current frame search area; the original YOLOV8S neck is removed, leaving only the backbone and head to improve the model's inference speed.

[0073] Step 3: During relocation, template updates are prohibited to avoid including incorrect detection boxes in the template library until the confidence level is greater than the set threshold, at which point template updates are resumed.

[0074] Step IV: Kalman filter state estimation involves defining the state vector. ,in, The coordinates of the target center are For velocity components, The target width and height are defined; the observation noise covariance matrix Q is dynamically adjusted according to the tracking confidence: when the confidence conf is less than the set threshold, the diagonal elements of Q are increased to reduce the impact of erroneous observations on state estimation; the smoothed target state is output by Kalman filtering to correct the jitter of the relocation results and eliminate the accidental offset of the YOLOv8s detection box.

[0075] In step S4, single-target tracking is affected by changes in target appearance, background interference, and small displacements in adjacent frames. The output bounding box confidence score is prone to high-frequency fluctuations due to the introduction of local noise, and the coordinates of adjacent tracking boxes exhibit jitter between frames. For example, changes in target pose cause pixel value shifts in the target box position. Existing technologies rely solely on Kalman filtering to handle jitter, failing to address the impact of confidence fluctuations on target selection. This method employs a two-stage post-processing approach—"Hanning window confidence suppression + EMA coordinate smoothing"—to synergistically improve the stability and reliability of the target tracking box. The specific steps are as follows:

[0076] Step A: The Hanning window suppresses the confidence level of the tracking box;

[0077] Record the confidence score and bounding box of the previous frame. When calculating the confidence score of the current frame, a two-dimensional Hanning window is superimposed for weighted suppression. The Hanning window is used for local sliding smoothing. The Hanning window suppression formula is as follows:

[0078] ;

[0079] in, Indicates the coordinates at the center of the target The weight value of the Hanning window at the location (ranging from 0 to 1); M and N represent the width and height of the window, respectively; The cosine term is used to generate the weight changes of the window shape (so that the weights at the window edges decay smoothly).

[0080] The weighted formula is as follows:

[0081] ;

[0082] in, For the final output of the first Confidence at each position (the fused result); Represents the target center coordinates of the current window position The confidence level at the point (the confidence level of the model's direct prediction);

[0083] This represents the window suppression factor, with a value range of [0,1]. Represents the coordinates of the target center The weight of the Hanning window at that location; It represents the multiplication operation, that is, multiplying by in mathematics;

[0084] when When the value is 0, the Hanning window is not used at all, and the final confidence level is equal to the original confidence level.

[0085] when When the value is 1, the confidence level is entirely determined by the Hanning window weights;

[0086] when When the value is 0.5, the Hanning window can be used, and the weight of the Hanning window can determine the confidence level.

[0087] By suppressing sidelobe noise in the confidence score and reducing score jitter caused by local texture interference, the highest confidence score output is the final target box score.

[0088] Step B: Smooth the tracking box using an EMA exponentially weighted average. This method defines a smoothing coefficient. When the confidence level is greater than the set threshold, the smoothing formula is as follows:

[0089] ;

[0090] in, This indicates the location information of the target bounding box in the previous frame. This represents the target bounding box output by the current tracking model. This provides the target bounding box position information for the current frame.

[0091] Using the modified tracking process detection method described above, the entire link of detection-tracking-filtering was optimized. The final target detection speed on this test platform was 8ms. Combined with the multi-threaded inference of the tracking model, and processed in parallel with the tracking algorithm which took 22ms, it did not affect the overall inference time and speed.

[0092] In addition, on the HiSilicon 3403 chip, the camera outputs a raw stream of 1920×1080@30fps. The IVE unit crops a 128×128 template image and a 256×256 search image, which are then sent to the NPU and SVP_NPU with zero copying. The improved AVTrack-INT8 model inference takes 22ms, image IVE cropping takes 4ms, and post-processing smoothing takes 1-2ms. The overall pipeline takes 28ms / frame, with a measured 35FPS.

[0093] This embodiment runs on the HiSilicon Hi3403 chip, which has a computing power of 10.4 TOPSINT8, integrates an ISP and two NPU units, supports the IVE intelligent engine module to scale and crop images, supports passive heat dissipation, and is suitable for use in edge scenarios such as drones.

[0094] Example 2;

[0095] Based on Example 1, such as Figure 3 As shown, this embodiment provides a single-target real-time tracking system for edge devices, including:

[0096] The model inference module is configured to run the AVTrack single-target lightweight tracking model and the optimized YOLOv8S target detection model, and perform INT8 precision quantization inference.

[0097] The preprocessing module is configured as an IVE integration, enabling template and search area cropping, scaling, and memory sharing through a zero-copy pipeline;

[0098] The collaborative control module is configured to execute collaborative logic for detection-tracking-state estimation, including relocation trigger judgment, template update control, and Kalman filter state estimation.

[0099] The post-processing module is configured to perform Hanning window confidence suppression and EMA exponential weighted average smoothing.

[0100] like Figure 4 As shown, this embodiment also provides an edge computing device, including:

[0101] At least one NPU computing unit is used to perform deep learning model inference;

[0102] The memory is used to store the target template, candidate box data, and image processing results in shared memory.

[0103] The CPU processor is used to trigger the re-detection thread, control template updates, and coordinate the scheduling of various modules.

[0104] The intelligent video engine performs hardware-accelerated processing for image cropping and scaling, and shares DDR memory with the NPU computing unit.

[0105] In addition, such as Figure 5 As shown, the edge computing device has a computer-readable storage medium, which is equivalent to a readable memory module. The storage medium stores a computer program, which, when executed by a processor, implements the single-target real-time tracking method of the edge device described in any of the above embodiments.

[0106] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope described in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A single-target real-time tracking method under edge devices, characterized in that, Comprising the following steps: Constructing an AVTrack single-target lightweight tracking model: replacing the two 16x16 convolutions on the input side of the original model with a 4x3x3 convolution stacking structure, and adding a residual structure branch with 1x1 convolution channel fusion, preserving the equivalent receptive field of single 16x16 convolution, performing INT8 precision quantization on the model, and reducing the inference time of edge devices; Executing the image preprocessing of the IVE-NPU zero-copy pipeline: migrating the cropping and scaling tasks of the template and search area to the IVE of the edge device, eliminating the CPU transit overhead through the zero-copy scheduling mode triggered by the shared DDR memory physical address and DMA synchronization of IVE and NPU; The execution steps of the IVE-NPU zero-copy pipeline include: IVE hardware adaptation: utilizing the built-in image processing unit and hardware acceleration instructions of IVE to parallel process the cropping and scaling of the template and search area; Memory sharing mechanism: IVE initializes and allocates the physical memory area overlapping with NPU, and directly stores the processed data to the shared area; Zero-copy pipeline scheduling: IVE and NPU are triggered by DMA synchronization, IVE dynamically calculates the coordinates and writes them to the shared memory, and NPU reads and processes the results; Executing the cooperative process of detection-tracking-state estimation: introducing the YOLOv8S target detection model as the repositioning module, setting a double-threshold trigger condition, triggering repositioning and pausing template updating when the tracking confidence is lower than the set threshold or the current frame and the previous frame target box IoU is lower than the set threshold, combining Kalman filter to dynamically adjust the observation noise covariance matrix, and outputting the smoothed target state; Adopting two-stage post-processing optimization: using an MxN size Hann window to suppress the tracking box confidence, reducing score jitter, and smoothing the tracking box coordinates based on EMA exponential weighted average to reduce adjacent frame jitter.

2. The single-target real-time tracking method under edge device according to claim 1, characterized in that, The parameter quantity calculation mode of the four 3x3 convolutions is C 3×3 =4x128x128x32, and four nonlinear activation functions are calculated to enhance the extraction of small target features.

3. The single-target real-time tracking method under edge device according to claim 1, characterized in that, The double-threshold trigger condition satisfies the following formula: ； wherein, represents triggering repositioning; represents a confidence score of a current tracking frame output; represents a threshold set currently; represents a previous current frame target frame area; represents a previous frame target frame area; is an intersection over union function; represents an intersection over union function a threshold set; When When the system starts the repositioning process, the target detection model is called to re-detect the target in the search area based on the position expansion of the previous frame. When System does not trigger when n is 0.

4. The single-target real-time tracking method under edge device according to claim 1, characterized in that, The optimization of the YOLOv8S target detection model includes: reducing the input resolution from 640x640 to 256x256 to realize fast target detection on the search area of the current frame; removing the neck structure, only keeping the backbone and detection head, and improving the model inference speed.

5. The single-target real-time tracking method under edge device according to claim 1, characterized in that, The state vector of the Kalman filter is wherein, is the target center coordinate, is the velocity component, is the target width-height; the diagonal elements of the observation noise covariance matrix Q are dynamically adjusted with the tracking confidence conf, and when the confidence conf is lower than a threshold, Q is increased to reduce the influence of false observations.

6. The single-target real-time tracking method under edge device according to claim 1, characterized in that, The way the Hann window suppresses the tracking box confidence is to record the confidence and target box of the previous frame, and when calculating the confidence of the current frame, a two-dimensional Hann window is superimposed for weighted suppression, and local sliding smoothing is performed using the Hann window.

7. A single-target real-time tracking system under edge device, configured to perform the single-target real-time tracking method under edge device according to any one of claims 1-6, characterized in that, Comprising: A model inference module configured to run the AVTrack single-target lightweight tracking model and the optimized YOLOv8S target detection model, and perform INT8 precision quantization inference; A preprocessing module configured as an IVE integration, realizing the cropping, scaling, and memory sharing of the template and search area through the zero-copy pipeline; A cooperative control module configured to execute the cooperative logic of detection-tracking-state estimation, including repositioning trigger judgment, template update control, and Kalman filter state estimation; A post-processing module configured to perform Hann window confidence suppression and EMA exponential weighted average smoothing processing.

8. An edge computing device configured to perform the single-target real-time tracking method under the edge device according to any one of claims 1-6. Comprising: At least one NPU computing unit for executing deep learning model inference; A memory for storing target templates, candidate box data, and image processing results in shared memory; A CPU processor for triggering re-detection threads, controlling template updates, and coordinating module scheduling; An intelligent video engine for performing hardware-accelerated image cropping and scaling, and sharing DDR memory with the NPU computing unit.

9. A computer-readable storage medium, characterized in that, The storage medium has a computer program stored thereon, and the computer program is executed by a processor to implement the single-target real-time tracking method under the edge device according to any one of claims 1-6.

Images