Lightweight target tracking method based on neural network architecture search and token clipping
By optimizing the transformer module structure through neural network architecture search and token pruning, the problem of high computational cost of transformer target tracking algorithm in edge computing environment is solved, and the model can achieve efficient tracking and dynamic adaptation on edge devices.
Patent Information
- Application Number
- CN202510962208.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-12-12
AI Technical Summary
Existing transformer-based target tracking algorithms have high computational costs in edge computing environments, and existing lightweight solutions rely on expert experience for design, lack dynamic adaptability, and result in high training overhead.
We employ a neural network architecture for searching and token pruning, and optimize the transformer module structure using a blacklist pruning strategy and a matching knowledge distillation method with token overlap as the indicator, thereby achieving a balance between model performance and computational efficiency.
Significantly reduce computational costs, improve model applicability on edge devices, enhance tracking accuracy and reduce training overhead, and enable dynamic adaptation of the model to different computing resource environments.
Smart Images
Figure CN121120686A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of visual target tracking technology, specifically to a lightweight target tracking method based on neural network architecture search and token pruning. Background Technology
[0002] Visual object tracking, a core research area in computer vision, aims to continuously and accurately locate specific targets within video sequences based on their initial states (such as position and scale). In recent years, object tracking algorithms based on convolutional neural networks (CNNs) have dominated this field due to their powerful feature extraction capabilities. However, with the breakthroughs in transformer architecture for computer vision tasks, the research paradigm is gradually shifting towards transformer-based trackers. These novel trackers, through their global modeling capabilities and long-range dependency capture characteristics, exhibit significant advantages in tracking accuracy and robustness, and are gradually becoming a hot research topic.
[0003] While transformer-based target tracking algorithms exhibit significant performance advantages, their high computational cost severely restricts their practical application in edge computing environments. Current mainstream lightweight solutions mainly fall into three technical paths: token pruning, CNN-Transformer hybrid architectures, and knowledge distillation. However, these methods suffer from two key limitations: first, they heavily rely on expert experience for meticulous manual design to achieve a speed-accuracy balance; second, their designs lack dynamic adaptability, often requiring retraining when the computational resources required by the deployment platform change, increasing training overhead. Neural Architecture Search (NAS) technology offers a new technical path to address these issues by automating the design of optimized architectures that conform to specific computational constraints. LightTrack pioneered the application of NAS to the design optimization of CNN-based trackers, but research on NAS for Transformer-based trackers remains largely unexplored. Summary of the Invention
[0004] In view of this, the present invention provides a lightweight target tracking method based on neural network architecture search and token pruning. By integrating pruning technology and neural network architecture search technology, it achieves multi-dimensional optimization of the Transformer module structure, thereby achieving the optimal trade-off between model performance and computational efficiency.
[0005] Furthermore, to effectively address the search space expansion problem caused by the token retention ratio, this invention proposes two innovative mechanisms: a blacklist pruning strategy based on prior knowledge, and a matching-based knowledge distillation method using token overlap (TOR) as an indicator. By optimizing the training strategy, the performance gap between each sub-model in the weight-sharing supernetwork and its corresponding independently trained model is reduced.
[0006] To solve the above-mentioned technical problems, the present invention is implemented as follows.
[0007] A lightweight target tracking method based on neural network architecture search and token pruning includes:
[0008] Step S1: Establish the hypernet search space: The structure of each transformer block in the tracking network OSTrack and the token retention rate, which is used as the basis for token pruning, are used as searchable options to form the hypernet search space; the token retention rate of each transformer block forms the token pruning scheme.
[0009] Step S2: Hypernetwork Training: Sample subnetworks from the hypernetwork search space as student networks. Based on the token overlap between the student network and the teacher network, match the teacher network from the elite pool for distillation learning. Dynamically update the elite pool by comparing the performance of the student network and the teacher network. Identify inappropriate token pruning schemes based on the performance of the student network and add them to the blacklist. These schemes will not be sampled again in subsequent training to reduce the hypernetwork search space.
[0010] Step S3: After completing the training of the super network, the best performing network in the elite pool is selected as the final lightweight tracker for target tracking.
[0011] Preferably, in step S1, the search range for the token retention rate is set to (1, 0.8, 0.7, 0.6).
[0012] Preferably, in step S2, the step of matching teacher networks from the elite pool for distillation learning based on the token overlap between the student network and the teacher network is as follows:
[0013] Based on the token pruning scheme, determine the remaining tokens in the last transformer block, calculate the importance score S of each remaining token, and filter elite tokens based on the importance score S.
[0014] The overlap rate of elite tokens between the two networks is used as the token overlap degree;
[0015] Determine the token overlap of each teacher network in the student network and the elite pool, and select the teacher network with the largest token overlap TOR as the matching teacher network for distillation learning.
[0016] Preferably, the step of selecting elite tokens based on importance score S is as follows: select k tokens with high importance based on importance score S as elite tokens; k is a set number.
[0017] Preferably, the token overlap ratio, which is the overlap rate of elite tokens between the two networks, is defined as follows:
[0018]
[0019] In the formula, TOR represents the token overlap, and P i P represents the set of elite token location indices of the student network. j This represents the set of elite token position indices for the j-th teacher network in the elite pool.
[0020] Preferably, in step S2, the dynamic updating of the elite pool is as follows:
[0021] By combining high network performance and low computational cost as comprehensive performance indicators, the student network is used to replace the teacher network in the elite pool, which has relatively weaker comprehensive performance.
[0022] Preferably, in step S2, adding improper token pruning schemes to the blacklist based on the performance of the student network is as follows:
[0023] Evaluate the validation IoU of the student network on the validation set; if the validation IoU is less than the set performance threshold T... IoU And the number of student network parameters is greater than or equal to the parameter number threshold T. Params If the token pruning scheme of the student network is deemed to significantly impair network performance, then the token pruning scheme will be added to the blacklist.
[0024] Preferably, in step S2, the step of adding improper token pruning schemes to the blacklist based on the performance of the student network includes:
[0025] First, obtain the training IoU of the student network; if the training IoU is less than a set performance threshold T... IoU And the number of student network parameters is less than the parameter number threshold T. Params If a token trimming scheme is found that may degrade network performance, it is considered a suspicious token trimming scheme.
[0026] Then, the validation IoU of the student network is evaluated on the validation set; if the validation IoU is less than the performance threshold T... IoU And the number of student network parameters is greater than or equal to the parameter number threshold T. Params If a suspicious token pruning scheme is deemed to significantly impair network performance, it will be added to the blacklist.
[0027] Preferably, in step S3, selecting the best-performing network in the elite pool as the final lightweight tracker means selecting the network with the largest verified IoU in the elite pool as the final lightweight tracker.
[0028] Preferably, in step S1, the token retention rate is set starting from the fourth layer transformer block, that is, the token pruning operation is implemented starting from the fourth layer.
[0029] Beneficial effects:
[0030] (1) The lightweight target tracking method based on neural architecture search and token pruning designed in this invention is the first efficient transformer target tracking framework based on neural architecture search (NAS). This framework integrates token pruning technology and effectively solves the key problem of model compression by jointly optimizing the topological connection relationship between transformer module layers and the hierarchical token retention rate.
[0031] (2) To address the problem of search space expansion caused by token retention rate, this invention optimizes the training strategy and proposes a blacklist pruning strategy based on prior knowledge. The blacklist mechanism is used to eliminate suboptimal pruning structures to compress the search space and reduce the performance gap between each sub-model and the corresponding independent training model in the weight-sharing supernetwork.
[0032] (3) In a preferred scheme, when determining the token pruning scheme to be included in the blacklist, the number of parameters (Params) and IoU are used as joint evaluation indicators. Compared with computational cost (FLOPs), Params are not affected by changes in the number of input tokens and can effectively reflect the network architecture performance, solving the coupling effect problem between network architecture and pruning strategy, and ensuring that the IoU is verified only when the number of network parameters is within a reasonable range. val The corresponding token pruning scheme will only be discarded when the value remains low. This strategy reduces the bias introduced by changes in network architecture and avoids the accidental deletion of potentially high-performance pruning structures.
[0033] (4) Similarly addressing the search space expansion problem introduced by token retention rate, this invention further proposes a matching-based knowledge distillation method using token overlap (TOR) as the metric. Employing a TOR-based teacher-student model matching criterion, it significantly accelerates network convergence. Extensive experiments demonstrate that this algorithm effectively improves the training quality of supernetworks, narrows the performance gap between each sub-model in a weight-sharing supernetwork and its corresponding independently trained model, and achieves optimal performance in multiple benchmark tests, significantly surpassing existing lightweight trackers.
[0034] (5) Considering the limitations of the feature representation ability of shallow networks, the pruning operation of the present invention is implemented from the fourth layer. This design has two advantages: 1) It significantly reduces the risk of accidental deletion due to inaccurate token importance scoring; 2) It effectively protects the key feature extraction ability of shallow networks.
[0035] (6) In a preferred embodiment, the search range of the token retention rate is set to (1, 0.8, 0.7, 0.6). The provided retention rate ratio constructs a search space from an unpruned baseline to aggressive pruning, providing sufficient flexibility for different application scenarios and performance requirements, so that it can find the best balance between model accuracy and computational efficiency. Attached Figure Description
[0036] Figure 1 This is a schematic diagram of the overall training framework and search space of the supernetwork of the present invention.
[0037] Figure 2 This is a sample diagram of the sub-network architecture of the present invention. Detailed Implementation
[0038] This invention provides a lightweight target tracking method based on neural network architecture search and token pruning. The basic idea is to add a token retention rate dimension to the hypernetwork search space of the tracking network OSTrack during hypernetwork training. The token retention rates of all transformer blocks in OSTrack constitute the token pruning scheme. This invention incorporates the token retention rate into the search space and uses Neural Architecture Search (NAS) to automatically find the optimal token retention rate for each layer, making the remaining tokens obtained through token pruning more relevant to target recognition. The addition of the token retention rate makes the NAS scheme more likely to find better networks, but it can lead to an excessively large search range. Therefore, to suppress this excessive expansion, this invention further adds a blacklist mechanism, adding token pruning schemes that significantly degrade network performance to the blacklist and removing them from subsequent training to reduce the hypernetwork search space. Simultaneously, computationally inexpensive and high-performance networks are added to an elite pool as candidate teacher networks. During hypernetwork training, suitable teachers are selected from the elite pool based on token overlap (TOR) to perform distillation learning on student models. After completing the training of the super network, the best-performing network in the elite pool is selected as the final lightweight tracker for target tracking.
[0039] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0040] Figure 1 The diagram shows the principle of the lightweight target tracking method based on neural network architecture search and token pruning of this invention. As shown in the figure, the method includes the following steps:
[0041] Step S1: Establish the hypernetic search space.
[0042] This invention uses the OSTrack tracking network as the basic structure of the tracking network. OSTrack includes multiple layers of transformer blocks. This invention uses the structure of each transformer block in the OSTrack tracking network and the token retention rate as searchable options to form a super-network search space.
[0043] The token retention rate represents the proportion of tokens retained in each transformer block. A token retention rate is set for each transformer block. The token retention rates of all transformer blocks constitute the token pruning scheme. The token retention rate for each layer is selected from a given range of token retention rates.
[0044] The tokens in the original input image are cropped according to the token retention rate set for each layer. After multiple layers of processing, a portion of the tokens are ultimately retained, such as... Figure 1 The remaining tokens in the upper middle section. This invention incorporates the token retention rate into the search space and uses Neural Architecture Search (NAS) to automatically find the optimal token retention rate for each layer, making the remaining tokens obtained through pruning more relevant to target recognition.
[0045] Preferably, the token retention rate search range is (1, 0.8, 0.7, 0.6). This design ensures that there are a large number of network architectures in the search space that balance model performance with computational efficiency.
[0046] Considering the limitations of shallow networks in feature representation, the pruning operation of this invention is implemented starting from the fourth layer. This design has two advantages: 1) it significantly reduces the risk of accidental deletion due to inaccurate token importance scoring; 2) it effectively protects the key feature extraction capabilities of shallow networks.
[0047] The structural parameters of each searchable transformer block can include the number of layers, the embedding layer dimension, the QKV dimension (QKV Dim), the number of self-attention heads, and the feed forward amplification factor. The search range of transformer block structural parameters and token retention rate together constitutes the hypernetwork search space. Figure 2 A diagram illustrating the searchable options is shown.
[0048] In practice, to accommodate networks with different numbers of parameters, corresponding hypernetwork search spaces can be designed based on the varying sizes of network parameters. Table 1 shows three hypernetwork search spaces designed for three parameter ranges to meet different computational resource constraints.
[0049] Table 1
[0050] <![CDATA[Hypernetwork 3,4 > <![CDATA[Hypernetwork 6,8 > <![CDATA[Hypernetwork 10,12 > number of floors (3,4) (6,8) (10,12) Embedding layer dimension (640,704,768) (640,704,768) (640,704,768) QKV dimension (640,704,768) (640,704,768) (640,704,768) Self-attention head count (10,11,12) (10,11,12) (10,11,12) Forward propagation layer magnification (3.0,3.5,4.0) (3.0,3.5,4.0) (3.0,3.5,4.0) Token retention rate (1,0.8,0.7,0.6) (1,0.8,0.7,0.6) (1,0.8,0.7,0.6) Network parameter range 18-35M 30-63M 47-92M
[0051] Step S2: Hypernetwork training.
[0052] Hypernetwork training refers to sampling a subnetwork from the hypernetwork search space as a student network, selecting a matching teacher network for one round of distillation learning; then selecting a new subnetwork and its teacher network for distillation learning again. During the distillation learning process, the training IoU is calculated based on the support set, denoted as IoU. train It will also calculate the verification IoU based on the verification set, denoted as IoU. val .
[0053] In each round of distillation learning, after sampling the student network, the present invention improves two operations: (1) a blacklist-based token pruning scheme is added to reduce the hypernet search space;
[0054] (2) Improved distillation scheme by designing knowledge distillation optimization based on dynamic elite pool and token overlap rate (TOR).
[0055] (1) A blacklist-based token pruning scheme
[0056] To address the training efficiency issues that may arise from incorporating token retention rate into the search space, this invention proposes a progressive search space optimization method. Based on performance evaluation during supernetwork training, token pruning schemes that degrade network performance due to token retention rate selection are identified and added to a blacklist, and are no longer sampled in subsequent training. This gradually eliminates invalid pruning structures, and by finely adjusting (reducing) the search space, the performance of the NAS method is significantly improved.
[0057] One specific implementation of the blacklist mechanism is as follows: A blacklist policy is set up, and for student networks, the validation IoU (Inter-Origin Value) is evaluated using a validation set sample. val If IoU val Less than the set performance threshold T IoU And the number of student network parameters Params(α) is greater than or equal to the parameter number threshold T. Params If the token pruning scheme of the student network is deemed to significantly impair network performance, then the token pruning scheme will be added to the blacklist.
[0058] blacklist The definition is as follows:
[0059]
[0060] in, This represents the token pruning scheme, where α represents the student network. Describing the student network α Network parameter values.
[0061] This scheme introduces the number of parameters (Params) and IoU as a joint evaluation metric. Compared to computational cost (FLOPs), Params are unaffected by changes in the number of input tokens and effectively reflect network architecture performance. This solves the coupling problem between network architecture and pruning strategies, ensuring that IoU only increases when the number of network parameters is within a reasonable range. val The corresponding token pruning scheme will only be discarded if the level remains low. This strategy reduces biases introduced by changes in network architecture and avoids the accidental deletion of potentially high-performance pruning structures.
[0062] Another specific implementation of the blacklist mechanism is: because of the IoU of the student network, i.e., IoU train Since the data is already available, no further calculations based on the support set are needed. Therefore, we first use the training IoU in conjunction with the number of student network parameters to make a judgment. If the training IoU (IoU) of the student network α is... train (Below the preset performance threshold T) IoU And the number of parameters is less than the threshold T. Params If a token pruning scheme that might be causing network performance degradation is identified, the current pruning configuration is recorded. This is a suspected token pruning scheme. In this case, the network performance degradation based on the training IoU might be due to a small number of parameters or pruning configuration issues, but this is inaccurate. It is necessary to combine the verification IoU to further evaluate the suspected token pruning scheme, in order to rule out performance degradation caused by a small number of parameters, and thus filter out performance degradation caused by token pruning configuration issues.
[0063] Then, evaluate the validation IoU (IoU) of the student network on the validation set. val If IoU val Still below the performance threshold T IoU And the number of student network parameters is greater than or equal to the parameter number threshold T. Params If a suspicious token pruning scheme is found, it will be added to the blacklist and will not be sampled again in subsequent training.
[0064] This approach of training IoU first and then validating IoU eliminates the need to evaluate and verify IoU using a validation set for each student network. Instead, it uses existing data to make initial judgments on suspicious cases, thereby reducing the computational load.
[0065] (2) Knowledge distillation optimization based on dynamic elite pool and token overlap ratio (TOR)
[0066] To accelerate the convergence speed of the subnetwork, this invention provides a knowledge distillation optimization method based on dynamic elite pool and token overlap ratio (TOR), the specific implementation process of which is as follows:
[0067] ● Regarding the construction of a dynamic elite pool:
[0068] During the training of supernetworks, high network performance (IoU) is combined. val The overall performance metrics consist of low computational cost (FLOPs), and the student network is used to replace the teacher network in the elite pool, which has a relatively weaker overall performance.
[0069] The elite pool ε can be dynamically constructed and updated using the following formula:
[0070]
[0071] In the formula, ε represents the teacher network in the elite pool; α represents the student network; This represents the verification IoU of the student network α. Indicating teacher network The verification IoU. Flops represents the computational cost of the network; & represents the AND relationship.
[0072] The above formula ensures that the networks in the elite pool simultaneously meet the high model performance (IoU). val The dual requirements of low computational cost (FLOPs) and low computational cost (FLOPs).
[0073] ● For distillation learning of teacher networks based on token overlap ratio (TOR) matching:
[0074] Based on the token overlap (TOR) between the student and teacher networks, the network with the highest TOR is selected from the elite pool as the matching teacher network for distillation learning. Specifically:
[0075] Based on the token pruning scheme, the remaining tokens in the last transformer block are determined, and the importance score S of each remaining token is calculated. Elite tokens are then selected based on the importance score S. Here, the token importance score S can be calculated using the attention score between the template and the search area.
[0076] When the number of tokens retained between two networks differs significantly, TOR becomes unreliable as a measure of token similarity. For example, even if one network's retained tokens are a subset of another network's, the TOR may still decrease due to the large difference in token numbers. To address this issue, this method selects the k most important tokens (S) from the remaining tokens as elite tokens and calculates the TOR based on the indices of these elite tokens, thus providing a more accurate measure of similarity. This design ensures that the number of tokens compared is consistent and focuses on the most important token features.
[0077] Let P i P represents the set of elite token location indices of the student network. j Let represent the set of elite token position indices for the j-th teacher network in the elite pool; then, to better match suitable teacher networks, this invention proposes the token overlap ratio (TOR) metric as follows:
[0078]
[0079] The specific implementation process for training the hypernet in this step is as follows:
[0080] 1) The sandwich rule was used to complete the initial training of the supernet;
[0081] 2) Build an elite pool;
[0082] 3) For the randomly sampled student network, calculate its TOR value relative to each network in the elite pool;
[0083] 4) Select the network with the highest TOR value as the teacher for knowledge distillation, and update the blacklist and elite pool. Repeat steps 3) and 4) until the supernetwork is trained.
[0084] This method significantly improves knowledge transfer efficiency through precise semantic feature matching. It also enhances the quality of the supernetwork by accelerating the convergence speed of the subnetworks.
[0085] Step S3: Analyze the performance (IoU) of the elite pool. val The strongest network is used as the final lightweight tracker to perform target tracking.
[0086] This concludes the process.
[0087] In a specific experiment, the designed neural network search algorithm used OSTrack as the baseline network. It was trained on the training set for 300 rounds with a learning rate of 0.0004. Starting from the 240th round, the learning rate decayed to 0.00004.
[0088] Table 2 shows the performance of the method of this invention (highlighted in gray) and the state-of-the-art (SOTA) method on the TrackingNet, LaSOT, UAV123, and TNL2K datasets. The best results are highlighted in bold with an underline.
[0089] Table 2
[0090]
[0091]
[0092] Table 2 compares the performance of the trackers discovered by the proposed method with the best current tracking methods on the LaSOT, TrackingNet, and UAV123 datasets. The proposed method outperforms existing real-time trackers in all performance metrics across these datasets. In terms of speed, the proposed trackers achieve a significant improvement over the baseline method OSTrack without sacrificing accuracy. Specifically, the proposed NASTrack-B achieves an AUC of 68.7% on the UAV123 dataset, 0.4% higher than the baseline OSTrack, while also achieving a 1.5x speedup. On the Nvidia Jetson Orin NX 16GB platform, all tracker variants of the proposed method achieve real-time performance even without TensorRT acceleration, demonstrating their suitability for deployment on edge devices.
[0093] The specific embodiments described above only illustrate the design principles of the present invention. The shapes and names of the components in this description may differ and are not limited. Therefore, those skilled in the art can modify or make equivalent substitutions to the technical solutions described in the foregoing embodiments; and such modifications and substitutions do not depart from the inventive spirit and technical solutions of the present invention, and should all fall within the protection scope of the present invention.
Claims
1. A lightweight target tracking method based on neural network architecture search and token pruning, characterized in that, include: Step S1: Establish the hypernet search space: The structure of each transformer block in the tracking network OSTrack and the token retention rate, which is used as the basis for token pruning, are used as searchable options to form the hypernet search space; the token retention rate of each transformer block forms the token pruning scheme. Step S2: Hypernetwork Training: Sample subnetworks from the hypernetwork search space as student networks. Based on the token overlap between the student network and the teacher network, match the teacher network from the elite pool for distillation learning. Dynamically update the elite pool by comparing the performance of the student network and the teacher network. Identify inappropriate token pruning schemes based on the performance of the student network and add them to the blacklist. These schemes will not be sampled again in subsequent training to reduce the hypernetwork search space. Step S3: After completing the training of the super network, the best performing network in the elite pool is selected as the final lightweight tracker for target tracking.
2. The method as described in claim 1, characterized in that, In step S1, the search range for the token retention rate is set to (1, 0.8, 0.7, 0.6).
3. The method as described in claim 1, characterized in that, In step S2, the step of matching teacher networks from the elite pool for distillation learning based on the token overlap between the student network and the teacher network is as follows: Based on the token pruning scheme, determine the remaining tokens in the last transformer block, calculate the importance score S of each remaining token, and filter elite tokens based on the importance score S. The overlap rate of elite tokens between the two networks is used as the token overlap degree; Determine the token overlap of each teacher network in the student network and the elite pool, and select the teacher network with the largest token overlap TOR as the matching teacher network for distillation learning.
4. The method as described in claim 3, characterized in that, The process of selecting elite tokens based on importance score S is as follows: select k tokens with high importance based on importance score S as elite tokens; k is a set number. The overlap rate of elite tokens between the two networks is used as the token overlap degree: In the formula, TOR represents the token overlap, and P i P represents the set of elite token location indices of the student network. j This represents the set of elite token position indices for the j-th teacher network in the elite pool.
5. The method as described in claim 1, characterized in that, In step S2, the dynamic update of the elite pool is as follows: By combining high network performance and low computational cost as comprehensive performance indicators, the student network is used to replace the teacher network in the elite pool, which has relatively weaker comprehensive performance.
6. The method as described in claim 1, characterized in that, In step S2, based on the performance of the student network, the improper token pruning scheme is identified and added to the blacklist as follows: Evaluate the validation IoU of the student network on the validation set; if the validation IoU is less than the set performance threshold T... IoU And the number of student network parameters is greater than or equal to the parameter number threshold T. Params If the token pruning scheme of the student network is deemed to significantly impair network performance, then the token pruning scheme will be added to the blacklist.
7. The method as described in claim 1, characterized in that, In step S2, the step of adding improper token pruning schemes to the blacklist based on the performance of the student network includes: First, obtain the training IoU of the student network; if the training IoU is less than a set performance threshold T... IoU And the number of student network parameters is less than the parameter number threshold T. Params If a token trimming scheme is found that may degrade network performance, it is considered a suspicious token trimming scheme. Then, the validation IoU of the student network is evaluated on the validation set; if the validation IoU is less than the performance threshold T... IoU And the number of student network parameters is greater than or equal to the parameter number threshold T. Params If a suspicious token pruning scheme is deemed to significantly impair network performance, it will be added to the blacklist.
8. The method as described in claim 1, characterized in that, In step S3, the step of selecting the best-performing network in the elite pool as the final lightweight tracker is: selecting the network with the largest verified IoU in the elite pool as the final lightweight tracker.
9. The method as described in claim 1, characterized in that, In step S1, the token retention rate is set starting from the fourth-level transformer block, meaning the token pruning operation is implemented starting from the fourth level.