Visual target tracking method and system based on dynamic hyperbolic tangent normalization
By introducing a dynamic hyperbolic tangent normalization visual target tracking method on the UAV platform to replace the traditional layer normalization layer, and combining self-attention and cross-attention mechanisms, the problem of heavy computational burden in UAV target tracking is solved, and efficient and lightweight real-time tracking effect is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2025-12-18
- Publication Date
- 2026-04-24
AI Technical Summary
Unmanned aerial vehicle (UAV) target tracking technology faces challenges such as heavy computational burden and high computational complexity, especially on UAV platforms where efficient and lightweight real-time tracking is difficult to achieve.
A visual target tracking method based on dynamic hyperbolic tangent normalization is adopted. By introducing a dynamic hyperbolic tangent (DyT) module to replace the traditional layer normalization layer, a novel Transformer structure is constructed. Combined with self-attention and cross-attention mechanisms, feature enhancement and interaction are performed to reduce the computational burden.
While maintaining target tracking accuracy, it significantly reduces computational complexity, improves applicability and real-time response capabilities on UAV platforms, and enhances the model's generalization ability and robustness.
Smart Images
Figure CN121921487A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of target tracking technology, and in particular relates to a visual target tracking method and system based on dynamic hyperbolic tangent normalization. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] The main challenges currently facing UAV target tracking technology include: the limited computing power of UAV onboard computing units, requiring tracking algorithms to be efficient and lightweight; drastic changes in target scale, with rapid ascents, descents, and movements causing rapid and dramatic changes in target scale; complex aerial backgrounds, including similar objects and fast-moving background interference; and the need for real-time feedback for flight control decisions, requiring tracking algorithms to have a high frame rate.
[0004] For drone target tracking, trackers employing a Siamese network structure are currently the mainstream approach, offering a good balance between tracking performance and inference speed. With the development of deep learning, especially the introduction of Transformers into target tracking, the ability to model long-term dependencies and global contextual information has been enhanced. This allows for efficient computation of relationships between pixels in image sequences, enabling the model to maintain robust tracking performance even in complex scenarios.
[0005] Existing Transformer-based trackers have achieved breakthroughs in accuracy by effectively fusing template and search region features through attention mechanisms. However, their core modules heavily rely on layer normalization operations. Because they must dynamically calculate the mean and variance of features during each forward propagation, this leads to significant computational redundancy and frequent memory read / write operations. This not only increases the overall computational burden of the model but also complicates the computation graph, making efficient optimization difficult on dedicated edge computing chips for UAVs. Therefore, the LN (layer normalization) layer becomes a key factor limiting model inference speed and affecting the real-time response and endurance of UAV tracking systems. Summary of the Invention
[0006] To address the technical problems mentioned above, this invention provides a visual target tracking method and system based on dynamic hyperbolic tangent normalization. In the feature fusion stage, it introduces feature enhancement and interaction based on an attention mechanism and introduces dynamic hyperbolic tangent (DyT) to construct a novel Transformer structure. This significantly reduces the computational burden while maintaining target tracking accuracy and improves the applicability of the model on UAV platforms.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a visual target tracking method based on dynamic hyperbolic tangent normalization, comprising: Acquire template and search images of the target being tracked; For the template image and the search image, feature extraction networks are used to extract features, and after convolution and unrolling, the template feature vector and the search feature vector are obtained. For the template feature vector and the search feature vector, several feature enhancement and fusion processes are performed to obtain the template features and the tracking features. In each feature enhancement and fusion process, two encoders and two decoders are used. The two encoders process the template image and the search image respectively, and the two decoders process the outputs of the two encoders simultaneously. Both the encoder and the decoder use dynamic hyperbolic tangent to achieve layer normalization. Template features and tracking features are fused by the decoder and then input into the prediction head network for target localization to obtain the bounding box of the tracked target.
[0008] Furthermore, after the encoder performs sinusoidal function position learning encoding on the feature vector, it expands and overlays the encoded feature map to obtain the query value, key value, and evaluation value. The query value, key value, and evaluation value are then processed by a multi-head attention function, followed by residual connection, and normalization is achieved using dynamic hyperbolic tangent.
[0009] Furthermore, after the decoder performs sinusoidal function position learning encoding on the two input vectors respectively, it expands and channels the encoded feature maps to obtain query values, key values and evaluation values. After fusing the query values, key values and evaluation values of the two input vectors through a residual multi-head cross-attention mechanism, it normalizes them through dynamic hyperbolic tangent and performs residual connections.
[0010] Furthermore, the dynamic hyperbolic tangent implementation layer normalization is expressed by the formula: ; in, It is a learnable scalar parameter. and These are learnable channel vector parameters, where x represents the input vector and DyT represents the dynamic hyperbolic tangent.
[0011] A second aspect of the present invention provides a visual target tracking system based on dynamic hyperbolic tangent normalization, comprising: The image acquisition module is configured to acquire template images and search images of the target being tracked. The feature extraction module is configured to extract features from the template image and the search image using a feature extraction network, and then obtain the template feature vector and the search feature vector after convolution and unrolling. The feature enhancement module is configured to perform several feature enhancement and fusion processes on the template feature vector and the search feature vector to obtain template features and tracking features. In each feature enhancement and fusion process, two encoders and two decoders are used. The two encoders process the template image and the search image respectively, and the two decoders process the outputs of the two encoders simultaneously. Both the encoder and the decoder use dynamic hyperbolic tangent to achieve layer normalization. The target tracking module is configured to fuse template features and tracking features through a decoder, and then input them into the prediction head network for target localization to obtain the bounding box of the tracked target.
[0012] Furthermore, after the encoder performs sinusoidal function position learning encoding on the feature vector, it expands and overlays the encoded feature map to obtain the query value, key value, and evaluation value. The query value, key value, and evaluation value are then processed by a multi-head attention function, followed by residual connection, and normalization is achieved using dynamic hyperbolic tangent.
[0013] Furthermore, after the decoder performs sinusoidal function position learning encoding on the two input vectors respectively, it expands and channels the encoded feature maps to obtain query values, key values and evaluation values. After fusing the query values, key values and evaluation values of the two input vectors through a residual multi-head cross-attention mechanism, it normalizes them through dynamic hyperbolic tangent and performs residual connections.
[0014] Furthermore, the dynamic hyperbolic tangent implementation layer normalization is expressed by the formula: ; in, It is a learnable scalar parameter. and These are learnable channel vector parameters, where x represents the input vector and DyT represents the dynamic hyperbolic tangent.
[0015] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a visual target tracking method based on dynamic hyperbolic tangent normalization as described above.
[0016] A fourth aspect of the present invention provides a computer device including a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, wherein the processor executes the program to implement the steps of a visual target tracking method based on dynamic hyperbolic tangent normalization as described above.
[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention introduces attention-based feature enhancement and interaction in the feature fusion stage, and introduces dynamic hyperbolic tangent to construct a novel Transformer structure, thereby significantly reducing the computational burden while maintaining target tracking accuracy and improving the applicability of the model on UAV platforms. Attached Figure Description
[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0019] Figure 1 This is a flowchart of a visual target tracking method based on dynamic hyperbolic tangent normalization according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the structure of the improved Encoder module according to Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the structure of the improved Decoder module according to Embodiment 1 of the present invention; Figure 4 This is a schematic diagram comparing the input-output curves of the DyT operation in Embodiment 1 of the present invention with those of the traditional normalization operation; Figure 5 This is a comparison curve of the loss function during the training process of the normalization operation and the DyT operation in Embodiment 1 of the present invention. Figure 6 This is a schematic diagram of the tracking effect in a typical UAV tracking scenario according to Embodiment 1 of the present invention; Figure 7 This is a schematic diagram of the structure of a computer device according to Embodiment 4 of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.
[0021] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0022] Example 1 This embodiment provides a visual target tracking method based on dynamic hyperbolic tangent normalization.
[0023] Dynamic hyperbolic tangent (DyT) has recently been proposed as a statistically insensitive normalization alternative. Its core idea is to use a learnable scaling parameter.α The sigmoid saturation characteristic of the hyperbolic tangent function is used to simulate the normalization and nonlinear transformation effects of traditional normalization layers on features. Specifically, the DyT operation is defined as: , These are learnable scalar parameters used to dynamically adjust the scale of the input feature x. and These are channel-level affine parameters used to recover the expressive power of features. Unlike LN, DyT does not rely on the calculation of statistics on input features; it achieves feature normalization only through element-level operations, thereby significantly reducing computational complexity while maintaining the model's expressive power.
[0024] Recent studies have shown that DyT can achieve performance comparable to or even better than LN in various tasks such as vision, language, and speech, and the training process is more stable, the convergence speed is faster, and there is basically no need to adjust the original hyperparameters.
[0025] However, to date, no research has introduced DyT into computer vision target tracking, a specific field that demands both high efficiency and accuracy, especially in achieving real-time high-performance tracking on UAV platforms.
[0026] This embodiment provides a visual target tracking method based on dynamic hyperbolic tangent normalization. Based on the Transformer mechanism architecture and dynamic hyperbolic tangent normalization operation, it is suitable for real-time single-target tracking on UAV platforms.
[0027] To address the dual requirements of high accuracy and low computational complexity in UAV target tracking, this embodiment provides a visual target tracking method based on dynamic hyperbolic tangent normalization. In the feature extraction stage, ResNet (residual neural network) is used as the backbone network. In the feature fusion stage, a feature enhancement and interaction module based on an attention mechanism is introduced, along with a DyT module, to construct a novel Transformer structure. This significantly reduces the computational burden while maintaining tracking accuracy, thereby improving the model's applicability on UAV platforms.
[0028] This embodiment provides a visual target tracking method based on dynamic hyperbolic tangent normalization. It uses dynamic hyperbolic tangent instead of traditional layer normalization to construct a novel Transformer architecture for target tracking, thereby making the entire target tracking process lightweight.
[0029] This embodiment provides a visual target tracking method based on dynamic hyperbolic tangent normalization. It adopts a high-speed feature extraction network to improve the extraction speed while maintaining feature representation ability; it introduces an attention-based feature fusion mechanism to enhance the model's utilization of global context information; and it replaces the standard normalization layer with a DyT module to reduce computational redundancy and improve operating efficiency.
[0030] This embodiment provides a visual target tracking method based on dynamic hyperbolic tangent normalization, such as... Figure 1 As shown, it includes the following steps: Step 1: Set the initial parameters of the monocular camera and LiDAR, load the pre-trained model required for the target tracking task, and continuously monitor the unlock command of the UAV. Once the unlock is successful, the UAV will autonomously start the target tracking task.
[0031] The specific steps are as follows: initialize the drone to start the mission; manually select the target bounding box to be tracked; the model extracts template features of the selected target bounding box; the drone camera acquires real-time tracking images; and features are extracted from the acquired image information.
[0032] Step 2: First, preprocess the template image (the first frame of the target image) and the search image (each subsequent frame of the target image); then use the ResNet50 feature extraction network to extract features from the search image (3×128×128) and the template image (3×256×256) respectively, obtain the feature maps of their corresponding scales, and expand them into template feature vectors and tracking feature vectors (256×256, 1024×256) respectively through 1×1 convolution.
[0033] In step 2, since the ResNet50 network is too redundant, this embodiment uses ResNet50 as the backbone network, removes the last stage, adjusts the stride, and uses dilated convolutions to maintain high resolution, thereby improving the speed of the model's feature extraction stage. Specifically: the standard ResNet50 contains an initial convolutional layer (stage0) and four main stages (stage1 to stage4) consisting of residual blocks. This embodiment removes the last residual block, stage4, and only retains and utilizes the residual blocks from stage0 to stage3 for feature extraction. To compensate for the loss of receptive field and maintain the high resolution of the feature map, the convolution stride of the downsampled residual block in stage3 is adjusted from 2 to 1. At the same time, dilated convolutions are used in this stage to maintain high resolution and expand the receptive field, thereby improving the feature extraction speed while ensuring feature representation capability.
[0034] Therefore, using the Transformer architecture with integrated Dynamic Hyperbolic Tangent (DyT) module, feature enhancement and feature fusion are performed on the template features and search features extracted in step 2 through its self-attention encoder and cross-attention decoder.
[0035] Step 3: Input the obtained template feature vector and tracking feature vector into the improved Encoder to enhance the contextual information within the input features. This process uses positional encoding to preserve spatial information, which can enhance the feature representation of a single branch.
[0036] In step 3, in order to enhance the contextual information within a single branch (template or search region) so that the model can extract more general and important feature information, a self-attention mechanism is used to input the extracted single-branch vector into the improved Encoder.
[0037] like Figure 2 The image shows the position where DyT replaces LN, as detailed below: Step 301: Perform sinusoidal function position learning encoding on the input feature vector X to obtain spatial position encoding.
[0038] Step 302: Based on the spatial location encoding and the input feature vector, perform a Flatten operation and channel overlay to obtain the query value Q, key value K, and evaluation value V.
[0039] Step 303: Pass Q, K, and V through a multi-head attention function MultiHead (Q,K,V) and residual connections are used to obtain output features with contextual feature enhancement and global dependency relationships. This allows for better capture of correlations within sequences; the calculation formula related to the multi-head attention function is: ; ;in, This represents a vector concatenation operation; This represents the i-th attention module; nh is the total number of heads; It is a parameter matrix; This represents the weight of each head in the attention module; , , These are parameter matrices representing the query value, key value, and value item, respectively. This represents a single attention operation. ; The dimension of the key value is represented; residual connections are used to directly add input data to the network layer output, thereby effectively conveying information. This structure can alleviate problems such as gradient vanishing and network degradation, and improve model performance.
[0040] Step 304: The feature vector X obtained in step 303 through multi-head self-attention and residual connection EA The input vector X is normalized by the DyT module: The DyT module normalizes the input vector X. EA The nonlinear transformation and normalization are performed, and the output is denoted as vector X.EC =DyT(X EA ), X EC It is the enhanced feature vector obtained from the input.
[0041] The DyT module is used to replace the layer normalization operation, and its definition is as follows: ; in, These are learnable scalar parameters, which are the core of learnable dynamic modules; and These are learnable channel vector parameters, the same as those used in all normalization layers, which enable the output to scale back to arbitrary scales; regarding the scaling parameters, this embodiment always uses... Initialize as a vector of all 1s. Initialize as a vector of all zeros.
[0042] As one implementation method, for scaling factors The default initial value is 0.5.
[0043] As another implementation method, based on the linear interval properties of the tanh function, an adaptive scaling factor initialization formula is designed: Where: hyperparameter k=r(log2(d) / log2(d0)), d is the dimension of x, d0 is the baseline dimension, and r is the baseline correction coefficient. Different scaling is applied based on the input range to adapt to the variable scale of x; tanh in | When |<1, it is approximately linear, through Scaling This can offset the influence of the feature dimension d on the input x scale; by using the logarithmic ratio of the current layer dimension to the baseline dimension, k is adaptively adjusted, with k increasing accordingly as the dimension increases, thus avoiding insufficient scaling; because The initial values already place tanh within the linear operating region, resulting in smoother and more efficient gradients (tanh's gradient is close to 1 in the linear region and close to 0 in the oversaturation region). Parameter updates are more stable in the early stages of training, allowing for rapid convergence without additional adjustments to the initial values. During training... It remains a learnable scalar parameter, and the scaling degree can be adjusted autonomously according to changes in data distribution. This ensures both the rationality of initialization and retains the core advantage of DyT's dynamic adaptation to the input scale.
[0044] like Figure 4 As shown, the input-output curves of DyT with different scaling degrees and the traditional normalization operation of a representative ViT model are compared, demonstrating that the effect of layer normalization can be achieved by learning and adjusting the parameters during training.
[0045] Step 4: Combine the obtained template features and tracking features, and perform feature fusion twice using two improved decoders. That is, the decoder input includes features from two branches, and the output is the fused features. Position encoding and FFN (feedforward neural network) are also used to enhance the representation ability. The fused features output from the two improved decoders are then fused again through an improved decoder.
[0046] In step 4, in order to fuse the features of the template and the search region and achieve global semantic alignment to improve the robustness of the model, the two extracted vectors are input into the improved Decoder using a cross-attention mechanism.
[0047] like Figure 3 The image shows the position where DyT replaces LN, as detailed below: Step 401: For the current branch feature X q and another branch feature X KV Similar to the Encoder, the Decoder also employs a spatial location encoding mechanism.
[0048] Step 402: Perform a Flatten operation on the encoded feature map and stack the channels to obtain the key value Q (query), K (key), and evaluation value V (value).
[0049] Step 403: Fuse the feature vectors from the two inputs using a residual-based multi-head cross-attention mechanism. The corresponding formula is: ; in, It is a feature from the current branch. It is a feature from another branch. yes With the corresponding spatial location code, yes With the corresponding spatial location code, the Decoder calculates... and The attention map is generated by the multi-scale product, and then the weights are recalculated based on the map. and overlay the results to To enhance the representational power of feature maps.
[0050] Step 404: Input the feature vector fused by cross-attention into the DyT module for normalization to obtain the output vector; then, pass the DyT output vector through a multilayer perceptron and perform a residual connection with the DyT output vector. This residual connection helps stabilize training and promote gradient flow; after that, pass it through another DyT module to obtain the output X. CFSubsequently, a fully connected feedforward network (FFN) module is used to enhance the fitting ability.
[0051] Steps 3 and 4 involve enhancing and fusing the extracted features, which typically requires N iterations. In this embodiment, N=4 and N=2 tracking models were used in the verification experiment to compare with commonly used trackers.
[0052] Step 5: Input the obtained feature vector into the dual predictor network for classification and regression to obtain the final target tracking result. In other words, input the obtained fused feature vector into the predictor network for target localization, and finally output the bounding box of the tracked target.
[0053] The prediction head consists of classification and regression branches, each a three-layer perceptron with a hidden layer dimension d and employing the ReLU activation function. For the feature map generated by the feature fusion network, the prediction head predicts each vector to obtain foreground / background classification results, as well as normalized coordinates related to the search region size.
[0054] This invention proposes a dynamic hyperbolic tangent normalized Transformer UAV tracking method. It ultimately utilizes the pytracking (visual tracking) framework, employing classification loss, weighted binary cross-entropy (negative sample weight is 1 / 16), and the AdamW (weight decay adaptive moment estimation algorithm) optimizer, with a learning rate of... The two models for LN and DyT operations were trained in parallel on two Nvidia 2080Ti graphics processors on three common datasets (COCO, GOT-10K, and Lasot datasets), with batch = 38 and epcho = 800.
[0055] The comparison curves of the loss functions during training for normalization and DyT operations are shown below. Figure 5 As shown, it is evident that replacing the LN operation with the DyT operation results in faster convergence of the loss function and smaller training loss, verifying that the DyT module can significantly reduce training and inference time and enhance the model's generalization ability and robustness.
[0056] The performance of the LN normalized model with N=2 and N=4 was compared on the OTB100 and UAV123 datasets, as shown in Table 1.
[0057] Table 1. Comparative Experiment Results
[0058] Among them, SiamRPN++ is a target tracking algorithm based on Siamese networks, SiamCAR is a visual tracking method based on Siamese networks, DiMP is a discriminative model prediction, ARTrack is an autoregressive visual tracking algorithm, and TransT is a target tracker based on the Transformer architecture.
[0059] OTB100 is a widely used and classic benchmark dataset in the field of object tracking, containing 100 challenging video sequences. This dataset covers 11 common challenging properties in object tracking, such as illumination variation, scale variation, occlusion, deformation, motion blur, fast motion, in-plane rotation, out-of-plane rotation, out-of-view, background interference, and low resolution. OTB100 provides detailed annotation information, including the object's bounding box and motion trajectory, and is often used to evaluate the overall performance of tracking algorithms.
[0060] UAV123 is a large-scale target tracking dataset focused on the perspective of unmanned aerial vehicles (UAVs). It contains 123 video sequences captured by low-altitude UAVs, totaling over 110,000 frames. Unlike many traditional ground-based datasets, the videos in UAV123 feature unique perspectives, altitude variations, and more complex camera motions, presenting the challenge of depicting targets at different scales, from different angles, and with different motion patterns from an aerial viewpoint. This dataset provides accurate bounding box annotations and can be used to evaluate the performance of tracking algorithms in UAV applications. Its evaluation metrics are generally consistent with those of the OTB dataset, such as success rate and accuracy.
[0061] The experimental results in Table 1 demonstrate that: Experimental results demonstrate that the feature fusion mechanism based on the collaborative operation of self-attention and cross-attention is highly effective. It enabled the model to achieve leading or near-optimal tracking accuracy (success rate and precision) on two mainstream datasets, OTB100 and UAV123. This verifies that the mechanism can better model the complex semantic relationship between the target and the background, thereby improving the model's discriminative ability.
[0062] Replacing the traditional normalization layer with the DyT module significantly accelerated inference in both N=2 and N=4 model configurations without compromising tracking accuracy. Figure 5 As shown in the training loss curve, the DyT module converges faster and the loss decreases more stably during training, indicating that it plays a positive role in improving the training efficiency and generalization ability of the model.
[0063] like Figure 6Common drone tracking scenarios are presented, where the green box represents the target's ground truth bounding box and the yellow box represents the tracker's predicted bounding box. Effective tracking can be achieved in typical drone tracking scenarios. Therefore, this embodiment achieves a good balance between target tracking speed and tracking accuracy, and is more advanced than existing models.
[0064] This embodiment provides a visual target tracking method based on Dynamic Hyperbolic Tangent Normalization. By embedding the Dynamic Hyperbolic Tangent (DyT) module into a single-target tracking architecture based on Transformer feature fusion, it replaces the traditional normalization layer, significantly improving computational efficiency while maintaining high accuracy. Structurally, the DyT module uses learnable scaling parameters... By leveraging the sigmoid saturation characteristics of the hyperbolic tangent function, the model simulates the standardized and nonlinear feature transformation effects of traditional normalization layers, avoiding the calculation of mean and variance, and significantly reducing the computational complexity and memory access overhead of the model.
[0065] This embodiment provides a visual target tracking method based on dynamic hyperbolic tangent normalization. Through the synergistic effect of self-attention and cross-attention mechanisms, the model can adaptively capture long-distance semantic dependencies between the target and the background, improve the ability to distinguish difficult situations such as edge features and similar targets, and thus significantly improve the classification and regression accuracy of the tracking algorithm.
[0066] This embodiment provides a visual target tracking method based on dynamic hyperbolic tangent normalization. It uses the DyT module to replace the traditional normalization layer, avoiding the calculation of mean and variance. It achieves feature normalization only through learnable scaling parameters and tanh activation, which significantly reduces computational complexity while maintaining model performance, speeds up training and inference, and enhances the model's generalization ability.
[0067] This embodiment provides a visual target tracking method based on dynamic hyperbolic tangent normalization. The DyT module has a simple structure and few hyperparameters, and can achieve good results with default initialization, reducing the difficulty of model design and optimization, and facilitating rapid integration and deployment in practical applications.
[0068] Experiments show that this embodiment achieves state-of-the-art performance on several mainstream tracking datasets such as OTB100 and UAV123, with both success rate and accuracy reaching advanced levels. Simultaneously, its inference speed is significantly faster than traditional layer normalization (LN)-based models. Furthermore, the DyT module converges faster during training, and the loss function decreases more stably, enhancing the model's generalization ability and robustness. This embodiment is suitable for real-time visual tracking tasks on UAV platforms that require both high precision and low latency, and has promising engineering application prospects.
[0069] Example 2 This embodiment provides a visual target tracking system based on dynamic hyperbolic tangent normalization, comprising: The image acquisition module is configured to acquire template images and search images of the target being tracked. The feature extraction module is configured to extract features from the template image and the search image using a feature extraction network, and then obtain the template feature vector and the search feature vector after convolution and unrolling. The feature enhancement module is configured to perform several feature enhancement and fusion processes on the template feature vector and the search feature vector to obtain template features and tracking features. In each feature enhancement and fusion process, two encoders and two decoders are used. The two encoders process the template image and the search image respectively, and the two decoders process the outputs of the two encoders simultaneously. Both the encoder and the decoder use dynamic hyperbolic tangent to achieve layer normalization. The target tracking module is configured to fuse template features and tracking features through a decoder, and then input them into the prediction head network for target localization to obtain the bounding box of the tracked target.
[0070] Furthermore, after the encoder performs sinusoidal function position learning encoding on the feature vector, it expands and overlays the encoded feature map to obtain the query value, key value, and evaluation value. The query value, key value, and evaluation value are then processed by a multi-head attention function, followed by residual connection, and normalization is achieved using dynamic hyperbolic tangent.
[0071] Furthermore, after the decoder performs sinusoidal function position learning encoding on the two input vectors respectively, it expands and channels the encoded feature maps to obtain query values, key values and evaluation values. After fusing the query values, key values and evaluation values of the two input vectors through a residual multi-head cross-attention mechanism, it normalizes them through dynamic hyperbolic tangent and performs residual connections.
[0072] Furthermore, the dynamic hyperbolic tangent implementation layer normalization is expressed by the formula: ; in, It is a learnable scalar parameter. and These are learnable channel vector parameters, where x represents the input vector and DyT represents the dynamic hyperbolic tangent.
[0073] This embodiment provides a visual target tracking system based on dynamic hyperbolic tangent normalization, which can complete the entire process from visual perception to motion control to target tracking.
[0074] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.
[0075] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a visual target tracking method based on dynamic hyperbolic tangent normalization as described in Embodiment 1 above.
[0076] Example 4 This embodiment provides a computer device, such as... Figure 7 As shown, the system includes a computer-readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored on the computer-readable storage medium 1003 and executable on the processor 1001. The processor 1001, communication interface 1002, and computer-readable storage medium 1003 can be connected via a bus or other means. The communication interface 1002 is used to receive and send data. When the processor 1001 executes the program, it implements the steps in the visual target tracking method based on dynamic hyperbolic tangent normalization as described in Embodiment 1 above.
[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A visual target tracking method based on dynamic hyperbolic tangent normalization, characterized in that, include: Acquire template and search images of the target being tracked; For the template image and the search image, feature extraction networks are used to extract features, and after convolution and unrolling, the template feature vector and the search feature vector are obtained. For the template feature vector and the search feature vector, several feature enhancement and fusion processes are performed to obtain the template features and the tracking features. In each feature enhancement and fusion process, two encoders and two decoders are used. The two encoders process the template image and the search image respectively, and the two decoders process the outputs of the two encoders simultaneously. Both the encoder and the decoder use dynamic hyperbolic tangent to achieve layer normalization. Template features and tracking features are fused by the decoder and then input into the prediction head network for target localization to obtain the bounding box of the tracked target.
2. The visual target tracking method based on dynamic hyperbolic tangent normalization as described in claim 1, characterized in that, The encoder performs sinusoidal function position learning encoding on the feature vector, expands the encoded feature map and overlays channels to obtain query value, key value and evaluation value. The query value, key value and evaluation value are processed by multi-head attention function, residual connection is performed, and normalization is achieved by dynamic hyperbolic tangent.
3. The visual target tracking method based on dynamic hyperbolic tangent normalization as described in claim 1, characterized in that, The decoder performs sinusoidal function position learning encoding on the two input vectors respectively, expands the encoded feature maps and performs channel stacking to obtain query value, key value and evaluation value. After fusing the query value, key value and evaluation value of the two input vectors through a residual multi-head cross-attention mechanism, normalization is achieved through dynamic hyperbolic tangent and residual connection is performed.
4. The visual target tracking method based on dynamic hyperbolic tangent normalization as described in claim 1, characterized in that, The dynamic hyperbolic tangent implementation layer normalization is expressed by the formula: ； in, It is a learnable scalar parameter. and These are learnable channel vector parameters, where x represents the input vector and DyT represents the dynamic hyperbolic tangent.
5. A visual target tracking system based on dynamic hyperbolic tangent normalization, characterized in that, include: The image acquisition module is configured to acquire template images and search images of the target being tracked. The feature extraction module is configured to extract features from the template image and the search image using a feature extraction network, and then obtain the template feature vector and the search feature vector after convolution and unrolling. The feature enhancement module is configured to perform several feature enhancement and fusion processes on the template feature vector and the search feature vector to obtain template features and tracking features. In each feature enhancement and fusion process, two encoders and two decoders are used. The two encoders process the template image and the search image respectively, and the two decoders process the outputs of the two encoders simultaneously. Both the encoder and the decoder use dynamic hyperbolic tangent to achieve layer normalization. The target tracking module is configured to fuse template features and tracking features through a decoder, and then input them into the prediction head network for target localization to obtain the bounding box of the tracked target.
6. The visual target tracking system based on dynamic hyperbolic tangent normalization as described in claim 5, characterized in that, The encoder performs sinusoidal function position learning encoding on the feature vector, expands the encoded feature map and overlays channels to obtain query value, key value and evaluation value. The query value, key value and evaluation value are processed by multi-head attention function, residual connection is performed, and normalization is achieved by dynamic hyperbolic tangent.
7. A visual target tracking system based on dynamic hyperbolic tangent normalization as described in claim 5, characterized in that, The decoder performs sinusoidal function position learning encoding on the two input vectors respectively, expands the encoded feature maps and performs channel stacking to obtain query value, key value and evaluation value. After fusing the query value, key value and evaluation value of the two input vectors through a residual multi-head cross-attention mechanism, normalization is achieved through dynamic hyperbolic tangent and residual connection is performed.
8. A visual target tracking system based on dynamic hyperbolic tangent normalization as described in claim 5, characterized in that, The dynamic hyperbolic tangent implementation layer normalization is expressed by the formula: ； in, It is a learnable scalar parameter. and These are learnable channel vector parameters, where x represents the input vector and DyT represents the dynamic hyperbolic tangent.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the visual target tracking method based on dynamic hyperbolic tangent normalization as described in any one of claims 1-4.
10. A computer device comprising a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the visual target tracking method based on dynamic hyperbolic tangent normalization as described in any one of claims 1-4.