Single target tracking method and storage medium with dynamic dual template update

Abstract

The present invention discloses a single target tracking method and storage medium with dynamic dual template updating, the method comprising step 10, reading a video sequence; step 20, reading the first frame image of the video sequence and its annotation data, initializing the template to obtain a template image; step 30, reading the next frame image, processing it as a search area image; step 40, simultaneously sending the template image and the search area image into a preset strong feature extraction network to obtain a template feature map and a search area feature map respectively; step 50, performing similarity matching calculation on the template feature map and the search area feature map, outputting a similarity score response map, and obtaining a preliminary tracking prediction result. The single target tracking method with dynamic dual template updating of the present invention is a target tracking method combining a strong feature extraction network and a dynamic dual template updating strategy, and when the video sequence where the tracked target is located is relatively complex, it still has high accuracy and robust tracking.

Description

Technical Field

[0001] The present invention relates to the technical field of target tracking of a visual depth feature extraction network, and in particular to a single target tracking method and a storage medium for dynamic dual-template updating. Background Art

[0002] Achieving robust and high-precision target tracking in complex video sequences has always been a challenging problem in the field of visual target tracking. In recent years, target tracking algorithms based on twin networks have received widespread attention and in-depth research, and have become the mainstream benchmark framework for single target tracking. The target tracking algorithm based on the twin network regards the target tracking task as a similarity matching problem, and predicts the center position and tracking box of the target by calculating the similarity scores of the template branch and the search branch. The template image is initialized by cropping the target annotation data of the first frame of the video sequence, and the search area image is cropped from the current frame image. The template branch and the search branch share a deep feature extraction network, so the visual target tracking algorithm based on the twin network has achieved very excellent performance in both tracking accuracy and tracking speed.

[0003] However, most target tracking algorithms based on twin networks rely on complex and large convolutional neural networks as feature extraction networks, and rely on multiple large data sets to perform offline training on the feature extraction networks. In the tracking stage, a fixed template initialized from the first frame is used. When the video sequence of the tracked target is relatively complex, the deep feature extraction network will have a relatively poor feature extraction effect on the target, resulting in a serious decrease in the final tracking accuracy. Summary of the invention

[0004] The present invention aims at the technical problem in the prior art that when the video sequence of the tracked target is relatively complex, the deep feature extraction network has poor feature extraction effect on the target, resulting in a decrease in the final tracking accuracy, and proposes the following technical solution.

[0005] A technical solution adopted by the present invention is:

[0006] The single target tracking method with dynamic dual template update includes:

[0007] Step 10: Read the video sequence;

[0008] Step 20, read the first frame image of the video sequence and its annotation data, initialize the template to obtain a template image;

[0009] Step 30, read the next frame image and process it into a search area image;

[0010] Step 40: Send the template image and the search area image to a preset strong feature extraction network at the same time to obtain a template feature map and a search area feature map respectively;

[0011] Step 50: perform similarity matching calculation on the template feature map and the search area feature map, output a similarity score response map, and obtain a preliminary tracking prediction result.

[0012] Furthermore, the step 20 specifically includes:

[0013] Read the target center point coordinates, bounding box width and bounding box length provided by the annotation data of the first frame image of the video sequence, that is, obtain the real bounding box of the target;

[0014] Mapping the target center point coordinates to the first frame image, and cropping the first frame image based on the target center point coordinates to obtain an image block corresponding to the bounding box width and the bounding box length;

[0015] The cropped image block is resampled and adjusted to a graphic block of a preset size, and the adjusted graphic block is used as a template image.

[0016] Furthermore, after step 50, the method further includes:

[0017] Step 60: Calculate the confidence of the response graph, and determine the dynamic dual template update strategy according to the confidence:

[0018] If the confidence meets the template backup condition, the current frame image is cropped according to the preliminary tracking prediction result and saved as a backup template, and the process proceeds to step 70;

[0019] If the confidence meets the dual template starting condition, the initial template and the backup template are weightedly fused to obtain a new template, and after repeating steps 40 and 50, the process directly proceeds to step 70;

[0020] If it is other case, go directly to step 70;

[0021] Step 70: Use the preliminary tracking prediction result as the final tracking prediction result.

[0022] Further, the determining of the dynamic dual-template update strategy according to the confidence level specifically includes:

[0023] Determine the confidence of the current frame response graph: record the confidence of the current frame response graph as T i , the preset confidence upper limit threshold is denoted as T u , the preset confidence lower limit threshold is recorded as T d The dynamic dual template update strategy is as follows:

[0024] The other situations are as follows: i >T d When the preliminary tracking prediction result is used as the final tracking prediction result;

[0025] The confidence level meets the template backup condition specifically: when T i >T u When the initial tracking prediction result is used as the final tracking prediction result, the current frame image is cropped using the initial tracking prediction result, the cropping parameter setting is consistent with the initialization process of the initialization template, and the cropped image is saved as the backup template;

[0026] The confidence level meets the dual template startup condition: when T i <T d When the backup template and the initial template are weighted and fused into the new template, after repeating steps 40 and 50, the process directly proceeds to step 70.

[0027] Furthermore, after step 70, the method further includes:

[0028] Steps 30 to 70 are repeated until all image frames of the video sequence are processed and the tracking is completed.

[0029] Furthermore, the preset strong feature extraction network includes a strong feature extraction network pre-trained offline, and the steps of constructing the strong feature extraction network pre-trained offline include:

[0030] Step 41, constructing an improved Bottleneck module, connecting a center cropping module at the end of the improved Bottleneck module to eliminate the target position offset effect caused by feature map filling, thereby forming a strong feature extraction network main sub-network;

[0031] Step 42: stacking a plurality of the strong feature extraction network main body sub-networks to form a main body of the strong feature extraction network;

[0032] Step 43: Connect the channel space attention network after the main body of the strong feature extraction network to form the preset strong feature extraction network.

[0033] Furthermore, the improved Bottleneck module is specifically constructed as follows:

[0034] The number of output channels of the intermediate convolutional layer of the Bottleneck module is doubled to obtain the Bottleneck module of the strong feature extraction network body.

[0035] Furthermore, the channel space attention network includes a channel attention subnetwork and a space attention subnetwork: the step 43 specifically includes:

[0036] The input features first pass through the channel attention sub-network to output a channel-weighted feature map, and the channel-weighted feature map then passes through the spatial attention sub-network to output a spatially weighted feature map based on the channel weighting.

[0037] Furthermore, the training steps of the pre-offline trained strong feature extraction network include:

[0038] The strong feature extraction network is trained offline using a target tracking data set, and the weight parameters of the strong feature extraction network are obtained by training according to preset parameter settings. Another technical solution adopted by the present invention is: a computer-readable storage medium, wherein the computer-readable storage medium stores program data, and when the program data is executed by a processor, it is used to implement the method as described above.

[0039] The beneficial effects of the present invention are as follows: Different from the prior art, the present invention provides a single target tracking method with dynamic dual template update, including step 10, reading a video sequence; step 20, reading the first frame image of the video sequence and its annotation data, initializing the template; step 30, reading the next frame image, processing it as a search area image; step 40, simultaneously sending the template image and the search area image into a preset strong feature extraction network to obtain a template feature map and a search area feature map respectively; step 50, performing similarity matching calculation on the template feature map and the search area feature map, outputting a similarity score response map, and obtaining a preliminary tracking prediction result. The single target tracking method with dynamic dual template update of the present invention is a target tracking method that combines a strong feature extraction network and a dynamic dual template update strategy. When the video sequence where the tracked target is located is relatively complex, it still has high accuracy and robust tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0041] Figure 1 It is a flow chart of an embodiment of a single target tracking method with dynamic dual template updating provided by the present invention;

[0042] Figure 2 It is a flow chart of another embodiment of the single target tracking method with dynamic dual template update provided by the present invention;

[0043] Figure 3 It is a flow chart of another embodiment of the single target tracking method with dynamic dual template update provided by the present invention;

[0044] Figure 4 It is a structural block diagram of a preset strong feature extraction network of a single target tracking method with dynamic dual template update provided by the present invention;

[0045] Figure 5 It is a structural block diagram of an embodiment of a single target tracking device with dynamic dual template update provided by the present invention;

[0046] Figure 6 It is a structural schematic diagram of an embodiment of a computer-readable storage medium provided by the present invention. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It will be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only some but not all methods and processes related to the present invention are shown in the drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] The terms "including" and "having" and any variations thereof in the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products or devices.

[0049] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0050] Most target tracking algorithms based on twin networks rely on complex and large convolutional neural networks as feature extraction networks, and rely on multiple large data sets for offline training of feature extraction networks. In the tracking stage, a fixed template initialized from the first frame is used. When the video sequence of the tracked target is more complex, or the target category is not included in the large data set for offline training, the deep feature extraction network will have very poor feature extraction effect on the target, resulting in a serious decrease in the final tracking accuracy.

[0051] In response to the above-mentioned existing technical problems, the present invention designs a target tracking method that combines a strong feature extraction network and a dynamic dual-template updating strategy, which can achieve high-precision and robust tracking under complex challenges such as illumination changes, background interference, fast motion and motion blur.

[0052] The existing target tracking method based on twin networks relies on large data sets to learn the universal target matching ability of the feature extraction network through offline end-to-end training. There is no online update and fine-tuning processing in the tracking stage. Only the template initialized by the first frame is used, which makes it unable to adapt to tracking tasks in more complex scenes such as lighting changes, background interference, fast motion and motion blur, and has low tracking accuracy and non-robust tracking.

[0053] In order to solve the above problems existing in the prior art, the present invention designs a target tracking method that combines a strong feature extraction network and a dynamic dual-template updating strategy, which can achieve high-precision and robust tracking under complex challenges.

[0054] See also Figure 1 , Figure 1 It is a flow chart of an embodiment of a single target tracking method with dynamic dual template updating provided by the present invention.

[0055] The single target tracking method with dynamic dual template update according to the embodiment of the present invention comprises:

[0056] Step 10: Read the video sequence;

[0057] Step 20, read the first frame image of the video sequence and its annotation data, initialize the template to obtain a template image;

[0058] Step 20 may specifically include:

[0059] Read the target center point coordinates, bounding box width and bounding box length provided by the annotation data of the first frame image of the video sequence, that is, obtain the real bounding box of the target;

[0060] Mapping the target center point coordinates to the first frame image, and cropping the first frame image based on the target center point coordinates to obtain an image block corresponding to the bounding box width and the bounding box length;

[0061] The cropped image block is resampled and adjusted to a graphic block of a preset size (eg, 127×127), and the adjusted graphic block is used as a template image.

[0062] Step 30, read the next frame image and process it into a search area image; specifically, the processing into the search area image may be performed by image processing measures such as cropping, etc. The search area image is also the input image of the search branch.

[0063] Step 40: Send the template image and the search area image to a preset strong feature extraction network at the same time to obtain a template feature map and a search area feature map respectively;

[0064] Step 50: perform similarity matching calculation on the template feature map and the search area feature map, output a similarity score response map, and obtain a preliminary tracking prediction result.

[0065] Furthermore, combined with Figure 2 After step 50, the method further comprises:

[0066] Step 60: Calculate the confidence of the response graph, and determine the dynamic dual template update strategy according to the confidence:

[0067] If the confidence meets the template backup condition, the current frame image is cropped according to the preliminary tracking prediction result and saved as a backup template, and the process proceeds to step 70;

[0068] If the confidence meets the dual template starting condition, the initial template and the backup template are weightedly fused to obtain a new template, and after repeating steps 40 and 50, the process directly proceeds to step 70;

[0069] If it is other cases, directly go to step 70; specifically, the other cases here refer to cases other than the first and second cases.

[0070] Step 70: Use the preliminary tracking prediction result as the final tracking prediction result.

[0071] In one embodiment, determining the dynamic dual-template update strategy according to the confidence level specifically includes:

[0072] Determine the confidence of the current frame response graph; record the confidence of the current frame response graph as T i , the preset confidence upper limit threshold is denoted as T u , the preset confidence lower limit threshold is recorded as T d , the dynamic dual template update strategy is as follows:

[0073] The other situations are as follows: i >T d When the preliminary tracking prediction result is used as the final tracking prediction result;

[0074] The confidence level meets the template backup condition specifically: when T i >T u When the initial tracking prediction result is used as the final tracking prediction result, the current frame image is cropped using the initial tracking prediction result, the cropping parameter setting is consistent with the initialization process of the initialization template, and the cropped image is saved as the backup template;

[0075] The confidence level meets the dual template startup condition: when T i <T d When the backup template and the initial template are weighted and fused into the new template, after repeating steps 40 and 50, the process directly proceeds to step 70.

[0076] Furthermore, after step 70, the method further includes:

[0077] Steps 30 to 70 are repeated until all image frames of the video sequence are processed and the tracking is completed.

[0078] In some embodiments, the step of calculating the confidence of the response graph and entering the dynamic dual-template update strategy judgment branch specifically includes:

[0079] The average peak correlation energy of the response graph is calculated and saved as the confidence of the current frame response graph, and the confidence of the current frame response graph is placed in the dynamic dual-template update strategy judgment branch.

[0080] In some embodiments, after the step of outputting the similarity score response graph, the method further comprises:

[0081] Get the maximum response value of the response plot.

[0082] The above embodiment is an online tracking process of a single target tracking method with dynamic dual template update. The following is an explanation of the construction of a strong feature extraction network and the training process of the strong feature extraction network specifically included in the single target tracking method with dynamic dual template update of the present invention.

[0083] In some embodiments, in combination Figure 3 and Figure 4 , the preset strong feature extraction network includes a strong feature extraction network pre-trained offline. Optionally, the steps of constructing the pre-trained offline strong feature extraction network include:

[0084] Step 41, constructing an improved Bottleneck module, connecting a center cropping module at the end of the improved Bottleneck module to eliminate the target position offset effect caused by feature map filling, thereby forming a strong feature extraction network main sub-network;

[0085] Step 42: stacking a plurality of the strong feature extraction network main body sub-networks to form a main body of the strong feature extraction network;

[0086] Step 43: Connect the channel space attention network after the main body of the strong feature extraction network to form the preset strong feature extraction network.

[0087] Furthermore, combined with Figure 4 , the channel space attention network includes a channel attention subnetwork and a spatial attention subnetwork: the step 43 specifically includes:

[0088] The input features first pass through the channel attention sub-network to output a channel-weighted feature map, and the channel-weighted feature map then passes through the spatial attention sub-network to output a spatially weighted feature map based on the channel weighting.

[0089] In some embodiments, the training step of the pre-offline trained strong feature extraction network includes:

[0090] The target tracking data set is used to perform offline training on the strong feature extraction network, and the weight parameters of the strong feature extraction network are obtained by training according to the preset parameter settings. Specifically, the predetermined data set is a large data set GOT10K.

[0091] The preset parameter settings are as follows:

[0092] 1. Use nearly one million pairs (e.g., 937,500 pairs) of images in GOT10K as training images, train for multiple cycles (e.g., 50 cycles), and each cycle contains more than ten thousand pairs (e.g., 18,750 pairs) of images.

[0093] 2. Set the template branch input image size (i.e., set the template image size), for example, to 127×127, and set the search branch input image size (i.e., set the search area image size), for example, to 255×255.

[0094] 3. Use SDG for training (use SDG function to train the network), set the parameters of momentum, batchsize and learning rate. For example, set momentum to 0.9, batchsize to 32, and the learning rate to decay exponentially from 0.01 to 0.00001.

[0095] 4. Set the loss function to be used, for example, BCEWithLogitsLoss.

[0096] According to the above settings, the final weighted feature extraction network is obtained after training.

[0097] The beneficial effect of the present invention is that, different from the prior art, the single target tracking method with dynamic dual template update provided by the present invention is a target tracking method that combines a strong feature extraction network and a dynamic dual template update strategy, which can achieve high-precision and robust tracking under complex challenges such as illumination changes, background interference, rapid motion and motion blur.

[0098] The single target tracking method with dynamic dual template updating provided by the present invention is a target tracking method combining a strong feature extraction network and a dynamic dual template updating strategy, which specifically includes three parts: construction of a strong feature extraction network, training of a strong feature extraction network, and online tracking.

[0099] 1. Construction of strong feature extraction network

[0100] The present invention is based on the twin network tracking method, and the overall framework consists of a template branch, a search branch, a similarity calculation module, and a tracking prediction module. Among them, the template branch and the search branch share a strong feature extraction network, and the construction of the strong feature extraction network is described in the following steps:

[0101] 1. Construct an improved Bottleneck module, specifically by doubling the number of output channels of the intermediate convolutional layer of the Bottlenec module to obtain the Bottleneck submodule of the main body of the strong feature extraction network.

[0102] 2. A center cropping module is connected to the end of the improved Bottleneck module constructed in step 1 to form a sub-network of the strong feature extraction network. The center cropping module specifically crops the outer circle of the input feature map and retains the central part of the feature map to eliminate the impact of the significant offset of the target position caused by the filling operation of the feature map as the network deepens.

[0103] 3. The sub-networks of the strong feature extraction network main body constructed in step 2 are stacked into the main body of the strong feature extraction network, and the specific number of stacking is 7.

[0104] 4. The main end of the strong feature extraction network constructed in step 3 is connected to the channel spatial attention network to construct the final strong feature extraction network. The channel spatial attention network is as follows: the input feature first passes through the channel attention sub-network, outputs the channel weighted feature map, and then the channel weighted feature map passes through the spatial attention sub-network, outputs the spatial weighted feature map based on the channel weighting.

[0105] 2. Training of Strong Feature Extraction Network

[0106] The strong feature extraction network designed by the present invention is trained offline on the large dataset GOT10K. The parameter settings are as follows:

[0107] 1. A total of 937,500 pairs of images from GOT10K are used as training images. A total of 50 cycles are trained, and each cycle contains 18,750 pairs of images.

[0108] 2. The template branch input image size is 127×127, and the search branch input image size is 255×255.

[0109] 3. Use SDG for training, with momentum set to 0.9, batchsize set to 32, and learning rate decayed exponentially from 0.01 to 0.00001.

[0110] 4. The loss function uses BCEWithLogitsLoss.

[0111] According to the above settings, the final weighted feature extraction network is obtained after training.

[0112] 3. Online Tracking

[0113] 1. Read the video sequence.

[0114] 2. Read the first frame image of the video sequence and its annotation data to initialize the template.

[0115] 3. Read the next frame image and crop it to be the search branch input image.

[0116] 4. Send the template image and the search area image to the preset strong feature extraction network at the same time, and output the template feature map and the search area feature map respectively.

[0117] 5. Perform similarity matching calculation on the template feature map and the search area feature map, specifically, perform cross-correlation calculation on the two feature maps and output a response map.

[0118] 6. Obtain the maximum response value of the response graph and obtain preliminary tracking prediction results.

[0119] 7. Calculate the average peak correlation energy of the response map and save it as the confidence of the response map of the current frame.

[0120] 8. Determine the confidence of the current frame response graph. The confidence of the current response graph is recorded as T i , the preset confidence upper limit threshold is denoted as T u , the preset confidence lower limit threshold is recorded as T d The dynamic dual-template update strategy designed by the present invention is specifically described as follows:

[0121] (1) When T i >T d When the initial tracking prediction result obtained in step 6 is used as the final tracking prediction result.

[0122] (2) When T i >T u When the initial tracking prediction result obtained in step 6 is used as the final tracking prediction result, the current frame image is cropped using the tracking prediction result, the cropping parameter setting is consistent with the initial template initialization process, and the cropped image is saved as a backup template.

[0123] (3) When T i <T d When the backup template and the initial template are weighted and fused into a new template, steps 4-6 are repeated to obtain the final tracking prediction result.

[0124] 9. Repeat steps 3-8 until all image frames of the video sequence are processed and tracking is completed.

[0125] The present invention designs a strong feature extraction network, which can extract feature maps with strong representation ability and strong discrimination for target tracking tasks, significantly improving the success rate and accuracy of target tracking.

[0126] The present invention designs a dynamic dual-template updating strategy for online tracking, which achieves high-precision and robust tracking under complex challenges such as illumination changes, background interference, fast motion and motion blur.

[0127] The test and evaluation results of the inventors on the dataset OTB2015 show that the single target tracking method with dynamic dual template update designed by the present invention has an overall tracking success rate (AUC) of 0.666 and a tracking accuracy of 0.879, which is better than mainstream advanced algorithms such as SiamFC, SiamRPN, and DaSiamRPN.

[0128] According to the above embodiments, the present invention proposes a single target tracking device with dynamic dual template update. Figure 5 , Figure 5 It is a structural schematic diagram of an embodiment of a single target tracking device with dynamic dual template updating provided by the present invention.

[0129] The single target tracking device 100 with dynamic dual template update includes a processor 110 and a memory 120. The processor 110 and the memory 120 are coupled. The memory 120 stores a computer program, which is used to execute the single target tracking method with dynamic dual template update.

[0130] Specific reference Figure 6 , Figure 6 It is a structural schematic diagram of an embodiment of a computer-readable storage medium provided by the present invention.

[0131] The computer-readable storage medium 200 includes program data 210. When the program data 210 is executed by the processor, the single target tracking method with dynamic dual template update can be implemented.

[0132] Different from the prior art, the present invention provides a single target tracking method with dynamic dual template updating.

[0133] The beneficial effects of the present invention are as follows: Different from the prior art, the present invention provides a single target tracking method with dynamic dual template update, including step 10, reading a video sequence; step 20, reading the first frame image of the video sequence and its annotation data, initializing the template; step 30, reading the next frame image, processing it as a search area image; step 40, simultaneously sending the template image and the search area image into a preset strong feature extraction network to obtain a template feature map and a search area feature map respectively; step 50, performing similarity matching calculation on the template feature map and the search area feature map, outputting a similarity score response map, and obtaining a preliminary tracking prediction result. The single target tracking method with dynamic dual template update of the present invention is a target tracking method that combines a strong feature extraction network and a dynamic dual template update strategy. When the video sequence where the tracked target is located is relatively complex, it still has high accuracy and robust tracking.

[0134] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A single target tracking method with dynamic dual template update, characterized in that: include: Step 10: Read the video sequence; Step 20: read the first frame image of the video sequence and its annotation data, and initialize the template to obtain a template image; Step 30, read the next frame image and process it into a search area image; Step 40: Send the template image and the search area image to a preset strong feature extraction network at the same time to obtain a template feature map and a search area feature map respectively; Step 50: perform similarity matching calculation on the template feature map and the search area feature map, output a similarity score response map, and obtain a preliminary tracking prediction result; After step 50, the method further includes: Step 60: Calculate the confidence of the response graph, and determine the dynamic dual template update strategy according to the confidence: If the confidence meets the template backup condition, the current frame image is cropped according to the preliminary tracking prediction result and saved as a backup template, and the process proceeds to step 70; If the confidence meets the dual template starting condition, the initial template and the backup template are weightedly fused to obtain a new template, and after repeating steps 40 and 50, the process directly proceeds to step 70; If it is other case, go directly to step 70; Step 70: taking the preliminary tracking prediction result as the final tracking prediction result; Determining the dynamic dual-template update strategy according to the confidence level specifically includes: Determine the confidence of the current frame response graph: record the confidence of the current frame response graph as , the preset confidence upper limit threshold is recorded as , the preset confidence lower limit threshold is recorded as The dynamic dual template update strategy is as follows: The other circumstances are: > When the preliminary tracking prediction result is used as the final tracking prediction result; The confidence level meets the template backup condition specifically when: > When the initial tracking prediction result is used as the final tracking prediction result, the current frame image is cropped using the initial tracking prediction result, the cropping parameter setting is consistent with the initialization process of the initialization template, and the cropped image is saved as the backup template; The confidence level meets the dual template startup condition specifically: < When the backup template and the initial template are weighted and fused into the new template, after repeating steps 40 and 50, the process directly proceeds to step 70.

2. The method according to claim 1, characterized in that The step 20 specifically includes: Read the target center point coordinates, bounding box width and bounding box length provided by the annotation data of the first frame image of the video sequence, that is, obtain the real bounding box of the target; Mapping the target center point coordinates to the first frame image, and cropping the first frame image based on the target center point coordinates to obtain an image block corresponding to the bounding box width and the bounding box length; The cropped image block is resampled and adjusted to a graphic block of a preset size, and the adjusted graphic block is used as a template image.

3. The method according to claim 1, characterized in that After step 70, the method further includes: Steps 30 to 70 are repeated until all image frames of the video sequence are processed and the tracking is completed.

4. The method according to any one of claims 1 to 3, characterized in that: The preset strong feature extraction network includes a strong feature extraction network pre-trained offline, and the steps of constructing the strong feature extraction network pre-trained offline include: Step 41, constructing an improved Bottleneck module, connecting a center cropping module at the end of the improved Bottleneck module to eliminate the target position offset effect caused by feature map filling, thereby forming a strong feature extraction network main sub-network; Step 42: stacking a plurality of the strong feature extraction network main body sub-networks to form a main body of the strong feature extraction network; Step 43: Connect the channel space attention network after the main body of the strong feature extraction network to form the preset strong feature extraction network.

5. The method according to claim 4, characterized in that The Bottleneck module of the construction improvement is specifically as follows: The number of output channels of the intermediate convolutional layer of the Bottleneck module is doubled to obtain the Bottleneck module of the strong feature extraction network body.

6. The method according to claim 5, characterized in that The channel space attention network includes a channel attention sub-network and a space attention sub-network: the step 43 specifically includes: The input features first pass through the channel attention sub-network to output a channel-weighted feature map, and the channel-weighted feature map then passes through the spatial attention sub-network to output a spatially weighted feature map based on the channel weighting.

7. The method according to claim 6, characterized in that The training steps of the pre-offline trained strong feature extraction network include: The target tracking dataset is used to perform offline training on the strong feature extraction network, and the weight parameters of the strong feature extraction network are obtained through training according to the preset parameter settings.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program data, and when the program data is executed by a processor, it is used to implement the method according to any one of claims 1 to 7.

Images