A method, device, electronic device and storage medium for tracking a target

By extracting and fusing feature maps from video frames in the target tracking method, and using aligned convolutional neural networks for bounding box filtering and adjustment, the problem of inaccurate target state prediction in existing technologies is solved, achieving higher tracking accuracy.

CN112950675BActive Publication Date: 2024-11-26SHENZHEN SENSETIME TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110291485.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-03-18
Publication Date
2024-11-26
Estimated Expiration
2041-03-18

AI Technical Summary

Technical Problem

Existing target tracking methods only provide the initial state of the target, resulting in poor accuracy in predicting the target state and poor tracking performance when faced with similar objects, complex backgrounds, deformation of non-rigid objects, and rapid movement.

Method used

By acquiring initial video frames and video frames to be tracked labeled with target bounding boxes, first and second feature maps are extracted, the fused feature maps are used to generate initial predicted bounding boxes, and convolution operations are performed using an aligned convolutional neural network to filter and adjust the bounding boxes to determine the location of the target object.

Benefits of technology

It improves the accuracy of target state prediction and enhances the effect of target tracking, enabling more accurate capture of the target object's position in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112950675B_ABST
    Figure CN112950675B_ABST
Patent Text Reader

Abstract

The present disclosure provides a method, device, electronic device and storage medium for target tracking, wherein the method includes: obtaining an initial video frame marked with a target bounding box and a target video frame to be tracked; the target bounding box contains a target object; extracting a first feature map and a second feature map from the initial video frame and the target video frame respectively; determining a fused feature map based on the first feature map and the second feature map, and generating at least one initial predicted bounding box for the target video frame based on the fused feature map; determining a target tracking result for the target video frame according to at least one initial predicted bounding box of the target video frame and the fused feature map; wherein the target tracking result includes a target bounding box selected from at least one initial predicted bounding box. The present disclosure uses the fused feature map to screen the initial predicted bounding box, thereby selecting a more accurate position of the target object, and achieving a better tracking effect.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the technical field of target tracking, and in particular to a method, device, electronic device and storage medium for target tracking. Background Art

[0002] As an important research direction in computer vision, object tracking is widely used in various fields, such as video surveillance, human-computer interaction, and unmanned driving. Existing object tracking tasks can be based on the object state given in the initial frame of the video, for example, using a bounding box to represent the location of the object to predict the object state in subsequent frames.

[0003] However, the existing methods only give the initial state of the target, and there are problems such as similar objects, complex background, deformation and rapid motion of non-rigid objects, which makes the predicted target state less accurate and the tracking effect poor. Summary of the invention

[0004] The embodiments of the present disclosure at least provide a method, device, electronic device and storage medium for target tracking, which improve the accuracy of target state prediction and achieve better target tracking effect.

[0005] In a first aspect, an embodiment of the present disclosure provides a method for tracking a target, the method comprising:

[0006] Acquire an initial video frame marked with a target bounding box and a target video frame to be tracked; the target bounding box contains a target object;

[0007] Extracting a first feature map and a second feature map from the initial video frame and the target video frame respectively;

[0008] Determine a fused feature map based on the first feature map and the second feature map, and generate at least one initial prediction bounding box for the target video frame based on the fused feature map;

[0009] Determine a target tracking result for the target video frame according to at least one initial predicted bounding box of the target video frame and the fused feature map; wherein the target tracking result includes a target bounding box selected from the at least one initial predicted bounding box and containing the target object.

[0010] By adopting the above-mentioned target tracking method, when the initial video frame marked with the target bounding box and the target video frame to be tracked are obtained, feature extraction can be performed first, so that a fused feature map can be obtained based on the extracted first feature map and the second feature map. The fused feature map is fused with the relevant information of the initial video frame and the target bounding box in the initial video frame, and is also fused with the relevant information of the target video frame. In this way, at least one initial prediction bounding box for the target video frame can be generated based on the fused feature map, and the initial prediction bounding box can indicate the position where the target object may appear in the target video frame. In order to determine the target bounding box containing the target object in the target video frame, the fused feature map can be correlated with the initial prediction bounding box, that is, the fused feature map can be further used to screen the initial prediction bounding box, and then a more accurate position of the target object can be selected, so that the tracking effect is better.

[0011] In a possible implementation, determining the target tracking result for the target video frame according to at least one initial predicted bounding box of the target video frame and the fused feature map includes:

[0012] Performing a first convolution operation on the fused feature map based on a first preset convolution parameter to obtain a first convolution feature map;

[0013] Using the trained first aligned convolutional neural network and at least one initial predicted bounding box of the target video frame, performing an aligned convolution operation on the first convolutional feature map to obtain a first convolutional feature map output by the network;

[0014] Determine a probability value that each of the at least one initial prediction bounding box indicated by the first post-convolution feature map output by the network contains a target object;

[0015] An initial prediction bounding box with a maximum probability value is selected from the at least one initial prediction bounding box as the target bounding box containing the target object.

[0016] Here, the trained first aligned convolutional neural network can be used to perform aligned convolution operations. The aligned convolution operation here is to perform a convolution operation on the feature map after the first convolution using the initial prediction bounding box. That is, the convolution sampling position can be determined based on the initial prediction bounding box. Compared with direct convolution, the context information of the initial prediction bounding box is referred to here, so that the position of the determined target object is more accurate.

[0017] In a possible implementation, using the trained first aligned convolutional neural network and at least one initial predicted bounding box of the target video frame to perform an aligned convolution operation on the first convolutional feature map to obtain a first convolutional feature map output by the network includes:

[0018] Determine the convolution kernel parameters corresponding to the trained first aligned convolutional neural network;

[0019] Using the convolution kernel parameters to traverse the first convolution feature map;

[0020] When traversing to a first convolution feature region corresponding to any initial prediction bounding box of the target video frame, convolution operation is performed on the convolution kernel indicated by the convolution kernel parameter and the first convolution feature region to obtain a first convolution feature map output by the network;

[0021] The first post-convolution feature area is contained in the first post-convolution feature map and has the same size as the initial prediction bounding box.

[0022] Here, the aligned convolution operation can be implemented by traversing convolution. During the operation, the convolution kernel indicated by the convolution kernel parameters can be convolved with the first post-convolution feature area. Since the first post-convolution feature area has the same size as the initial predicted bounding box, the vertex positions corresponding to the initial predicted bounding box in the first post-convolution feature map can be involved in the convolution operation. Through this operation process, the geometric information of the bounding box can be well captured, so that the position of the target object determined is more accurate.

[0023] In a possible implementation, the method further includes:

[0024] Performing a second convolution operation on the fused feature map based on a second preset convolution parameter to obtain a second convolution feature map;

[0025] Using the trained second aligned convolutional neural network and the target bounding box containing the target object in the target video frame, an aligned convolution operation is performed on the second convolutional feature map to obtain a second convolutional feature map output by the network;

[0026] Determining an adjustment amount for the target bounding box indicated by a second post-convolution feature map output by the network;

[0027] The target bounding box is adjusted based on the adjustment amount output by the network to obtain an adjusted target bounding box.

[0028] Here, the trained second aligned convolutional neural network can be used to perform an aligned convolution operation. The aligned convolution operation here is that the target bounding box performs a convolution operation on the second convolution feature map, that is, the target bounding box can be adjusted based on the reference second convolution feature map, so that the adjusted target bounding box can be closer to the target object.

[0029] In a possible implementation manner, the target video frame to be tracked is annotated with a target bounding box containing the target object; and the first aligned convolutional neural network and the second aligned convolutional neural network are trained according to the following steps:

[0030] Inputting the generated at least one initial prediction bounding box for the target video frame and the first convolution feature map into a first aligned convolutional neural network to be trained, and obtaining a first convolution feature map output by the network; determining a probability value that each of the at least one initial prediction bounding box indicated by the first convolution feature map output by the network contains a target object; comparing the position information of the initial prediction bounding box with the largest probability value with the position information of the target bounding box marked for the target video frame to be tracked, and obtaining a first comparison result;

[0031] And, inputting the generated at least one initial prediction bounding box for the target video frame and the second convolution feature map into a second aligned convolutional neural network to be trained, to obtain a second convolution feature map output by the network; determining the adjustment amount required for each initial prediction bounding box in the at least one initial prediction bounding box indicated by the second convolution feature map output by the network; adjusting each initial prediction bounding box according to the adjustment amount to obtain position information of the adjusted initial prediction bounding box, and comparing the position information of the target bounding box annotated for the target video frame to be tracked, to obtain a second comparison result;

[0032] If any comparison result of the first comparison result and the second comparison result is inconsistent, parameters of at least one of the first aligned convolutional neural network and the second aligned convolutional neural network are adjusted.

[0033] In a possible implementation, extracting the first feature map and the second feature map from the initial video frame and the target video frame respectively includes:

[0034] The trained feature extraction network is used to perform feature extraction on the initial video frame and the target video frame respectively to obtain a first feature map and a second feature map.

[0035] In a possible implementation, the method further includes:

[0036] Determining a target mask feature corresponding to the initial video frame; wherein the target mask feature of the video frame region within the target bounding box in the initial video frame is different from the target mask feature of the video frame region outside the target bounding box;

[0037] The first feature map is updated based on the determined target mask feature to obtain an updated first feature map.

[0038] Here, the image features of the video frame area where the target object is located can be emphasized based on the target mask features, that is, only the target object can be used as the foreground and all non-target objects can be used as the background. In this way, in the subsequent target object tracking stage, more focus can be placed on tracking the target object, and the tracking effect can be better.

[0039] In a possible implementation, determining a fused feature map based on the first feature map and the second feature map includes:

[0040] The trained cross-correlation network is used to perform a cross-correlation operation on the first feature map and the second feature map to obtain a fused feature map.

[0041] In a possible implementation, the using a trained cross-correlation network to perform a cross-correlation operation on the first feature map and the second feature map to obtain a fused feature map includes:

[0042] Using the first feature map as the convolution kernel corresponding to the trained cross-correlation network;

[0043] Using the convolution kernel to traverse the second feature map, and determine the convolution result at each traversal position;

[0044] The convolution results at each traversal position are combined to obtain a fused feature map.

[0045] Here, the first feature map corresponding to the initial video frame marked with the target bounding box can be used as the convolution kernel, and then the video frame area related to the initial video frame in the target video frame is determined by traversing the convolution on the second feature map. These highly correlated video frame areas are more likely to have the target object to a certain extent, so that the fused feature map can better guide the determination of the initial predicted bounding box.

[0046] In a second aspect, an embodiment of the present disclosure further provides a target tracking device, the device comprising:

[0047] An acquisition module, used to acquire an initial video frame marked with a target bounding box and a target video frame to be tracked; the target bounding box contains a target object;

[0048] An extraction module, used to extract a first feature map and a second feature map from the initial video frame and the target video frame respectively;

[0049] A fusion module, configured to determine a fused feature map based on the first feature map and the second feature map, and generate at least one initial prediction bounding box for the target video frame based on the fused feature map;

[0050] A tracking module is used to determine a target tracking result for the target video frame based on at least one initial predicted bounding box of the target video frame and the fused feature map; wherein the target tracking result includes a target bounding box selected from the at least one initial predicted bounding box and containing the target object.

[0051] In a third aspect, an embodiment of the present disclosure further provides an electronic device, comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the steps of the target tracking method as described in the first aspect and any of its various embodiments are performed.

[0052] In a fourth aspect, an embodiment of the present disclosure further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, executes the steps of the target tracking method as described in the first aspect and any of its various embodiments.

[0053] For a description of the effects of the above-mentioned target tracking device, electronic device, and computer-readable storage medium, please refer to the description of the above-mentioned target tracking method, which will not be repeated here.

[0054] In order to make the above-mentioned objectives, features and advantages of the present disclosure more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following is a brief introduction to the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can also be obtained based on these drawings without creative work.

[0056] Figure 1 A flow chart of a target tracking method provided by an embodiment of the present disclosure is shown;

[0057] Figure 2 A schematic diagram of calculation of aligned convolution in a target tracking method provided by an embodiment of the present disclosure is shown;

[0058] Figure 3 A schematic diagram showing an application of a target tracking method provided by an embodiment of the present disclosure is shown;

[0059] Figure 4 A schematic diagram of a target tracking device provided by an embodiment of the present disclosure is shown;

[0060] Figure 5 A schematic diagram of an electronic device provided by an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical scheme and advantages of the embodiments of the present disclosure clearer, the technical scheme in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The components of the embodiments of the present disclosure generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the present disclosure for protection, but merely represents the selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present disclosure.

[0062] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0063] The term "and / or" herein only describes an association relationship, indicating that three relationships may exist. For example, A and / or B may represent the following three situations: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C may represent including any one or more elements selected from the set consisting of A, B, and C.

[0064] It has been found that the existing target tracking task can be based on the target state given in the initial frame of the video, for example, using a bounding box to represent the location of the target to predict the target state in subsequent frames.

[0065] However, the existing methods only give the initial state of the target, and there are problems such as similar objects, complex background, deformation and rapid motion of non-rigid objects, which makes the predicted target state less accurate and the tracking effect poor.

[0066] Based on the above research, the present disclosure provides a method, device, electronic device and storage medium for target tracking, which improve the accuracy of target state prediction and achieve better target tracking effect.

[0067] To facilitate understanding of the present embodiment, a target tracking method disclosed in the embodiment of the present disclosure is first introduced in detail. The execution subject of the target tracking method provided in the embodiment of the present disclosure is generally a computer device with certain computing capabilities, and the computer device includes, for example: a terminal device or a server or other processing device, and the terminal device can be a user equipment (User Equipment, UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (Personal Digital Assistant, PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. In some possible implementations, the target tracking method can be implemented by a processor calling a computer-readable instruction stored in a memory.

[0068] See also Figure 1 FIG. 1 is a flowchart of a target tracking method provided by an embodiment of the present disclosure, the method comprising steps S101 to S104, wherein:

[0069] S101: Acquire an initial video frame marked with a target bounding box and a target video frame to be tracked; the target bounding box contains a target object;

[0070] S102: extracting a first feature map and a second feature map from an initial video frame and a target video frame respectively;

[0071] S103: determining a fused feature map based on the first feature map and the second feature map, and generating at least one initial prediction bounding box for the target video frame based on the fused feature map;

[0072] S104: Determine a target tracking result for the target video frame according to at least one initial predicted bounding box of the target video frame and the fused feature map; wherein the target tracking result includes a target bounding box containing the target object selected from the at least one initial predicted bounding box.

[0073] Here, in order to facilitate understanding of the target tracking method provided by the embodiment of the present disclosure, the application scenario of the method can be briefly described first. The above-mentioned target tracking method can be mainly applied to target tracking in various fields, for example, it can be used to track the target object in the field of human-computer interaction, the field of video surveillance, the field of intelligent driving, and the field of robots. The target object here can be a person or an object, for example, a vehicle. The embodiment of the present disclosure does not make specific restrictions on this.

[0074] Considering that the current target tracking method can predict the target state of subsequent frames based on the initial state of the target determined in the initial frame, since only the initial state of the target is given, it is difficult to accurately predict the state of the target in subsequent frames in the presence of similar targets, background clutter, rapid target movement and other difficult situations.

[0075] Precisely to solve the above problems, the embodiments of the present disclosure provide a solution for screening the initial predicted bounding box based on feature fusion to determine the final target tracking result, so that the predicted target state has better accuracy and better tracking effect.

[0076] Among them, for the initial video frame marked with a target bounding box, the target bounding box can be manually selected or marked using relevant annotation tools. In specific applications, a rectangular box can be used as the target bounding box. The initial video frame here can be a video frame in which the target object first appears in a video clip. The video frame can be the first video frame (i.e., the first frame) in the video clip, or a video frame in the middle of the video clip. It can be determined according to different application scenarios and is not specifically limited here.

[0077] In addition, the target video frame to be tracked here can be a video frame after the initial video frame, for example, it can be the second video frame, the third video frame, etc. after the initial video frame. The target tracking method provided in the embodiment of the present disclosure aims to determine information such as the position status of the target object from the target video frame.

[0078] It should be noted that, in practical applications, considering that the resolution of the initial video frame and the target video frame is usually relatively high, if the subsequent related operations are directly performed, it is likely to result in a large amount of operations. Therefore, the embodiments of the present disclosure can perform shearing processing on the initial video frame and the target video frame in combination with the actual scene. For example, for the initial video frame, a local image centered on the target bounding box where the target object is located can be intercepted. For another example, when the size of the target object is relatively small, the above-mentioned intercepted local image can be enlarged, and at the same time, the target video frame can also be synchronously enlarged.

[0079] Considering the key role of the features of the target object in the initial video frame in determining the relevant information of the target object in the target video frame, here, the first feature map and the second feature map can be extracted from the initial video frame and the target video frame respectively (that is, the first feature map is extracted from the initial video frame, and the second feature map is extracted from the target video frame), and then the fused feature map is determined.

[0080] The above two feature maps (i.e., the first feature map and the second feature map) can be a set of features extracted from the corresponding video frame. For example, they can be extracted features such as the position, content, color, etc. of the target object, or they can be related features of the background of the video frame. The embodiments of the present disclosure do not impose specific restrictions on this.

[0081] The feature extraction process of the initial video frame and the target video frame may be obtained based on a trained feature extraction network, or may be directly extracted from the video frame based on an image processing method.

[0082] Here, considering that the feature extraction network can dig out deeper features, the embodiment of the present disclosure can select a trained feature extraction network to extract the above two feature maps, wherein the above feature extraction network can be obtained based on ResNet-50 training.

[0083] In the process of feature extraction, in order to make full use of the target bounding box in the initial video frame, here, the corresponding target mask feature can be determined for the initial video frame. Here, the target mask feature of the video frame area within the target bounding box in the initial video frame is different from the target mask feature of the video frame area outside the target bounding box. For example, the value of the target mask feature of the video frame area within the target bounding box can be set to 1, and the value of the target mask feature of the video frame area outside the target bounding box can be set to -1. At this time, the target object can be well distinguished from the background. In this way, the updated first feature map can more focusedly reflect the characteristics of the target object.

[0084] In order to better determine information such as the position state of the target object in the target video frame, the embodiment of the present disclosure may first perform feature fusion on the two extracted feature maps to obtain a fused feature map.

[0085] Considering that the purpose of feature fusion here is to search for information about the target object in the target video frame, fusion can be achieved here based on cross-correlation operation. In the embodiment of the present disclosure, cross-correlation operation can be performed on the above two feature maps based on a cross-correlation network to obtain a fused feature map.

[0086] The embodiment of the present disclosure can implement the cross-correlation operation according to the following steps:

[0087] Step 1: Use the first feature map as the convolution kernel corresponding to the trained cross-correlation network;

[0088] Step 2: Use the convolution kernel to traverse the second feature map and determine the convolution result at each traversal position;

[0089] Step 3: Combine the convolution results at each traversal position to obtain the fused feature map.

[0090] Here, the first feature map can be used as the convolution kernel corresponding to the cross-correlation network to traverse the convolution of the second feature map. In this way, for each traversal position, a convolution result can be determined, and the convolution result can characterize the correlation between the target object in the initial video frame and the target video frame to a certain extent. The larger the convolution value, the higher the corresponding correlation. Similarly, the smaller the convolution value, the lower the corresponding correlation. The convolution results are combined according to the traversal positions to obtain the fused feature map.

[0091] Based on the fused feature map, several target positions with a relatively high correlation can be determined from the target video frame, and the target position can correspond to the initial predicted bounding box for the target video frame.

[0092] For each initial prediction bounding box, the embodiment of the present disclosure can further determine the possibility that the initial prediction bounding box contains the target object based on the fused feature map, and the initial prediction bounding box with the highest possibility can be selected as the target bounding box. Specifically, it can be implemented according to the following steps:

[0093] Step 1: performing a first convolution operation on the fused feature map based on a first preset convolution parameter to obtain a first convolution feature map;

[0094] Step 2: Using the trained first aligned convolutional neural network and at least one initial predicted bounding box of the target video frame, an aligned convolution operation is performed on the first convolutional feature map to obtain the first convolutional feature map output by the network;

[0095] Step 3: Determine the probability value of each initial prediction bounding box in at least one initial prediction bounding box indicated by the first post-convolution feature map output by the network containing the target object;

[0096] Step 4: Select an initial prediction bounding box with the largest probability value from at least one initial prediction bounding box as the target bounding box containing the target object.

[0097] Here, the operation between the initial prediction bounding box and the fused feature map mainly refers to the alignment convolution operation between the first post-convolution feature map determined based on the fused feature map and the initial prediction bounding box. The aligned convolution operation can capture the geometric information and context information of the first post-convolution feature area corresponding to the initial prediction bounding box, so that the first post-convolution feature map output by the determined network can determine the probability value of the target object contained in each initial prediction bounding box.

[0098] For an initial predicted bounding box with a larger probability value of containing a target object, the position state information of the target object in the target video frame determined by the initial predicted bounding box is more accurate.

[0099] In the disclosed embodiment, an aligned convolution operation may be performed according to the following steps to determine a first post-convolution feature map output by a first aligned convolutional neural network.

[0100] Step 1: Determine the convolution kernel parameters corresponding to the trained first aligned convolutional neural network;

[0101] Step 2: Use the convolution kernel parameters to traverse the feature map after the first convolution;

[0102] Step 3. When traversing to the first convolution feature area corresponding to any initial prediction bounding box of the target video frame, convolution operation is performed on the convolution kernel indicated by the convolution kernel parameters and the first convolution feature area to obtain the first convolution feature map output by the network; wherein the first convolution feature area is included in the first convolution feature map and has the same size as the initial prediction bounding box.

[0103] Here, when the convolution kernel parameters corresponding to the first aligned convolutional neural network are determined, the convolution kernel parameters can be used to perform traversal convolution on the first convolution feature map. When traversing to the first convolution feature area corresponding to any initial prediction bounding box of the target video frame, the convolution kernel indicated by the convolution kernel parameters is convolved with the first convolution feature area. That is, sampling can be performed based on the initial prediction bounding box, and then a convolution operation can be performed.

[0104] In order to explain the above aligned convolution process, we can combine Figure 2 Provide an example.

[0105] When a 3*3 convolution kernel is determined for the first aligned convolutional neural network, the convolution kernel is used to perform traversal convolution on the first convolution feature map. Figure 2In the case of the first convolution feature area corresponding to the initial prediction bounding box shown, here, for the upper left vertex of the convolution kernel, the corresponding convolution can be the upper left vertex of the first convolution feature area, as shown in the arrow diagram, for other vertices of the convolution kernel, the convolution operation can be performed according to the sampling position shown in the figure.

[0106] Taking the traversal position (x, y) corresponding to the upper left vertex of a convolution kernel as an example, the distance from the current position to the left, top, right and bottom of the initial prediction bounding box is l', t', r', b'. The nine sampling positions of aligned convolution will be (x-l', y-t'), (x, y-t'), (x+r', y-t'), (x-l', y), (x, y), (x+r', y), (x-l', y+b'), (x, y+b'), (x+r', y+b'). It can be seen that by capturing the geometric information of the feature area after the first convolution corresponding to the initial prediction bounding box and the context information, the values ​​of each probability value in the determined probability matrix are more accurate, thereby making the determined target bounding box more accurate.

[0107] The target tracking method provided by the embodiment of the present disclosure can not only determine a more accurate target bounding box, but also adjust the target bounding box through a bounding box adjustment strategy to obtain an adjusted target bounding box, which can be specifically achieved through the following steps:

[0108] Step 1: performing a second convolution operation on the fused feature map based on a second preset convolution parameter to obtain a second convolution feature map;

[0109] Step 2: Using the trained second aligned convolutional neural network and the target bounding box containing the target object in the target video frame, an aligned convolution operation is performed on the second convolution feature map to obtain the second convolution feature map output by the network;

[0110] Step 3: Determine the adjustment amount for the target bounding box indicated by the second convolution feature map output by the network;

[0111] Step 4: Adjust the target bounding box based on the adjustment amount output by the network to obtain an adjusted target bounding box.

[0112] Here, the operation between the target bounding box and the fused feature map mainly refers to the alignment convolution operation between the second convolution feature map determined based on the fused feature map and the target bounding box. The alignment convolution operation can capture the geometric information and context information of the second convolution feature area corresponding to the target bounding box, so that the second convolution feature map output by the determined network can determine the adjustment amount of the target bounding box.

[0113] The target bounding box may be adjusted based on the determined adjustment amount to obtain an adjusted target bounding box.

[0114] Among them, the alignment convolution operation between the above-mentioned target bounding box and the second convolution feature map can refer to the above-mentioned alignment convolution operation process between the initial prediction bounding box and the first convolution feature map, which will not be repeated here.

[0115] It should be noted that although the second convolution operation here is also obtained on the fused feature map, since the second preset convolution parameters here are different from the first preset convolution parameters mentioned above, the feature values ​​of the second convolution feature map obtained by the second convolution operation and the first convolution feature map obtained by the first convolution operation are not the same.

[0116] In a specific application, considering that both the first convolution feature map and the second convolution feature map need to take into account the bounding box to perform the aligned convolution operation, the sizes of the first convolution feature map and the second convolution feature map can be the same, for example, both can be 25*25*256. Regarding the first preset convolution parameter and the second preset convolution parameter, the corresponding convolution kernels can be the same, for example, both can be 3*3, but the specific values ​​of the convolution kernels can be different.

[0117] In the embodiment of the present disclosure, when it is determined that a target bounding box containing a target object is marked in a target video frame to be tracked, the first aligned convolutional neural network and the second aligned convolutional neural network may be trained, which may be specifically achieved by the following steps:

[0118] Step 1: inputting at least one initial prediction bounding box generated for the target video frame and the first convolution feature map into the first aligned convolutional neural network to be trained, and obtaining the first convolution feature map output by the network; determining the probability value of each initial prediction bounding box in at least one initial prediction bounding box indicated by the first convolution feature map output by the network containing the target object; comparing the position information of the initial prediction bounding box with the largest probability value with the position information of the target bounding box marked for the target video frame to be tracked, and obtaining a first comparison result;

[0119] And, inputting the generated at least one initial prediction bounding box for the target video frame and the second convolution feature map into the second aligned convolution neural network to be trained, to obtain the second convolution feature map output by the network; determining the adjustment amount required for each initial prediction bounding box in the at least one initial prediction bounding box indicated by the second convolution feature map output by the network; adjusting each initial prediction bounding box according to the adjustment amount to obtain position information of the adjusted initial prediction bounding box, and comparing the position information of the target bounding box annotated for the target video frame to be tracked, to obtain a second comparison result;

[0120] Step 2: If any comparison result of the first comparison result and the second comparison result is inconsistent, adjusting the parameters of at least one of the first aligned convolutional neural network and the second aligned convolutional neural network.

[0121] Here, in the process of training the first aligned convolutional neural network and the second aligned convolutional neural network, on the one hand, when generating at least one initial prediction bounding box and a first convolutional feature map for the target video frame according to the above method, the generated at least one initial prediction bounding box and the first convolutional feature map for the target video frame can be input into the first aligned convolutional neural network to be trained to obtain the first convolutional feature map output by the network; determine the probability value of each initial prediction bounding box in at least one initial prediction bounding box indicated by the first convolutional feature map output by the network containing the target object; compare the position information of the initial prediction bounding box with the largest probability value with the position information of the target bounding box marked for the target video frame to be tracked, and obtain a first comparison result.

[0122] The first comparison result here indicates the comparison result of the network prediction of the location information of the initial predicted bounding box that is most likely to contain the target object and the actual annotated location information. If the comparison results are consistent, it means that the accuracy of the current bounding box prediction is high. If the comparison results are inconsistent, it means that the accuracy of the current bounding box prediction is poor. At this time, the network parameters can be adjusted to ensure that the prediction result can be close to the annotated true value result.

[0123] On the other hand, when at least one initial prediction bounding box and a second convolutional feature map for a target video frame are generated according to the above method, the generated at least one initial prediction bounding box and the second convolutional feature map for the target video frame can be input into a second aligned convolutional neural network to be trained to obtain a second convolutional feature map output by the network; the adjustment amount required for each initial prediction bounding box in the at least one initial prediction bounding box indicated by the second convolutional feature map output by the network is determined; the position information of the adjusted initial prediction bounding box obtained by adjusting each initial prediction bounding box according to the adjustment amount is compared with the position information of the target bounding box marked for the target video frame to be tracked, to obtain a second comparison result.

[0124] The second comparison result here indicates the comparison result between the position information of the adjusted target bounding box determined by the network and the actual marked position information. If the comparison results are consistent, it means that the current adjustment effect is good. If the comparison results are inconsistent, it means that the current adjustment effect is poor. At this time, the network parameters can be adjusted to ensure that the adjustment result can approach the true value result of the annotation.

[0125] The network parameters of the two aligned convolutional neural networks can be adjusted synchronously or individually, and can be constrained in combination with specific training conditions, and no specific restrictions are made here.

[0126] At the same time, in the embodiment of the present disclosure, the network parameters can also be adjusted based on the comparison result between the position information of the initial prediction bounding box output by the network and the annotated true value result to further improve the accuracy of the model prediction.

[0127] Considering that the target tracking method provided by the embodiment of the present disclosure not only involves the above two aligned convolutional neural networks, but also involves a feature extraction network for feature extraction and a cross-correlation network for realizing feature fusion, the actual training process may also be a joint training of the above multiple networks. Figure 3 Provide an example.

[0128] For the initial video frame and the target video frame, the corresponding first feature map and the second feature map can be first extracted through the feature extraction network. When the fused feature map is determined by using the cross-correlation network, the first convolution operation and the second convolution operation can be performed based on the first preset convolution parameter and the second preset convolution parameter, respectively, to obtain the first convolution feature map and the second convolution feature map.

[0129] For the first convolution feature map, the first aligned convolution neural network (i.e., the first aligned network) can be used to perform an aligned convolution operation based on the initial predicted bounding box, and the network output can be the classification score corresponding to the initial predicted bounding box, and the initial predicted bounding box with the highest classification score is selected as the target bounding box that is most likely to contain the target object. For the second convolution feature map, the second aligned convolution neural network (i.e., the second aligned network) can be used to perform an aligned convolution operation based on the target bounding box to obtain the adjusted target bounding box output by the network.

[0130] It can be seen that the embodiment of the present disclosure accurately captures the features of the bounding box through a two-step optimization mechanism and further optimizes the bounding box, thereby making the determined target state more accurate and the tracking effect better.

[0131] Those skilled in the art will appreciate that, in the above method of specific implementation, the order in which the steps are written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of the steps should be determined by their functions and possible internal logic.

[0132] Based on the same inventive concept, an embodiment of the present disclosure also provides a target tracking device corresponding to the target tracking method. Since the principle of solving the problem by the device in the embodiment of the present disclosure is similar to the above-mentioned target tracking method in the embodiment of the present disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0133] Reference Figure 4 FIG. 4 is a schematic diagram of a target tracking device provided by an embodiment of the present disclosure, wherein the device includes: an acquisition module 401, an extraction module 402, a fusion module 403 and a tracking module 404; wherein,

[0134] The acquisition module 401 is used to acquire an initial video frame marked with a target bounding box and a target video frame to be tracked; the target bounding box contains a target object;

[0135] An extraction module 402 is used to extract a first feature map and a second feature map from an initial video frame and a target video frame respectively;

[0136] A fusion module 403 is used to determine a fused feature map based on the first feature map and the second feature map, and generate at least one initial prediction bounding box for the target video frame based on the fused feature map;

[0137] The tracking module 404 is used to determine the target tracking result for the target video frame based on at least one initial predicted bounding box of the target video frame and the fused feature map; wherein the target tracking result includes a target bounding box containing the target object selected from the at least one initial predicted bounding box.

[0138] In the disclosed embodiment, when an initial video frame marked with a target bounding box and a target video frame to be tracked are obtained, feature extraction can be performed first, so that a fused feature map can be obtained based on the extracted first feature map and the second feature map. The fused feature map is fused with relevant information of the initial video frame and the target bounding box in the initial video frame, and is also fused with relevant information of the target video frame. In this way, at least one initial predicted bounding box for the target video frame can be generated based on the fused feature map, and the initial predicted bounding box can indicate the position where the target object may appear in the target video frame. In order to determine the target bounding box containing the target object in the target video frame, the fused feature map can be correlated with the initial predicted bounding box, that is, the fused feature map can be further used to screen the initial predicted bounding box, and then a more accurate position of the target object can be selected, so that the tracking effect is better.

[0139] In a possible implementation, the tracking module 404 is configured to determine a target tracking result for the target video frame according to at least one initial predicted bounding box and a fused feature map of the target video frame in the following steps:

[0140] Performing a first convolution operation on the fused feature map based on a first preset convolution parameter to obtain a first convolution feature map;

[0141] Using the trained first aligned convolutional neural network and at least one initial predicted bounding box of the target video frame, an aligned convolution operation is performed on the first convolutional feature map to obtain a first convolutional feature map output by the network;

[0142] Determine a probability value that each of at least one initial prediction bounding box indicated by a first post-convolution feature map output by the network contains a target object;

[0143] An initial prediction bounding box with a maximum probability value is selected from at least one initial prediction bounding box as a target bounding box containing the target object.

[0144] In a possible implementation, the tracking module 404 is configured to perform an aligned convolution operation on the first convolution feature map using the trained first aligned convolution neural network and at least one initial predicted bounding box of the target video frame according to the following steps to obtain the first convolution feature map output by the network:

[0145] Determine the convolution kernel parameters corresponding to the trained first aligned convolutional neural network;

[0146] Use the convolution kernel parameters to traverse the feature map after the first convolution;

[0147] When traversing to a first convolution feature region corresponding to any initial prediction bounding box of the target video frame, convolution operation is performed on the first convolution feature region with the convolution kernel indicated by the convolution kernel parameter to obtain a first convolution feature map output by the network;

[0148] The first convolution feature area is contained in the first convolution feature map and has the same size as the initial prediction bounding box.

[0149] In a possible implementation manner, the above device further includes:

[0150] The adjustment module 405 is used to:

[0151] Performing a second convolution operation on the fused feature map based on a second preset convolution parameter to obtain a second convolution feature map;

[0152] Using the trained second aligned convolutional neural network and the target bounding box containing the target object in the target video frame, an aligned convolution operation is performed on the second convolution feature map to obtain the second convolution feature map output by the network;

[0153] Determining an adjustment amount for the target bounding box indicated by a second post-convolution feature map output by the network;

[0154] The target bounding box is adjusted based on the adjustment amount output by the network to obtain an adjusted target bounding box.

[0155] In a possible implementation manner, a target bounding box containing a target object is marked in a target video frame to be tracked; and the above-mentioned device further includes:

[0156] The training module 406 is used to:

[0157] Inputting at least one initial prediction bounding box generated for the target video frame and the first convolution feature map into a first aligned convolutional neural network to be trained, obtaining a first convolution feature map output by the network; determining a probability value that each of the at least one initial prediction bounding box indicated by the first convolution feature map output by the network contains the target object; comparing the position information of the initial prediction bounding box with the largest probability value with the position information of the target bounding box marked for the target video frame to be tracked, obtaining a first comparison result;

[0158] And, inputting the generated at least one initial prediction bounding box for the target video frame and the second convolution feature map into the second aligned convolution neural network to be trained, to obtain the second convolution feature map output by the network; determining the adjustment amount required for each initial prediction bounding box in the at least one initial prediction bounding box indicated by the second convolution feature map output by the network; adjusting each initial prediction bounding box according to the adjustment amount to obtain position information of the adjusted initial prediction bounding box, and comparing the position information of the target bounding box annotated for the target video frame to be tracked, to obtain a second comparison result;

[0159] If any comparison result of the first comparison result and the second comparison result is inconsistent, parameters of at least one of the first aligned convolutional neural network and the second aligned convolutional neural network are adjusted.

[0160] In a possible implementation, the extraction module 402 is configured to extract the first feature map and the second feature map from the initial video frame and the target video frame respectively according to the following steps:

[0161] The trained feature extraction network is used to extract features from the initial video frame and the target video frame respectively to obtain a first feature map and a second feature map.

[0162] In a possible implementation manner, the above device further includes:

[0163] The updating module 407 is used to determine the target mask features corresponding to the initial video frame; wherein the target mask features of the video frame area within the target bounding box in the initial video frame are different from the target mask features of the video frame area outside the target bounding box; and the first feature map is updated based on the determined target mask features to obtain an updated first feature map.

[0164] In a possible implementation, the fusion module 403 is configured to determine a fused feature map based on the first feature map and the second feature map according to the following steps:

[0165] The trained cross-correlation network is used to perform cross-correlation operation on the first feature map and the second feature map to obtain a fused feature map.

[0166] In a possible implementation, the fusion module 403 is used to perform a cross-correlation operation on the first feature map and the second feature map using a trained cross-correlation network according to the following steps to obtain a fused feature map:

[0167] The first feature map is used as the convolution kernel corresponding to the trained cross-correlation network;

[0168] Using the convolution kernel to traverse the second feature map, and determining the convolution result at each traversal position;

[0169] The convolution results at each traversal position are combined to obtain a fused feature map.

[0170] For descriptions of the processing flow of each module in the device and the interaction flow between each module, reference may be made to the relevant descriptions in the above method embodiment, which will not be described in detail here.

[0171] The present disclosure also provides an electronic device, such as Figure 5 , which is a schematic diagram of the structure of an electronic device provided by an embodiment of the present disclosure, includes: a processor 501, a memory 502, and a bus 503. The memory 502 stores machine-readable instructions executable by the processor 501 (for example, Figure 4 In the device, the acquisition module 401, the extraction module 402, the fusion module 403, the tracking module 404 corresponding execution instructions, etc., when the electronic device is running, the processor 501 communicates with the memory 502 through the bus 503, and the machine-readable instructions are executed by the processor 501 to perform the following processing:

[0172] Acquire an initial video frame marked with a target bounding box and a target video frame to be tracked; the target bounding box contains the target object;

[0173] Extracting a first feature map and a second feature map from an initial video frame and a target video frame respectively;

[0174] Determine a fused feature map based on the first feature map and the second feature map, and generate at least one initial prediction bounding box for the target video frame based on the fused feature map;

[0175] According to at least one initial predicted bounding box of the target video frame and the fused feature map, a target tracking result for the target video frame is determined; wherein the target tracking result includes a target bounding box containing the target object selected from the at least one initial predicted bounding box.

[0176] The present disclosure also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the target tracking method described in the above method embodiment are executed. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0177] The present disclosure also provides a computer program product that carries a program code. The program code includes instructions that can be used to execute the steps of the target tracking method described in the above method embodiment. For details, please refer to the above method embodiment, which will not be repeated here.

[0178] The computer program product may be implemented in hardware, software or a combination thereof. In one optional embodiment, the computer program product is implemented as a computer storage medium. In another optional embodiment, the computer program product is implemented as a software product, such as a software development kit (SDK).

[0179] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the device or unit can be electrical, mechanical or other forms.

[0180] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0181] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0182] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0183] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The protection scope of the present disclosure is not limited thereto. Although the present disclosure is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed in the present disclosure, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. A method for target tracking, characterized in that: The method comprises: Acquire an initial video frame marked with a target bounding box and a target video frame to be tracked; the target bounding box contains a target object; Extracting a first feature map and a second feature map from the initial video frame and the target video frame respectively; Determine a fused feature map based on the first feature map and the second feature map, and generate at least one initial prediction bounding box for the target video frame based on the fused feature map; Determine a target tracking result for the target video frame according to at least one initial predicted bounding box of the target video frame and the fused feature map; wherein the target tracking result includes a target bounding box selected from the at least one initial predicted bounding box and containing the target object; The step of determining the target tracking result for the target video frame according to at least one initial predicted bounding box of the target video frame and the fused feature map includes: Performing a first convolution operation on the fused feature map based on a first preset convolution parameter to obtain a first convolution feature map; Using the trained first aligned convolutional neural network and at least one initial predicted bounding box of the target video frame, performing an aligned convolution operation on the first convolutional feature map to obtain a first convolutional feature map output by the network; Determine a probability value that each of the at least one initial prediction bounding box indicated by the first post-convolution feature map output by the network contains a target object; An initial prediction bounding box with a maximum probability value is selected from the at least one initial prediction bounding box as the target bounding box containing the target object.

2. The method according to claim 1, characterized in that: The method uses the trained first aligned convolutional neural network and at least one initial predicted bounding box of the target video frame to perform an aligned convolution operation on the first convolutional feature map to obtain a first convolutional feature map output by the network, including: Determine the convolution kernel parameters corresponding to the trained first aligned convolutional neural network; Using the convolution kernel parameters to traverse the first convolution feature map; When traversing to a first convolution feature region corresponding to any initial prediction bounding box of the target video frame, convolution operation is performed on the convolution kernel indicated by the convolution kernel parameter and the first convolution feature region to obtain a first convolution feature map output by the network; The first post-convolution feature area is contained in the first post-convolution feature map and has the same size as the initial prediction bounding box.

3. The method according to claim 1 or 2, characterized in that: The method further comprises: Performing a second convolution operation on the fused feature map based on a second preset convolution parameter to obtain a second convolution feature map; Using the trained second aligned convolutional neural network and the target bounding box containing the target object in the target video frame, an aligned convolution operation is performed on the second convolutional feature map to obtain a second convolutional feature map output by the network; Determining an adjustment amount for the target bounding box indicated by a second post-convolution feature map output by the network; The target bounding box is adjusted based on the adjustment amount output by the network to obtain an adjusted target bounding box.

4. The method according to claim 3, characterized in that The target video frame to be tracked is marked with a target bounding box containing the target object; the first aligned convolutional neural network and the second aligned convolutional neural network are trained according to the following steps: Inputting the generated at least one initial prediction bounding box for the target video frame and the first convolution feature map into a first aligned convolutional neural network to be trained, and obtaining a first convolution feature map output by the network; determining a probability value that each of the at least one initial prediction bounding box indicated by the first convolution feature map output by the network contains a target object; comparing the position information of the initial prediction bounding box with the largest probability value with the position information of the target bounding box marked for the target video frame to be tracked, and obtaining a first comparison result; And, inputting the generated at least one initial prediction bounding box for the target video frame and the second convolution feature map into a second aligned convolutional neural network to be trained, to obtain a second convolution feature map output by the network; determining the adjustment amount required for each initial prediction bounding box in the at least one initial prediction bounding box indicated by the second convolution feature map output by the network; adjusting each initial prediction bounding box according to the adjustment amount to obtain position information of the adjusted initial prediction bounding box, and comparing the position information of the target bounding box annotated for the target video frame to be tracked, to obtain a second comparison result; If any comparison result of the first comparison result and the second comparison result is inconsistent, parameters of at least one of the first aligned convolutional neural network and the second aligned convolutional neural network are adjusted.

5. The method according to claim 1 or 2, characterized in that: The extracting a first feature map and a second feature map from the initial video frame and the target video frame respectively includes: The trained feature extraction network is used to perform feature extraction on the initial video frame and the target video frame respectively to obtain a first feature map and a second feature map.

6. The method according to claim 1 or 2, characterized in that: The method further comprises: Determining a target mask feature corresponding to the initial video frame; wherein the target mask feature of the video frame region within the target bounding box in the initial video frame is different from the target mask feature of the video frame region outside the target bounding box; The first feature map is updated based on the determined target mask feature to obtain an updated first feature map.

7. The method according to claim 1 or 2, characterized in that: The determining a fused feature map based on the first feature map and the second feature map includes: The trained cross-correlation network is used to perform a cross-correlation operation on the first feature map and the second feature map to obtain a fused feature map.

8. The method according to claim 7, characterized in that The method of performing a cross-correlation operation on the first feature map and the second feature map using the trained cross-correlation network to obtain a fused feature map includes: Using the first feature map as the convolution kernel corresponding to the trained cross-correlation network; Using the convolution kernel to traverse the second feature map, and determine the convolution result at each traversal position; The convolution results at each traversal position are combined to obtain a fused feature map.

9. A target tracking device, characterized in that: The device comprises: An acquisition module, used to acquire an initial video frame marked with a target bounding box and a target video frame to be tracked; the target bounding box contains a target object; An extraction module, used to extract a first feature map and a second feature map from the initial video frame and the target video frame respectively; A fusion module, configured to determine a fused feature map based on the first feature map and the second feature map, and generate at least one initial prediction bounding box for the target video frame based on the fused feature map; A tracking module, configured to determine a target tracking result for the target video frame according to at least one initial predicted bounding box of the target video frame and the fused feature map; wherein the target tracking result includes a target bounding box selected from the at least one initial predicted bounding box and containing the target object; The tracking module is used to determine the target tracking result for the target video frame according to at least one initial predicted bounding box of the target video frame and the fused feature map according to the following steps: Performing a first convolution operation on the fused feature map based on a first preset convolution parameter to obtain a first convolution feature map; Using the trained first aligned convolutional neural network and at least one initial predicted bounding box of the target video frame, performing an aligned convolution operation on the first convolutional feature map to obtain a first convolutional feature map output by the network; Determine a probability value that each of the at least one initial prediction bounding box indicated by the first post-convolution feature map output by the network contains a target object; An initial prediction bounding box with a maximum probability value is selected from the at least one initial prediction bounding box as the target bounding box containing the target object.

10. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the steps of the target tracking method as described in any one of claims 1 to 8 are performed.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the target tracking method as claimed in any one of claims 1 to 8 are executed.