Electronic device and method for object tracking

Abstract

The present application provides an electronic device and method for object tracking, the method comprising: generating a first vector of a first scale object and a second vector of a second scale object of an object detection frame corresponding to a specific bounding box through a specific object model; and generating an identity label of an object in the object detection frame according to the first vector, the second vector, and M first scale reference vectors and M second scale reference vectors stored in an object vector database.

Description

Technical Field

[0001] This invention relates to an electronic device and method for object tracking. Background Technology

[0002] Generally speaking, existing object tracking technologies are mainly divided into two types. The first type is one-stage object tracking, which uses deep learning methods to directly train and generate a model that can perform object detection and object tracking simultaneously. The second type is two-stage object tracking, which separates the object detection and object tracking operations by designing an object detector and an object tracker independently. The results detected by the object detector are then tracked by the self-designed object tracker.

[0003] However, practical object tracking operations present numerous challenges, increasing the difficulty of object matching, with occlusion being the most problematic. When the image segmented by existing object detectors is processed by existing trackers, recognition becomes difficult because the object feature vectors entering the tracker only retain half of the original information and are prone to errors. Furthermore, current object tracking technologies cannot effectively address the occlusion issue, resulting in low accuracy in correctly identifying objects during tracking. Summary of the Invention

[0004] Therefore, in order to improve the problem of difficulty in identifying objects when encountering occlusion during object tracking, one of the objectives of this invention is to propose an electronic device and method that can simultaneously consider the features of the whole and parts (e.g., the whole body and half body) of an object as well as the correlation between multiple frames, optimize all detection results, improve the accuracy of object re-identification during object tracking, and improve the recall rate.

[0005] According to an embodiment of the present invention, a method for object tracking is provided, the method comprising: generating a first vector of a first-scale object and a second vector of a second-scale object of an object detection box corresponding to a specific frame through a specific object model; and generating an identity tag of an object within the object detection box based on the first vector, the second vector, and M first-scale reference vectors and M second-scale reference vectors stored in an object vector database.

[0006] According to an embodiment of the present invention, an electronic device for object tracking is provided, the electronic device comprising a storage circuit and a processing circuit. The storage circuit is used to store a specific object model and an object vector database. The processing circuit is coupled to the storage circuit and is used to: generate a first vector of a first-scale object and a second vector of a second-scale object of an object detection box corresponding to a specific frame, based on the specific object model; and generate an identification tag of an object within the object detection box based on the first vector, the second vector, and M first-scale reference vectors and M second-scale reference vectors stored in the object vector database. Attached Figure Description

[0007] Figure 1 This is a simplified flowchart illustrating a method for tracking moving objects according to an embodiment of this application.

[0008] Figure 2 This is a block diagram of an electronic device / circuit for object / body tracking according to an embodiment of this application.

[0009] Figure 3 This is a schematic diagram of the identity tagging results of existing object tracking technologies.

[0010] Figure 4 This is an example schematic diagram of the identity tagging results of object tracking according to an embodiment of the present invention.

[0011] Figure 5 This is an embodiment of the present invention. Figure 2 The diagram shows the operation flow of the electronic device / circuit used for object / body tracking.

[0012] [Symbol Explanation]

[0013] 100: Electronic devices

[0014] 101: Object Detection Device

[0015] 105: Storage Circuit

[0016] 110: Processing Circuit

[0017] D1: Reference Character Vector Database

[0018] M1: Object Model

[0019] ID4, ID5, ID6, ID7, ID45: Identity Tags

[0020] S10~S35, S500~S565: Steps Detailed Implementation

[0021] This invention aims to provide a technical solution, electronic device, and corresponding method for accurately identifying and tracking an object. Please refer to... Figure 1 , Figure 1 This is a simplified flowchart illustrating a method for tracking moving objects according to an embodiment of this application. If the same result can be achieved substantially, it is not necessary to follow the same procedure exactly. Figure 1 The steps in the illustrated process are performed sequentially, and Figure 1 The steps shown do not necessarily have to be performed consecutively; other steps can be inserted as well. The detailed process steps are explained below:

[0022] Step S10: Begin;

[0023] Step S15: Input or obtain an object detecting bounding box on a specific frame;

[0024] Step S20: Based on the object detection box, using a specific object detection model, generate and obtain a first vector of a first-scale object and a second vector of a second-scale object. The first and second scale objects refer to a first scale and a second scale of a specific moving object within the object detection box, respectively. For example, the first scale of the moving object is an overall scale, and the first vector is an overall scale vector of the object. The second scale of the moving object is a partial scale, and the second vector is a partial scale vector of the object. For example, when the specific object detection model is a person object detection model and the moving object is a person, the first scale of the person is a full-body scale, and the second scale of the person is, for example, a half-body scale, which can be an upper-body scale or a lower-body scale. The first vector of the person is a full-body scale vector of the person, and the second vector of the person is a half-body scale vector of the person. However, this is not a limitation of this application. In addition, the partial scale can be 1 / 3 or 1 / 4, and is not limited.

[0025] Step S25: Based on an object vector database (e.g., a person vector database), compare the first vector with the second vector to identify the moving object; wherein the person vector database stores, for example, the full-body proportion reference vectors and half-body proportion reference vectors of multiple different reference persons, and by comparing the full-body and half-body vectors, determine which person's full-body or half-body image the moving object's image is corresponding to or similar to.

[0026] Step S30: When a similar person is identified, output the identity of that similar person as the identity of the moving object, assign a corresponding identity label (or object number) to the object detection box to complete identity labeling, and add the image / vector within the object detection box to the object vector database as the image or vector of the similar person; and

[0027] Step S35: End.

[0028] For implementation details, please refer to [link / reference]. Figure 2 , Figure 2 This is a block diagram of an electronic device / circuit 100 for object / object tracking according to an embodiment of this application. The electronic device 100 is externally coupled to an object detection device / circuit 101 and includes a storage circuit 105 and a processing circuit 110. The storage circuit 105 is used to store the aforementioned specific object model M1 (e.g., a person object model) and the object vector database D1 (e.g., a reference person vector database).

[0029] Processing circuit 110 is coupled to storage circuit 105. Electronic device 100 receives an input object detection box from object detection device 101. Object detection device 101 is used to detect images of one or more frames, for example, to detect a moving object in a specific frame, and to outline the image of the moving object using a specific shape, such as a quadrilateral or rectangle. The resulting rectangle is the object detection box. For an image of a frame, since there may be one or more moving objects, object detection device 101 can generate one or more different object detection boxes for electronic device 100. After receiving one or more object detection boxes and their images, electronic device 100 assigns a corresponding identity label to each object detection as its identity as a moving object. In the case of multiple object detection boxes of the same moving object in different frames, electronic device 100 assigns the same identity label or object number to these object detection boxes. For multiple object detection frames of different moving objects, the electronic device 100 assigns different and unique identification tags or object serial numbers to these object detection frames. In addition, the identification tag can also be regarded as an object's serial number or sequential number, which is used to track the moving object in real time across different frames.

[0030] Since a moving object, such as a person (but not limited to, it can also be other types of moving objects, such as machines, vehicles or animals), may be partially occluded by certain objects, in order to achieve the effect of effectively tracking the moving object even when a part of the moving object is occluded by other moving or non-moving objects, taking a specific frame of object detection box and the moving object as a person as an example, each reference person in the reference person vector database D1 stored in the storage circuit 105 has a first vector of a first proportion object (e.g., a full-body proportion object image), a second vector of a second proportion object (e.g., a half-body proportion object image), and / or one or more vectors corresponding to object images of different viewpoints / rotation angles. For example, the reference person vector database stores M people. A person can have a full-body proportion reference vector as the first vector, a half-body proportion reference vector as the second vector, and / or one or more reference vectors of different rotation angles (corresponding to different turning / rotation images of the person). Furthermore, the number of M is not limited and can be updated and increased by the processing circuit 110 itself; detailed operation is described below.

[0031] In one embodiment, the full-body proportion vector and half-body proportion vector of a person are used as examples (but are not limited). The reference person vector database D1 records M full-body proportion reference vectors and M half-body proportion reference vectors of M different people. When the processing circuit 110 receives an input image of an object detection box of a specific frame (corresponding to an input person), the processing circuit 110 will perform vector calculation based on the image of the object detection box and a person detection model to generate a full-body proportion detection vector and a half-body proportion detection vector of the input person. For example, the processing circuit 110 will convert at least a part of the image of the input person into the full-body proportion detection vector and the half-body proportion detection vector. In addition, the implementation of the person detection model can adopt a person feature point model, a person distance difference model, or a person re-identification model, etc., which is not a limitation of the present invention.

[0032] Next, the processing circuit 110 performs comparisons of the full-body proportion vector and the half-body proportion vector. It compares the full-body proportion detection vector with M full-body proportion reference vectors of M different reference figures in the specific character model. For example, it calculates the vector distance between the full-body proportion detection vector and each of the M full-body proportion reference vectors, generating M full-body proportion vector distance values ​​corresponding to the M different reference figures. Similarly, the processing circuit 110 compares the half-body proportion detection vector with M half-body proportion reference vectors of the M different reference figures in the specific character model. For example, it calculates the vector distance between the half-body proportion detection vector and each of the M half-body proportion reference vectors, generating M half-body proportion vector distance values ​​corresponding to the M different reference figures. The calculation of the vector distance can be implemented using Euclidean distance (L2 distance), but is not limited to this method. In one embodiment, the processing circuit 110 selects the smallest full-body proportion vector distance value and its corresponding reference figure from the M full-body proportion vector distance values, and selects the smallest half-body proportion vector distance value and its corresponding reference figure from the M half-body proportion vector distance values, wherein the reference figures corresponding to the two vector distance values ​​may be the same or different. The processing circuit 110 compares the minimum full-body proportion vector distance value with a threshold value Tmatch. When the minimum full-body proportion vector distance value is less than the threshold value Tmatch, the processing circuit 110 can determine that, in terms of full-body proportion image, the input person is similar to or close to the full-body image of a reference person corresponding to the minimum full-body proportion vector distance value. Therefore, it can determine that the identity of the input person is the same as the identity of the reference person corresponding to the minimum full-body proportion vector distance value. Thus, the identity tag containing the input object detection box can be marked as the identity of the reference person corresponding to the minimum full-body proportion vector distance value, and the full-body proportion reference image or full-body proportion reference vector of the person corresponding to the minimum full-body proportion vector distance value can be updated to the image of the input person or its full-body proportion detection vector to achieve real-time tracking of the full-body image of the person.

[0033] When the minimum full-body proportion vector distance value is greater than or equal to the threshold value Tmatch, the processing circuit 110 can determine that, in terms of full-body proportion image, the input person is not close to the full-body image of the reference person corresponding to the minimum full-body proportion vector distance value. Therefore, it will then perform a half-body proportion judgment to determine whether the current image of the input person is close to the result image of a reference person after partial occlusion. At this time, the processing circuit 110 compares the minimum half-body proportion vector distance value with another threshold value Tmu. When the minimum half-body proportion vector distance value is less than the threshold value Tmu, the processing circuit 110 can determine that, in terms of half-body proportion image, the input person is not close to the reference person's full-body image. If the input person is similar to or close to the half-body image of a reference person corresponding to the minimum half-body ratio vector distance value, it can be determined that the identity of the input person is the same as the identity of the reference person corresponding to the minimum half-body ratio vector distance value. Therefore, the identity label containing the input object detection box can be marked as the identity of the reference person corresponding to the minimum half-body ratio vector distance value, and the half-body ratio reference image or half-body ratio reference vector of the person corresponding to the minimum half-body ratio vector distance value can be updated to the half-body image of the input person or its half-body ratio detection vector to achieve real-time tracking of the half-body image of the person.

[0034] Furthermore, in another embodiment, when determining whether to use a full-body proportion image or a half-body proportion image, the Intersection-over-Union (IoU) calculation can be used to assist in the determination; the following explanation uses the determination of a full-body proportion image as an example. For instance, in one embodiment, the processing circuit 110 can also generate the proportion of a portion of the input person to the whole based on the person detection model, such as the half-body proportion of the input person, and the processing circuit 110 can be used to generate a velocity prediction box corresponding to the object detection box within the specific image frame. In implementation, the processing circuit 110 calculates speed based on multiple object detection boxes corresponding to one or more identical characters in multiple previous frames to generate one or more speed prediction boxes. Taking a speed prediction box as an example, the processing circuit 110 can calculate an intersection-union distance (IU / R) between the current object detection box in the specific frame and the speed prediction box based on the half-body proportion of the input character, the full-body proportion detection vector of the input character, and the reference character vector database. When the IU / R value is larger, it indicates that the current object detection box is closer to the speed prediction box; conversely, when the IU / R value is smaller, it indicates that the current object detection box is farther away from the speed prediction box. Furthermore, regarding the full-body proportion vector (and the same applies to the half-body proportion vector), the smaller the aforementioned minimum full-body proportion vector distance value, the closer the person is to a reference person corresponding to the minimum full-body proportion vector distance value in the reference person database. Therefore, the processing circuit 110 of this application generates M adjusted full-body proportion vector distance values ​​based on a specific weight value, the intersection-union distance, and each of the M full-body proportion vector distance values. Each adjusted full-body proportion vector distance value can be expressed by the following formula:

[0035] dm′=(1-d_iou)×α+dm×(1-a)

[0036] Where dm′ is an adjusted full-body proportion vector distance value, d_iou is the intersection-union ratio distance, α is the specific weight value which can be adjusted by the user, and dm is each full-body proportion vector distance value. In this embodiment, when the smallest adjusted full-body proportion vector distance value among the M adjusted full-body proportion vector distance values ​​is less than the specific threshold Tmatch, the processing circuit 110 will determine that the input character is similar to or close to a reference character corresponding to the smallest adjusted full-body proportion vector distance value, and mark the identity of the input character as the identity of the reference character, which is implemented by marking the label of the character detection box as the identity of the reference character.

[0037] In another embodiment, after the processing circuit 110 assigns a new identity tag to an input person, the processing circuit 110 does not immediately update the input person to the reference person vector database. In implementation, the processing circuit 110 updates the input person to the reference person vector database only after detecting N consecutive images of the input person with the same identity tag, in order to avoid misjudgment; where N is, for example, equal to 3, but is not limited to.

[0038] In implementation, the storage circuit 105 is further used to store a temporary object vector database, such as a temporary character vector database. This temporary character vector database stores the full-body proportion reference vectors and half-body proportion reference vectors of characters that currently or recently appear consecutively in at least one frame. For example, when the processing circuit 110 determines that the minimum adjusted full-body proportion vector distance value is not less than a specific threshold Tmatch and the minimum half-body proportion vector distance value is not less than the threshold Tmu, the processing circuit 110 will then compare the temporary character vector database. For example, the temporary character vector database currently stores the full-body proportion reference vectors of K different temporary characters and the number of consecutive appearances. The processing circuit 110 calculates a vector distance value between the full-body proportion reference vectors of the K different temporary characters and the full-body proportion detection vector of the input character, and then calculates the distance value between the K vector distance values. The system finds the smallest vector distance value and its corresponding temporary character, and determines whether the smallest vector distance value is less than a temporary threshold Ttep. If the smallest vector distance value is less than the temporary threshold Ttep, the processing circuit 110 increments the corresponding consecutive occurrence count by 1, and then compares whether the updated consecutive occurrence count is greater than or equal to N (the value of N is, for example, 3, but not limited). If the updated consecutive occurrence count is greater than or equal to N, the processing circuit 110 can determine that the corresponding temporary character has appeared consecutively in N frames. Then, the temporary character and its corresponding full-body proportion reference vector are added and updated to the reference character vector database. That is, in this case, the temporary character is determined to be a referenceable character object, and the identity tag or object number corresponding to the temporary character is also updated to the reference character vector database.

[0039] If the minimum vector distance value is determined to be not less than a temporary threshold Ttep, it means that the input character corresponding to the minimum vector distance value does not exist in the temporary character vector database nor in the reference character vector database. In this case, the processing circuit 110 will add the input character to the temporary character vector database and assign a different and unique identity tag or object number to the input character, indicating that the input character is a new and different character. If the minimum vector distance value is determined to be less than the temporary threshold Ttep, and the number of consecutive occurrences after the update is less than N, it means that the input character is similar to a temporary character in the temporary character vector database (that is, the input character may already exist in the temporary character vector database), but the number of consecutive occurrences has not exceeded N. Therefore, the processing circuit 110 will not update the information of that temporary character to the character vector database. In this way, excessively frequent updates to the reference character vector database can be avoided, thus preventing misjudgments.

[0040] In another embodiment, if two or more vector distance values ​​(or two or more adjusted full-body proportion vector distance values) are all less than the specific threshold Tmatch, the processing circuit 110 will use a specific optimization algorithm (e.g., the Hungarian algorithm, but not limited thereto) to calculate multiple combinations of person objects with different loss values, find the combination of person objects with the minimum loss, and check the pairing comparison results of individual person object combinations among all person object combinations. For example, if the vector distance value corresponding to the person object combination with the minimum loss is greater than or equal to the threshold TDLOSS, the processing circuit 110 will determine that the input person is an unknown identity (i.e., its identity does not appear in the reference person vector database) and assign a new, different identity label or object number to the input person.

[0041] If the vector distance value corresponding to the group of human objects with the minimum loss is less than the threshold TDLOSS, the processing circuit 110 will then determine whether the half-body ratio value of the image corresponding to the group of human objects with the minimum loss is less than a specific threshold Th. If the half-body ratio value is less than the threshold Th, the processing circuit 110 will update the half-body ratio vector corresponding to the vector distance value of the group of human objects with the minimum loss to the half-body ratio vector of a reference human in the human vector database corresponding to the vector distance value of the group of human objects with the minimum loss, replacing the half-body ratio vector of the reference human.

[0042] To help readers understand the effectiveness of the embodiments of the present invention, the following is provided: Figure 3 and Figure 4A comparison is provided between the object tracking identification results of existing technologies and the object tracking identification results of embodiments of the present invention. For example... Figure 3 As shown, the sequence from left to right is chronological. Existing object tracking technology labels the four object detection boxes of four different people in the left frame as ID5, ID4, ID6, and ID7 at time point t(n). The actual images of the people in these four object detection boxes are all in a seated position. However, as time progresses, in subsequent frames, such as the frame at time point t(n+k), the actual images of two of these people are still seated, while the actual images of the other two people change to standing. Figure 3 As shown in the right-hand frame, while existing object tracking technologies can still correctly label the detection boxes of two people maintaining a sitting posture with identity tags ID4 and ID6, they cannot overcome the problem of different perspectives, postures, or occlusions. This means that even if two other people only change from a sitting posture to a standing posture and do not become completely different people, existing object tracking technologies will label the detection boxes of these two people with different identity tags ID45 (different from ID5) and ID67 (different from ID7).

[0043] The effects of the embodiments of the present invention, such as Figure 4 As shown, the sequence from left to right is chronological. In this embodiment of the invention, at time point t(n), the four object detection boxes for the four different people in the left frame are labeled as ID5, ID4, ID6, and ID7, respectively. The actual images of the people in these four object detection boxes are all in a seated position. As time progresses, in subsequent frames, such as the frame at time point t(n+k), the actual images of two of these people are still seated, while the actual images of the other two people change to standing. Figure 4 As shown in the right-hand frame, this embodiment of the invention can still correctly assign the correct identity tags ID5 and ID7 to two characters changing different poses, without misjudging them as completely different characters due to differences in perspective, pose, or occlusion. Therefore, compared to Figure 3 Compared with existing technologies, the embodiments of the present invention can significantly improve the accuracy of real-time tracking of moving objects (such as people) and enhance the effectiveness of monitoring systems.

[0044] In addition, to make the detailed embodiments and operation sequence of the present invention clearer to the reader, this application provides Figure 5 The process will be explained separately. Figure 5 This is an embodiment of the present invention. Figure 2 The diagram shows the operation flowchart of the electronic device / circuit used for object / physical tracking. If a similar result can be achieved, it is not necessary to follow it exactly. Figure 5 The steps in the illustrated process are performed sequentially, and Figure 5 The steps shown do not necessarily have to be performed consecutively; other steps can be inserted as well. The detailed process steps are explained below:

[0045] Step S500: Begin;

[0046] Step S505: Receive an input object detection box;

[0047] Step S510: Based on the image of the input object detection box and according to a person detection model, perform vector calculation to generate a full-body proportion detection vector and a half-body proportion detection vector of the input person.

[0048] Step S515: Compare the full-body proportion detection vector with the M full-body proportion reference vectors of M different reference characters in the specific character model, calculate the vector distance between the full-body proportion detection vector and the M full-body proportion reference vectors, and generate M full-body proportion vector distance values.

[0049] Step S520: Compare the half-body proportion detection vector with the M half-body proportion reference vectors of M different reference characters in the specific character model, calculate the vector distance between the half-body proportion detection vector and the M half-body proportion reference vectors, and generate M half-body proportion vector distance values.

[0050] Step S525: Calculate the intersection-union distance between the current input object detection box and the velocity prediction box;

[0051] Step S530: Based on the intersection-union distance and the M whole-body proportion vector distance values, generate M adjusted whole-body proportion vector distance values, and select the smallest adjusted whole-body proportion vector distance value and its corresponding reference figure.

[0052] Step S535: From the M half-body proportion vector distance values, select the smallest half-body proportion vector distance value and its corresponding reference figure;

[0053] Step S540: Determine whether the minimum adjusted full-body proportion vector distance value is less than the critical value Tmatch; if the value is less than the critical value Tmatch, the process will proceed to step S545, and if the value is greater than or equal to the critical value Tmatch, the process will proceed to step S550.

[0054] Step S545: Determine that the input person is not an unknown person, but a reference person corresponding to the minimum adjusted full-body proportion vector distance value, and mark the image corresponding to the input person with the identity label of the corresponding reference person as the identity label of the input person;

[0055] Step S550: Determine whether the minimum half-body proportion vector distance value is less than the critical value Tmu; if the value is less than the critical value Tmu, the process will proceed to step S555, and if the value is greater than or equal to the critical value Tmu, the process will proceed to step S560.

[0056] Step S555: Determine that the input person is not an unknown person, but a reference person corresponding to the smallest half-body ratio vector distance value, and mark the image corresponding to the input person with the identity label of the corresponding reference person as the identity label of the input person;

[0057] Step S560: Determine that the input character does not exist in the reference character vector data, and perform a comparison with the temporary character database; and

[0058] Step S565: End.

[0059] In summary, the object tracking method of the present invention can successfully re-identify the correct identity of an object even when it is significantly occluded, and can also take into account the correlation between video frame images, thus greatly improving performance and stability.

[0060] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made in accordance with the claims of the present invention should be included within the scope of the present invention.

Claims

1. A method for object tracking, characterized by, Includes: Using a specific object model, generate a first vector of a first scale object and a second vector of a second scale object corresponding to an object detection box of a specific frame; as well as Based on the first vector, the second vector, and M first proportional reference vectors and M second proportional reference vectors stored in an object vector database, an identity label for an object within the object detection box is generated. The step of generating the identity label of the object within the object detection box further includes: Based on the object detection bounding box and a velocity prediction bounding box of the specific frame, determine an intersection-over-conclusion distance; and Based on the first vector, the second vector, the intersection-union distance, and the M first proportional reference vectors and M second proportional reference vectors, the identity label of the object within the object detection box is generated.

2. The method of claim 1, wherein, The first scale object corresponds to at least one first part of the object, and the second scale object corresponds to at least one second part of the object, wherein the at least one first part includes the at least one second part.

3. The method of claim 1, wherein, The first vector of the first proportional object is a full-body object vector of the object, and the second vector of the second proportional object is a half-body object vector of the object.

4. The method of claim 1, wherein, Also includes: Based on the first vector and the M first proportional reference vectors, calculate the distance between the M first vectors; Based on a specific weight value, the distances of the M first vectors are weighted and calculated with the intersection-union ratio distance to generate M first adjusted vector distances, and a specific first adjusted vector distance is selected from the M first adjusted vector distances. Based on the second vector and the M second proportional reference vectors, calculate the distances between the M second vectors, and select a specific second adjusted vector distance from the M second vector distances; and The identity label of the object within the object detection box is determined based on the specific first adjusted vector distance and the specific second adjusted vector distance.

5. The method as described in claim 4, characterized in that, The specific first adjusted vector distance is the smallest first adjusted vector distance among the M first adjusted vector distances, and the specific second adjusted vector distance is the smallest second adjusted vector distance among the M second adjusted vector distances.

6. The method as described in claim 5, characterized in that, Also includes: When the minimum first adjusted vector distance is greater than a first threshold and the minimum second adjusted vector distance is less than a second threshold, the identity label of a specific object in the object vector database corresponding to the minimum first adjusted vector distance is used as the identity label of the object in the object detection box.

7. The method as described in claim 6, characterized in that, Also includes: When the minimum first adjusted vector distance is greater than the first critical value and the minimum second adjusted vector distance is greater than the second critical value: Based on the first vector and K first proportional reference vectors stored in a temporary vector database, calculate K third vector distances, and select a specific third vector distance from the K third vector distances; and Determine whether the calculated distance of the selected specific third vector is less than a third threshold value to decide whether the temporary object corresponding to the specific third vector distance in the temporary vector database is close to the object in the object detection box.

8. The method as described in claim 7, characterized in that, Also includes: When at least N consecutive specific third vector distances are all less than the third threshold, it is determined that the temporary object in the temporary vector database corresponding to the N consecutive specific third vector distances is similar to the object in the object detection box, a specific identity label is assigned to the temporary object, and the temporary object and the specific identity label are stored in the object vector database. as well as When there are no consecutive N specific third vector distances that are all less than the third threshold, assign a specific identity label to a specific object corresponding to the latest specific third vector distance, and store the specific object and the specific identity label in the temporary vector database, instead of storing them in the object vector database.

9. An electronic device for object tracking, characterized in that, Includes: A storage circuit for storing a specific object model and an object vector database; and A processing circuit, coupled to the storage circuit, is used to: Using this specific object model, a first vector of a first-scale object and a second vector of a second-scale object are generated for an object detection box corresponding to a specific frame. as well as Based on the first vector, the second vector, and the M first proportional reference vectors and M second proportional reference vectors stored in the object vector database, an identity label for an object within the object detection box is generated. The identity label of the object within the object detection box includes: Based on the object detection bounding box and a velocity prediction bounding box of the specific frame, determine an intersection-over-conclusion distance; and Based on the first vector, the second vector, the intersection-union distance, and the M first proportional reference vectors and M second proportional reference vectors, the identity label of the object within the object detection box is generated.

Images