Hybrid target tracking method adaptive to complex background and related device

Abstract

The invention provides a mixed target tracking method adaptive to a complex background and a related device. The method comprises the following steps: initializing a first tracking model and a second tracking model by using a video image sequence and an initial position and size of a target; checking whether the second tracking model is initialized or not before tracking the video frame, if not, continuing to use the first tracking model to track, and if yes, switching the second tracking model as a working model to perform normal target tracking; inputting a target tracking position filtering response diagram to the working model, judging whether the target is lost or not, and if yes, setting a tracking state as target lost; and in the target tracking and losing state, expanding the search range, initializing the first tracking model to retrieve the target again, and after the target is successfully searched, converting the tracking state into the normal target tracking state. According to the method, the advantages of the two tracking models are fully exerted, the tracking target can be quickly locked, and the tracking target can be stably tracked under the complex background.

Description

Technical Field

[0001] The present invention belongs to the technical field of target tracking, and particularly relates to a hybrid target tracking method adapted to complex backgrounds and related devices. Background Art

[0002] Tracking methods based on correlation filtering have good tracking effects and fast calculation efficiency, and are one of the mainstream target tracking technologies today.

[0003] The correlation filtering tracking method uses image filtering technology to track targets. It extracts manually designed features such as color histograms from the target region images, converts the time-domain convolution of the target features into a frequency-domain dot product through the fast Fourier transform, which can greatly improve the calculation efficiency, and uses kernel function technology to transform the target features into a higher-dimensional space, which can enhance the adaptability of the tracking model to target changes.

[0004] The DSST tracking algorithm and the ECO-HC tracking algorithm are two correlation filtering tracking methods widely used in the industry.

[0005] The DSST algorithm realizes the adaptation to target scale changes on the basis of the benchmark correlation filtering tracking method. The specific approach is to first use a traditional position tracking method to locate the position of the target in the new frame, then use the trained scale classifier at the located position to obtain the scale of the new frame, and at the same time update the filter parameters with the new position and new scale, and repeat this process to continuously track. The DSST tracking algorithm is very fast and can reach 200 frames per second on a conventional hardware platform.

[0006] The ECO-HC algorithm uses multi-channel deep learning convolutional features + histogram of oriented gradients features + color features, which greatly improves its tracking effect. At the same time, the ECO algorithm has done a lot of optimization work in terms of model size, training set size, and model update, making it have high calculation efficiency. ECO-HC only uses histogram of oriented gradients features + color features and has higher calculation efficiency with little reduction in effect, and can reach 60 frames per second on a conventional hardware platform.

[0007] Although the speed of the DSST tracking algorithm can reach 200 frames per second and it can be initialized within 50 milliseconds to quickly lock the target, it is easily interfered by complex backgrounds and loses the target.

[0008] The speed of the ECO-HC tracking algorithm can reach 60 frames per second, and it has strong adaptability to complex backgrounds. However, the initialization of the algorithm requires 500 - 1000 milliseconds, which is intolerable in scenarios where the target needs to be quickly locked. For example, on an airborne moving platform or when the target is moving rapidly, the target has long deviated far from its original position after such a long initialization time and it is no longer possible to track the target.

[0009] Conventional target tracking algorithms including DSST and ECO-HC do not have the ability to identify the target tracking state and retrieve the target again. Summary of the Invention

[0010] The object of the present invention is to provide a hybrid target tracking method and related device adapted to complex backgrounds to solve the problem that the prior art does not have the ability to identify the target tracking state and retrieve the target again.

[0011] To achieve the above object, the present invention adopts the following technical solutions: In the first aspect, the present invention provides a hybrid target tracking method adapted to complex backgrounds, including: Initializing a first tracking model and a second tracking model with a video image sequence and the initial position and size of the target. After the initialization of the first tracking model is completed, it is set as the working model for target tracking. The first tracking model is used for rapid initialization, and the second tracking model is used for target tracking in complex backgrounds; Before tracking the video frame, check whether the initialization of the second tracking model is completed. Otherwise, continue to use the first tracking model for tracking. If it is completed, switch the second tracking model to the working model for normal target tracking; Input the target tracking position filtering response map into the working model to determine whether the target is lost. If so, set the tracking state as the target is lost; In the state that the target is lost, expand the search range to initialize the first tracking model to retrieve the target again. After the target search is successful, change the tracking state to normal target tracking.

[0012] Further, the initialization of the first tracking model and the second tracking model with the video image sequence and the initial position and size of the target includes: Position search range coefficient , with a value of 3-9, that is, the search range is 3-9 times the target size; the first frame image of the target is , the initial position is and the size ; output the initial filter model ( , ); According to the target position and the size and the position search range coefficient , extract the multi-channel position filtering image features and the target position response map from the image , initialize the position filter model ( , ), and the calculation is as follows: Frequency domain position feature map:

[0013] Frequency domain position response diagram:

[0014] Position filter model:

[0015] .

[0016] Furthermore, after the initialization of the first tracking model is completed, it is set as the working model for target tracking, including: Using the same parameters as the initialization of the filter model, input the current frame image , the position of the target in the previous frame ; the position filter model of the previous frame ( , ); output the position of the target in the current frame ; According to the position of the previous frame and the scale as well as the position search range coefficient , extract the multi-channel position filter image features from the image , calculate: Frequency domain position feature map:

[0017] Frequency domain position response diagram:

[0018] Inverse Fourier transform to calculate the time domain response diagram: , from find the peak point as the new target position , and the position confidence .

[0019] Furthermore, before tracking the video frame, check whether the second tracking model is initialized. Otherwise, continue to use the first tracking model for tracking. If so, switch the second tracking model to the working model for normal target tracking, including: In the initial stage of tracking, use the first tracking model to track the target while the second tracking model is being initialized; before tracking the video frame, check whether the second tracking model is initialized. If not, continue to use the first tracking model for tracking. If it has been completed, set the second tracking model as the working model. For those who have switched to the second tracking model, continue to use the second tracking model for tracking; Use the working model to track the video frame, obtain the target position and scale, and output the target tracking box. If the tracking is successful, update the tracking model with the new target position and scale.

[0020] ​Further, the specific update method: Input the current frame image , the current frame target position and the size is ; the position filter model of the previous frame ( , ); output the new position filter model ( , ); According to the current frame position and the size as well as the position search range coefficient , extract the multi-channel position filter image features from the image , and calculate the frequency domain position feature map of the current frame , the frequency domain position response map , the frequency domain position response map ; Update the target tracking model:

[0021] .

[0022] Further, input the target tracking position filter response map into the working model, and determine whether the target is lost. If so, set the tracking status to target lost, including: Tracking confidence threshold , with a value range of 0.1 to 0.15; peak signal-to-noise ratio threshold , with a value range of 10 to 15; tracking confidence decline ratio threshold , with a value range of 5 to 6; Input the target tracking response map , and output: whether the target is lost; Take the pixel value of the peak point of the position tracking response map as the target tracking confidence ; Record the target tracking confidence of the previous frame and the target tracking confidence of the current frame , and calculate the tracking confidence decline rate ; Set the pixel values in the peak neighborhood of the tracking response map to 0 to obtain the background response map , calculate the mean and variance of the background response map , and calculate the peak signal-to-noise ratio ; If the conditions are met: < > , it is determined that the target is lost.

[0023] Furthermore, in the state of target loss, the search range is expanded to initialize the first tracking model to retrieve the target again. After the target search is successful, the tracking state is changed to normal target tracking, including: First, expand the search range to initialize the first tracking model, use the first tracking model to search for the position and scale of the tracking target. If the search and tracking are successful, set the normal tracking state; re-initialize the hybrid tracking model with the successfully tracked target box and the corresponding video frame, and enter the normal tracking state; Input the image successfully tracked in the previous frame and its target position and the size is , and use a larger search range coefficient , The value ranges from 9 to 11, that is, the search range is 9 to 11 times the target size; Continuously call the filter tracking algorithm. If the tracking fails for 100 consecutive frames, the tracking of the target is terminated; if the previous step of tracking is successful, output the target tracking box and enter the normal tracking state.

[0024] In a second aspect, the present invention provides a hybrid target tracking system adapted to complex backgrounds, including: A model initialization module for initializing the first tracking model and the second tracking model with a video image sequence and the initial position and size of the target. After the initialization of the first tracking model is completed, it is set as the working model for target tracking. The first tracking model is used for quick initialization, and the second tracking model is used for target tracking in complex backgrounds; A model update module for checking whether the second tracking model is initialized before tracking a video frame. Otherwise, continue to use the first tracking model for tracking. If so, switch the second tracking model to the working model for normal target tracking; A tracking state judgment module for inputting the filtered response map of the target tracking position to the working model to judge whether the target is lost. If so, set the tracking state to target loss; A target retrieval and output module for expanding the search range to initialize the first tracking model to retrieve the target again in the state of target loss. After the target search is successful, change the tracking state to normal target tracking.

[0025] In a third aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-mentioned hybrid target tracking method adapted to complex backgrounds.

[0026] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, characterized in that when the computer program is executed by a processor, the steps of the method for hybrid target tracking adapted to complex backgrounds are implemented.

[0027] Compared with the prior art, the present invention has the following technical effects: The present invention realizes a fast tracking algorithm adapted to complex backgrounds, which has the three capabilities of the above practical tracking algorithms: fastness, adaptability, and long-term performance, and solves the following technical problems.

[0028] Specifically, a hybrid tracking algorithm is realized by combining the first tracking model and the second tracking model. In the algorithm initialization stage, the first tracking model and the second tracking model are initialized in parallel, and the first tracking model is used to track the target in the initialization stage. After the second tracking model is initialized, the tracking is switched to the second tracking model. In this way, the advantages of the two tracking models are fully utilized, which can not only quickly lock the tracking target but also stably track the target in complex backgrounds.

[0029] Furthermore, on the basis of the hybrid tracking of the first tracking model and the second tracking model of the present invention, the target tracking position filtering response map is input into the working model to determine whether the target is lost. If so, the tracking state is set to target lost; in the target lost state, the search range is expanded to initialize the first tracking model to retrieve the target again. After the target search is successful, the tracking state is changed to normal target tracking, realizing the target lost judgment and re-retrieval mechanism, timely discovering that the target is lost, and starting the target search mechanism to retrieve the target again, so that the algorithm has the long-term tracking ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is the overall flowchart of the target tracking algorithm of the present invention.

[0031] Figure 2 It exemplifies the schematic diagram of obtaining the target position using the response map of image convolution filtering.

[0032] Figure 3 It is the initialization flowchart of the hybrid tracking model.

[0033] Figure 4 It is the tracking and updating flowchart of the hybrid tracking model.

[0034] Figure 5 It is the flowchart for judging the target tracking state.

[0035] Figure 6 The left figure is the schematic diagram of the normal target tracking state, the tracking confidence Fmax and the peak signal-to-noise ratio, and the right figure shows that the target is lost, and the tracking confidence Fmax and the peak signal-to-noise ratio PSR are shown.

[0036] Figure 7 It is the flowchart for the search and tracking model to retrieve the target again. Specific implementation manners

[0037] The present invention will be further described below in conjunction with the accompanying drawings: Example 1, please refer to Figure 1 , the present invention provides a hybrid target tracking method adapted to complex backgrounds, including: Initializing a first tracking model and a second tracking model with a video image sequence and the initial position and size of the target. After the initialization of the first tracking model is completed, it is set as the working model for target tracking. The first tracking model is used for rapid initialization, and the second tracking model is used for target tracking in complex backgrounds; Before tracking a video frame, check whether the second tracking model is initialized. Otherwise, continue to use the first tracking model for tracking. If so, switch the second tracking model to the working model for normal target tracking; Input the target tracking position filtering response map into the working model to determine whether the target is lost. If so, set the tracking status as the target being lost; In the state that the target is lost, expand the search range to initialize the first tracking model to retrieve the target again. After the target search is successful, convert the tracking status to normal target tracking.

[0038] Example 2, the present invention provides a hybrid target tracking method adapted to complex backgrounds, including: Taking the first tracking model as the DSST model and the second tracking model as the ECO-HC as an example: Figure 1 This is the overall process of the target tracking algorithm of the present invention, including three main steps: the hybrid tracking model normally tracks the target, the target tracking status is judged, and the search and tracking model retrieves the target again.

[0039] When the tracking status is normal tracking, use the hybrid tracking model combining ECO-HC and DSST to track the target of the video frame, and output the tracking frame when the target is not lost.

[0040] When the target changes rapidly or is occluded, the target tracking may fail. The target tracking status judgment module discovers this situation in time and converts the tracking status to the state that the target is lost.

[0041] When the tracking status is that the target is lost, use the search and tracking model to retrieve the target again. It expands the tracking range to search for the target. If the target search is successful, convert the tracking status to normal tracking.

[0042] This invention focuses on describing the hybrid tracking model. Both the DSST and ECO-HC object tracking models are based on correlation filters. For the algorithm details of the initialization, tracking, and update of these two tracking models, please refer to the relevant papers.

[0043] For correlation filter tracking, first, a filter model of the target is established. Then, the filter is used to perform a convolution operation on the image area to be tracked to obtain a filtered response map. Finally, the peak point is searched in the filtered response map as the target tracking position. According to the properties of the Fourier transform, convolution in the time domain corresponds to point multiplication in the frequency domain. Therefore, this property can be used to significantly reduce the computational complexity of image filtering. Specifically, the time-domain image patch and the filter are transformed into the frequency-domain image patch PATCH = FFT(patch) and the filter FILTER = FFT(filter) through the fast Fourier transform. The frequency-domain response RESPONSE = PATCH ⊙ FILTER is obtained by point multiplying the frequency-domain image patch PATCH and the filter FILTER. Then, the frequency-domain response RESPONSE is inverse Fourier transformed back to the time domain response = IFFT(RESPONSE). Finally, the peak point is searched in the filtered response map response as the target tracking position. Figure 2 An example of obtaining the target position using the response map of image convolution filtering is shown. The right figure is the filtered response map.

[0044] Initialization of the target correlation filter model Parameter: Position search range coefficient , with a value range of 3 to 9, that is, the search range is 3 to 9 times the target size.

[0045] Input: The first frame image of the target is , and the initial position is and the size .

[0046] Output: The initial filter model ( , ).

[0047] 1) According to the target position and the size and the position search range coefficient , extract the multi-channel position filtered image feature and the target position response map from the image , initialize the position filter model ( , ), and the calculation is as follows: Frequency-domain position feature map:

[0048] Frequency-domain position response map:

[0049] Position filter model:

[0050]

[0051] Target correlation filtering model tracking Parameters: Use the same parameters as the initialization of the correlation filtering model.

[0052] Input: Current frame image , the position of the target in the previous frame ; the position filter model of the previous frame ( , ).

[0053] Output: The position of the target in the current frame .

[0054] According to the position of the previous frame and scale as well as the position search range coefficient , extract the multi-channel position filter image features from the image , calculate, , Frequency domain position feature map:

[0055] Frequency domain position response map:

[0056] Inverse Fourier transform to calculate the time domain response map: , find the peak point from as the new target position , and the position confidence .

[0057] Target correlation filtering model update Parameters: Learning rate , with a value of 0.01 - 0.1, and the same parameters as the initialization of the correlation filtering model.

[0058] Input: Current frame image , the current frame target position and size of ; the position filter model of the previous frame ( , ).

[0059] Output: The new position filter model ( , ).

[0060] According to the current frame position and size and the position search range coefficient , extract multi-channel position filtered image features from the image and calculate the frequency domain position feature map of the current frame , the frequency domain position response map .

[0061] Update the target tracking model:

[0062]

[0063] Figure 3 This is the initialization process of the hybrid tracking model.

[0064] Initialize the DSST tracking model and the ECO-HC tracking model using the first frame image and the initial position and size of the target. The two models are initialized in parallel in their respective threads. The initialization time of the DSST tracking model is less than 50 milliseconds, and the initialization time of the ECO-HC tracking model is between 500 and 1000 milliseconds.

[0065] Since the initialization of the DSST model is much faster than that of the ECO-HC model, after the DSST model is initialized, set it as the working model, and then start to perform target tracking on subsequent video frames.

[0066] Based on the above initialization method of the hybrid model, use the DSST model to quickly lock the tracking target, and at the same time, the ECO-HC model with better tracking effect is being initialized.

[0067] Figure 4 This is the tracking and updating process of the hybrid tracking model.

[0068] Use the DSST model to track the target in the initial stage of tracking, and at the same time, the ECO-HC model is being initialized.

[0069] Check and switch the ECO-HC tracking model. Before tracking the video frame, check whether the ECO-HC model is initialized. If not, continue to use the DSST model for tracking. If it is completed, set the ECH-HC model as the working model and switch to use the ECO-HC model for tracking; if the ECO-HC model has been switched to the tracking model, always use the ECO-HC model for tracking.

[0070] Use the working model to track the video frame, obtain the target position and scale, output the target tracking box, and if the tracking is successful, update the tracking model with the new target position and scale. ​Through the above process, the hybrid tracking model can switch to using the ECO-HC model for target tracking within one second, and can achieve better tracking results than the DSST model under complex background conditions. Figure 5 is the target tracking status judgment process.

[0071] Input the target tracking position filtering response map as shown in Figure 2 to judge whether the target is lost based on the rules; If the target is lost, set the tracking status as target lost, and then try to retrieve the lost target. If the target is lost, also set the hybrid tracking model invalid, and the hybrid tracking model needs to be re-initialized after retrieving the target.

[0072] The rule-based target loss judgment algorithm is described in detail below.

[0073] As shown in Figure 6 The left figure is the normal target tracking status, and its tracking confidence (0.421) and peak signal-to-noise ratio (38.19) are both relatively large. In the right figure, the target is lost, and its tracking confidence (0.087) and peak signal-to-noise ratio (5.520) are both relatively small. The present invention uses these two tracking performance indicators and their change trends to judge whether the target is lost.

[0074] Target loss judgment algorithm Parameters: 1) Tracking confidence threshold , with a value range of 0.1 - 0.15; 2) Peak signal-to-noise ratio threshold , with a value range of 10 - 15; 3) Tracking confidence decline ratio threshold , with a value range of 5 - 6.

[0075] Input: Target tracking response map .

[0076] Output: Whether the target is lost.

[0077] Take the pixel value at the peak point of the position tracking response map as the target tracking confidence .

[0078] Record the target tracking confidence of the previous frame and the target tracking confidence of the current frame, and calculate the tracking confidence decline rate .

[0079] Set the pixel values of the 10x10 neighborhood near the peak of the tracking response map to 0 to obtain the background response map , calculate the background response map of the mean value and variance , calculate the peak signal-to-noise ratio .

[0080] If the condition is met: < > , then it is determined that the target is lost.

[0081] The search and tracking model retrieves the target again First, expand the search range and initialize the DSST tracking model; Use the DSST model to search for and track the target position and scale; If the search and tracking are successful, set the normal tracking state; Use the target box of the successful tracking and the corresponding video frame to re-initialize the hybrid tracking model and enter the normal tracking state.

[0082] The following details the target retrieval algorithm.

[0083] Call the relevant filter initialization algorithm to re-initialize the target tracking model, input the image of the previous successful tracking and its target position and the size is , and use a larger search range coefficient , The value ranges from 9 to 11, that is, the search range is 9 to 11 times the target size.

[0084] Continuously call the relevant filter tracking algorithm. If the tracking fails for 100 consecutive frames, terminate the tracking of the target.

[0085] If the previous step of tracking is successful, output the target tracking box and enter the normal tracking state.

[0086] Combine the DSST and ECO-HC tracking algorithms based on correlation filtering to track the target, and implement the target loss judgment and re-retrieval mechanism, which better solves the following practical problems faced by the tracking algorithm.

[0087] Quickness, can track the target in real time at a high frame rate, and can quickly initialize and lock the target in time; Adaptability, can adapt to complex backgrounds such as buildings, mountains, and trees, and stably track the target; Long-term performance, can identify the target tracking state, and can retrieve the target when the tracking fails, so as to track the target for a long time.

[0088] A hybrid tracking algorithm is implemented by combining the DSST tracking algorithm and the ECO-HC tracking algorithm. In the algorithm initialization stage, the DSST and ECO-HC tracking models are initialized in parallel, and the DSST model is used to track the target in the initialization stage. After the ECO-HC model is initialized, the tracking is switched to the ECO-HC model to track the target. In this way, the advantages of the two tracking models are fully utilized, which can not only quickly lock the tracking target but also stably track the target in complex backgrounds.

[0089] The mechanism for judging and re-finding the lost target is implemented to timely detect the lost target and start the target search mechanism to re-find the target, enabling the algorithm to have long-term tracking ability.

[0090] The method for judging the lost target uses multiple target tracking metrics to judge whether the target is lost, avoiding continuously tracking the wrong target.

[0091] The method for re-finding the target when it is lost is to re-find the target by expanding the search area after the target is lost, and then switch to the normal tracking state after re-finding.

[0092] In another embodiment of the present invention, a hybrid target tracking system adapted to complex backgrounds is provided, which can be used to implement the above-mentioned hybrid target tracking method adapted to complex backgrounds. Specifically, the system includes: A model initialization module, which is used to initialize the first tracking model and the second tracking model with the video image sequence and the initial position and size of the target. After the first tracking model is initialized, it is set as the working model for target tracking. The first tracking model is used for quick initialization, and the second tracking model is used for target tracking in complex backgrounds; A model update module, which is used to check whether the second tracking model is initialized before tracking the video frame. Otherwise, continue to use the first tracking model for tracking. If so, switch the second tracking model to the working model for normal target tracking; A tracking state judgment module, which is used to input the target tracking position filtering response map to the working model to judge whether the target is lost. If so, set the tracking state as the target being lost; A target re-finding output module, which is used to expand the search range and initialize the first tracking model to re-find the target in the state of the target being lost. After the target search is successful, change the tracking state to normal target tracking.

[0093] The division of modules in the embodiments of the present invention is illustrative, only a logical function division. In actual implementation, there may be other division methods. In addition, in each embodiment of the present invention, the functional modules can be integrated in one processor, or can exist separately physically, or two or more modules can be integrated in one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software functional modules.

[0094] In another embodiment of the present invention, a computer device is provided. The computer device includes a processor and a memory. The memory is used to store a computer program, and the computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function. The processor described in the embodiments of the present invention can be used for the operation of a hybrid target tracking method adapted to a complex background.

[0095] In another embodiment of the present invention, a storage medium is also provided, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and, of course, the extended storage medium supported by the computer device. The computer-readable storage medium provides a storage space that stores the operating system of the terminal. And in this storage space, one or more instructions suitable for being loaded and executed by the processor are also stored. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the hybrid target tracking method adapted to a complex background in the above embodiments.

[0096] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0097] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0098] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0099] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable devices provide steps for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the specific implementation manners of the present invention or make equivalent substitutions, and any modification or equivalent substitution that does not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A hybrid object tracking method adapted to complex backgrounds, characterized in that Including: Initializing a first tracking model and a second tracking model with a video image sequence and the initial position and size of the target. After the initialization of the first tracking model is completed, it is set as the working model for target tracking. The first tracking model is used for rapid initialization, and the second tracking model is used for target tracking in complex backgrounds; Before tracking a video frame, check whether the second tracking model is initialized. Otherwise, continue to use the first tracking model for tracking. If it is, switch the second tracking model to the working model for normal target tracking; Input the target tracking position filter response map into the working model to determine whether the target is lost. If so, set the tracking status as the target being lost; In the state of the target being lost, expand the search range to initialize the first tracking model to retrieve the target again. After the target search is successful, change the tracking status to normal target tracking.

2. The hybrid object tracking method adapted to a complex background according to claim 1, characterized in that The initializing the first tracking model and the second tracking model with a video image sequence and the initial position and size of the target includes: Position search range coefficient , with a value ranging from 3 to 9, that is, the search range is 3 to 9 times the target size; the first frame image of the target is , and the initial position is and the size ; output the initial filter model ( , ); According to the target location and size and the location search range factor , from the image Extracting multi-channel position filtering image features and target position response graph , initialize the position filter model ( , ), calculated as follows: Frequency domain position feature map: Frequency domain position response diagram: Position filter model: 。 3. A hybrid object tracking method adapted to complex backgrounds according to claim 1, characterized in that, After the initialization of the first tracking model is completed, setting it as the working model for target tracking includes: Input the current frame image using the same parameters as those for initializing the filtering model , the position of the target in the previous frame ; the position filter model of the previous frame ( , ); output the position of the target in the current frame ; Based on the position of the previous frame and scale as well as the position search range coefficient , extract multi-channel position filtered image features from the image , and calculate: ​ Frequency domain position feature map: Frequency domain position response diagram: Inverse Fourier transform calculates the time-domain response diagram: , starting from Find the peak point as the new target position , and the position confidence .

4. A hybrid object tracking method adapted to complex backgrounds according to claim 1, characterized in that Before tracking a video frame, checking whether the second tracking model is initialized. Otherwise, continue to use the first tracking model for tracking. If it is, switch the second tracking model to the working model for normal target tracking includes: Using the first tracking model to track the target in the initial stage of tracking, while the second tracking model is being initialized; before tracking a video frame, check whether the second tracking model is initialized. If not, continue to use the first tracking model for tracking. If it has been completed, set the second tracking model as the working model and continue to use the second tracking model for tracking for those who have switched the second tracking model; Using the working model to track the video frame to obtain the target position and scale, output the target tracking box. If the tracking is successful, update the tracking model with the new target position and scale.

5. A hybrid object tracking method adapted to complex backgrounds according to claim 4, characterized in that, Specific update method: Input the current frame image , the target position of the current frame and the size is ; the position filter model of the previous frame ( , ); output the new position filter model ( , ); According to the current frame position and size as well as the position search range coefficient , extract multi-channel position filtering image features from the image , and calculate the frequency domain position feature map and the frequency domain position response map ; Updating the target tracking model: 。 6. A hybrid object tracking method adapted to complex backgrounds according to claim 1, characterized in that Inputting the target tracking position filter response map into the working model to determine whether the target is lost. If so, setting the tracking status as the target being lost includes: Tracking confidence threshold , with a value range of 0.1 to 0.15; Peak signal-to-noise ratio threshold , with a value range of 10 to 15; Tracking confidence decline ratio threshold , with a value range of 5 to 6; Input target tracking response graph , Output: Whether the target is lost; Obtain the position tracking response graph The pixel value of the peak point is the target tracking confidence ; Record the target tracking confidence of the previous frame and the target tracking confidence of the current frame , and calculate the decline rate of the tracking confidence ; The tracking response map Set the pixel values in the peak neighborhood to 0 to obtain the background response map , and calculate the background response map mean value and variance , and calculate the peak signal-to-noise ratio ; If the condition is met: < > , then it is determined that the target is lost.

7. A hybrid object tracking method adapted to complex backgrounds according to claim 1, characterized in that, In the state of the target being lost, expanding the search range to initialize the first tracking model to retrieve the target again. After the target search is successful, changing the tracking status to normal target tracking includes: First, expand the search range to initialize the first tracking model, use the first tracking model to search for the target position and scale. If the search and tracking are successful, set the normal tracking status; re-initialize the hybrid tracking model with the successfully tracked target box and the corresponding video frame and enter the normal tracking state; Input the successfully tracked image of the previous frame and its target position and the size is and use a larger search range coefficient , The value ranges from 9 to 11, that is, the search range is 9 to 11 times the target size; Continuously call the filter tracking algorithm. If the tracking fails for 100 consecutive frames, terminate the tracking of the target; if the previous step is successful, output the target tracking box and enter the normal tracking state.

8. A hybrid object tracking system adapted to complex backgrounds, characterized in that, Including: A model initialization module for initializing a first tracking model and a second tracking model with a video image sequence and the initial position and size of the target. After the initialization of the first tracking model is completed, it is set as the working model for target tracking. The first tracking model is used for rapid initialization, and the second tracking model is used for target tracking in complex backgrounds; The model update module is used to check whether the second tracking model is initialized before tracking the video frame. Otherwise, continue to use the first tracking model for tracking. If it is, switch the second tracking model to the working model for normal target tracking; The tracking status judgment module is used to input the target tracking position filtering response map into the working model to judge whether the target is lost. If so, set the tracking status to target lost; The target recovery output module is used to expand the search range to initialize the first tracking model to recover the target under the target lost state. After the target search is successful, change the tracking status to normal target tracking.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a hybrid target tracking method adaptable to complex backgrounds as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of a hybrid target tracking method adaptable to complex backgrounds as described in any one of claims 1 to 7.

Images