Method and apparatus for tracking real-time moving objects using disruptor-suppressed response variation-aware multiregularized tracking filter in an image monitoring system

Abstract

The present invention relates to a method and apparatus for tracking a moving object using a rank-based context-aware multi-regularized correlation filter in a video surveillance system. The method comprises: a DSRVMRT configuration step of configuring a rank-based context-aware multi-regularized correlation filter (DSRVMRT) by adding an optional spatial normalization term, a group spatial channel normalization term, a temporal smoothness term, and a learning rank-based context awareness term to a DCF; a multi-feature extraction step of extracting multiple features including deep features and at least one manual feature from an input frame; a target tracking step of generating a response map corresponding to the multiple features using the DSRVMRT and determining a target location based on the response map; and an update step of updating the DSRVMRT and an appearance model according to the target location and resetting the search area of ​​the next frame.

Description

Technology Field

[0001] The present invention relates to a moving object tracking technology, and in particular to a moving object tracking method and apparatus using a disturber suppression response variation recognition multiple normalization tracking filter that enables more powerful tracking performance. Background Technology

[0002] Visual Object Tracking (VOT) is a computer vision technology that identifies and tracks objects in video streams or image sequences. It detects objects of interest, estimates their location and size, and updates this information as the object moves within the frame.

[0003] VOT can be used in various application fields, including surveillance, robotics, and augmented reality. To effectively track objects, algorithms must be able to handle changes in the object's appearance, such as occlusion and lighting variations.

[0004] Visual object tracking can be performed based on correlation filters (CF), discriminative correlation filters (DCF), and convolutional neural networks (CNN).

[0005] Correlation filter-based tracking is an object tracking method that uses a correlation filter to identify and track objects in a video sequence. It has the advantages of being fast and accurate, but has the disadvantage that the performance of the tracker degrades when an object undergoes a significant change in appearance or when an object is obscured by other objects in the frame.

[0006] DCF-based tracking is efficient, robust, and capable of handling fast-moving, deformed, and occluded objects, making it suitable for various applications such as video surveillance, robotics, and autonomous vehicles.

[0007] In correlation filter-based tracking, the ridge regression model is trained using the cyclic structure of the underlying samples, whereas DCF improves computational efficiency by utilizing the Fast Fourier Transform (FFT). Cyclic shift operations cause boundary effect problems, and damaged samples resulting from unwanted boundary effects can degrade the performance of the tracker during training.

[0008] However, DCF-based tracing uses only information from two consecutive frames, which has the problem of not being sufficiently robust due to information limitations and performance degradation resulting from the use of historical information without noise removal. Prior art literature

[0009] 1. J.F. Henriques, R. Caseiro, P. Martins, J. Batista, High-speed tracking with kernelized correlation filters, IEEE Transactions on Pattern Analysis and Machine Intelligence 37 (3) (2015) 583??596.2. M. Danelljan, G. H¨ager, F. Khan, M. Felsberg, Accurate scale estimation for robust visual tracking, in: British Machine Vision Conference, Nottingham, September 1-5, 2014, BMVA Press, March 2014. M. Mueller, N. Smith, B. Ghanem, Context-aware correlation filter tracking, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 1396??1404. The problem to be solved

[0010] Accordingly, in order to solve the aforementioned problems, the present invention aims to provide a method and apparatus for tracking moving objects using a multi-normalized tracking filter that recognizes disturbance suppression response variation in a video surveillance system, which integrates history interval information using responses during the filter training process to obtain more robust tracking performance.

[0011] The objectives of the present invention are not limited to those mentioned above, and other unmentioned objectives will be clearly understood by those skilled in the art to which the present invention pertains from the description below. means of solving the problem

[0012] As a means to solve the above problem, a method for tracking a moving object using a disruptor-suppressed response variation-aware multiregularized tracking filter in a video surveillance system according to one embodiment of the present invention provides: a DSRVMRT configuration step of configuring a disruptor-suppressed response variation-aware multiregularized tracking filter (DSRVMRT) by adding a salience recognition normalization term, a response variation recognition term, a channel reliability recognition normalization term, a disruption suppression correlation filter term, and an interval response mismatch element to the objective function of a BACF; a multi-feature extraction step of extracting multiple features including deep features and at least one manual feature from an input frame; a target tracking step of generating a response map corresponding to the multiple features using the DSRVMRT and determining a target location based on the response map; and an update step of updating the DSRVMRT and the appearance model according to the target location and resetting the search area of ​​the next frame.

[0013] The above DSRVMRT is " It has an objective function of ", and the above R SAR is the salience recognition normalization term, the above R RVAR The response variation recognition term, the above R CAR is the channel reliability recognition normalization term, the above R DSR The disturbance suppression correlation filter term, above is a vectorized feature map containing D channels of size N, the above The Gaussian shape label, above is a binary matrix that obtains more negative samples ( ), above is the number of channels, the above operator ⊙ is the correlation operator, the above The normalization parameter, the above It is characterized by representing each correlation filter used for object localization.

[0014] The above protrusion recognition normalization term is " It is expressed as ", and the above It is characterized by being a predefined protrusion normalizer.

[0015] The above response variation recognition term is " It is expressed as ", where the above λ represents the normalization parameter, and the above is two response maps and It is characterized by the difference between livers.

[0016] The above channel reliability recognition normalization term is " It is expressed as,

[0017] The above is a diagonal matrix composed of all D-channel weights, the above is a preset constant, the above Each is characterized by representing an initial weight distribution.

[0018] The above disturbance suppression correlation filter term is " It is expressed as ", and the above is the penalty mask of the (tf)th frame, above It is characterized by representing the response map of the (tf)th frame, respectively.

[0019] The above DSRVMRT configuration step is characterized by training a filter for the t-th frame by reflecting the response maps of the t-1-th frame to suppress interval response mismatch.

[0020] The above at least one manual feature is characterized by including at least one of HOG (Histogram of Oriented Gradients), ColorName (CN), Intensity (IC), and Saliency.

[0021] As a means to solve the above problem, the present invention provides a moving object tracking device using a disruptor-suppressed response variation-aware multiregularized tracking filter, comprising: a DSRVMRT component that configures a disruptor-suppressed response variation-aware multiregularized tracking filter (DSRVMRT) by adding a salience recognition normalization term, a response variation recognition term, a channel reliability recognition normalization term, a disturbance suppression correlation filter term, and an interval response mismatch element to the objective function of a Young BACF according to another embodiment of the present invention; a multi-feature extraction unit that extracts multiple features including deep features and at least one manual feature from an input frame; a target tracking unit that generates a response map corresponding to the multiple features using the DSRVMRT and determines a target location based on the response map; and an update unit that updates the DSRVMRT and the appearance model according to the target location and resets the search area of ​​the next frame. Effects of the invention

[0022] The present invention enables more robust tracking performance by integrating history interval information that uses responses during the filter training process.

[0023] In particular, it uses an efficient method based on a salience detection algorithm and normalizes response deviation and channel reliability to achieve smooth response deviation and an adaptive channel weight distribution.

[0024] In addition, a response bucketing-based interruptor recognition system is introduced to detect and penalize areas surrounding tracked objects that may be interrupted due to similar objects or partial occlusion, thereby minimizing the occurrence of tracking interruptions. Brief explanation of the drawing

[0025] FIGS. 1 and 2 are drawings for illustrating a method for tracking a moving object using a multiple normalization tracking filter that recognizes disturbance suppression response variation in an image surveillance system according to an embodiment of the present invention. FIG. 3 is a diagram illustrating a method for recognizing a disturbance according to an embodiment of the present invention. FIG. 4 is a diagram illustrating the object tracking results according to one embodiment of the present invention. FIG. 5 is a diagram illustrating a moving object tracking device using a multiple normalization tracking filter that recognizes disturbance suppression response variation in an image surveillance system according to an embodiment of the present invention. Specific details for implementing the invention

[0026] Hereinafter, preferred embodiments are described in detail with reference to the attached drawings so that those skilled in the art can easily practice the present invention. However, in describing the preferred embodiments of the present invention in detail, if it is determined that a detailed description of related known functions or configurations may unnecessarily obscure the essence of the present invention, such detailed description is omitted. Additionally, the same reference numerals are used throughout the drawings for parts having similar functions and operations.

[0027] Additionally, throughout the specification, when a part is described as being 'connected' to another part, this includes not only cases where they are 'directly connected,' but also cases where they are 'indirectly connected' with other elements in between. Furthermore, the description of a component as 'including' means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0028] FIGS. 1 and 2 are drawings for illustrating a method for tracking a moving object using a multiple normalization tracking filter that recognizes disturbance suppression response variation in an image surveillance system according to an embodiment of the present invention.

[0029] As illustrated in FIG. 1, the method of the present invention largely comprises a DSRVMRT configuration step (S1) for configuring a disruptor-suppressed response variation-aware multiregularized tracking filter (DSRVMRT) by adding a salience-aware normalization term, a response variation-aware term, a channel reliability-aware normalization term, a disturbance-suppressed correlation filter term, and an interval response mismatch element to the objective function of BACF; a multi-feature extraction step (S2) for extracting multiple features including deep features and at least one manual feature from an input frame; a target tracking step (S3) for generating a response map corresponding to the multiple features using the DSRVMRT and determining a target location based on the response map; and an update step (S4) for updating the DSRVMRT and the appearance model according to the target location and resetting the search area of ​​the next frame.

[0030] The present invention proposes a new correlation filter called DSRVMRT, which enables the dynamic generation of a protrusion recognition normalizer based on a target shape, allowing the tracker to perform filter penalty and background noise suppression during object deformation.

[0031] Furthermore, by utilizing response buckets to identify interrupted areas in the response and generating a custom penalty mask targeting those areas, it enables effective suppression of interfering background noise and improves the accuracy of object tracking results by considering the presence of disturbances.

[0032] In addition, it incorporates a new normalization method sensitive to changes in response. By normalizing the second-order difference of the response, it effectively smooths response fluctuations and improves the tracking recognition rate for changes in object appearance.

[0033] In addition, a new channel reliability awareness normalization is proposed to achieve automatic distribution of channel weights. This optimizes weights along with filters during the training phase, and as a result, the tracker's attention is directed toward the most reliable channel, thereby improving performance.

[0034] The method for tracking moving objects according to the present invention will be described in more detail below.

[0035] Step S1: DSRVMRT Configuration

[0036] 1) Selection of Baseline Tracker

[0037] In the present invention, the idea of ​​multiple feature channels and background recognition is combined, and a Background-Aware Correlation Filter (BACF) tracker using a multi-channel CF-based filter that can be expressed as ridge regression is considered as a baseline tracker.

[0038] BACF has an objective function expressed by mathematical formula 1.

[0039] [Mathematical Formula 1]

[0040]

[0041] Here, is a vectorized feature map containing D channels of size N, The Gaussian shape label, is a binary matrix that obtains more negative samples ( ), is the number of channels, and the operator ⊙ is the correlation operator, represents the normalization parameters respectively, and is a correlation filter used for object localization.

[0042] In the detection process, the response map is a search patch obtained from a newly captured frame and learned filter It is generated by calculating the spatial correlation between them.

[0043] The result value indicated by is used to determine the target location, It is updated based on the position of the maximum value.

[0044] The performance of ACF trackers is good, but they have several issues. The existing approach of traditional DCF trackers, which updates appearance model frames frame by frame using a fixed learning rate, fails to account for the current state of the tracked objects.

[0045] As a result, it fails to adapt to changes in the complex tracking environment and respond to anomalies or situations where tracking targets are lost, and additional optimization must be performed to address boundary effects.

[0046] 2) Objective function of DSRVMRT

[0047] The present invention presents a disturbance-suppressing response variation-recognizing multiple normalization CF for visual tracking, thereby addressing the weaknesses of BACF trackers and improving model stability and accuracy.

[0048] Figure 2 illustrates the general process and framework of the DSRVMRT algorithm. The presented method can adapt to changing tracking conditions and incorporates a method for updating the appearance model to generate more accurate and reliable tracking results. Therefore, the objective function of DSRVMRT It is as follows.

[0049] [Mathematical Formula 2]

[0050]

[0051] Here, R SAR , R RVAR , R CRAR , R DSR represents the salience recognition normalization term, the response variation recognition normalization term, the channel reliability recognition normalization term, and the disturbance suppression correlation filter term, respectively.

[0052] 2-1) Prominence Recognition Normalization Term (R SAR ) Introduction

[0053] Prominence-based methods are frequently used in DCF detection-based tracking methods to improve performance, as they can identify critical regions in images and accommodate irregular object shapes. To enable real-time updates of weight maps within correlation filters, DSAR-CF incorporates prominence methods, allowing the weight map to create synergy with the prominence map, thereby enabling real-time updates of the weight map within the correlation filter.

[0054] These updates enable an accurate representation of the target's current appearance and provide meaningful information about the target.

[0055] Therefore, the present invention utilizes protrusion information to construct a normalizer that better reflects changes in the target shape, and a corresponding protrusion recognition normalization term is R SAR It is expressed by mathematical formula 3.

[0056] [Mathematical Formula 3]

[0057]

[0058] It is a predefined salience normalizer that dynamically prioritizes object domain information through the regression optimization process and enables direct suppression of boundary effects that may occur during correlation operations.

[0059] 2-2) Response Variation Recognition Normalization Term (R RVAR ) Introduction

[0060] If the DCF filter is updated for every frame using the standard approach, the filter for the current frame may be overfitted.

[0061] In addition, if a trained filter is simply added to the previous filter model, a problem may arise where it cannot adapt to sudden changes in the target.

[0062] While the update method works well when the target changes continuously, when the target changes significantly, the trained filter differs greatly from the previous filter. Consequently, the combination of the two filters fails to apply effectively to the current or previous target state.

[0063] Furthermore, if occlusion and blurring frequently occur during the target tracking process, background information and noise are introduced into the filters learning these training samples. This acts as a factor in tracking drift and failure.

[0064] To enhance the robustness of the tracker, information from the target confidence region can be utilized by integrating response changes into the DCF learning process.

[0065] Therefore, in the present invention, the normalization term R for response variation RVAR By introducing [this], the tracker performance can be improved by adaptively considering the spatial difference of the previous frame. This is achieved by introducing a constraint on the second-order difference in response, which is expressed as Equation 4.

[0066] [Mathematical Formula 4]

[0067]

[0068] Here, λ represents the normalization parameter, and is two response maps and It is obtained by calculating the difference between them.

[0069] [Mathematical Formula 5]

[0070]

[0071] Here, two response maps and The peaks of are shifted to the center by the shift operator S(.).

[0072] 2-3) Channel Reliability Aware Normalization Term (R CRAR ) Introduction

[0073] To improve the accuracy of target localization, the contribution of each feature channel is emphasized by weighting the responses to target locations using channel reliability metrics. Accordingly, the present invention provides a channel weight distribution We propose integrating the learning of into the model training process to suppress misleading information. Consequently, the regularization term R that explains channel reliability. CRAR This is generated.

[0074] [Mathematical Formula 6]

[0075]

[0076] Here, is a diagonal matrix composed of all D-channel weights, is a preset constant, Each represents the initial weight distribution. The channel weights of the current frame are necessary to transition smoothly and simultaneously prioritize elastic channels.

[0077] 2-4) Disturbance-suppressing correlation filter term (R DSR ) Introduction

[0078] The process generates a response map by correlating the DCF with features within the search area.

[0079] In the new frame, the estimated localization of the target is determined by identifying the highest peak in the response map, and the resulting response map accurately reflects the state of the target. Consequently, changes in object shape are reflected in the response map early and immediately.

[0080] Accordingly, the present invention utilizes this response when designing temporal normalization.

[0081] The response may contain incorrect background noise.

[0082] Accordingly, the present invention proposes a new disturbance recognition method based on response buckets to identify and remove these disturbance elements as shown in FIG. 3.

[0083] In Fig. 3, the response map is shown using the bucket method for the t-f)th frame. It shows the process of dividing into α × α non-overlapping rectangles. Then, background noise (disturbance values) in individual buckets is detected based on the peaks of the response map. Based on this, penalty factors are identified and an adaptive penalty mask is applied. Develops.

[0084] The disturbance detection method calculates θ(i, j) as follows:

[0085] [Mathematical Formula 7]

[0086]

[0087] Here, m(i, j) is the local maximum of the i×j response bucket, and m max is the global maximum of the entire response map. Then, the penalty factor It is calculated as in mathematical formula 8.

[0088] [Mathematical Formula 8]

[0089]

[0090] Here, ρ represents a predefined constant, and is a threshold that checks whether the response bucket contains a disruptor.

[0091] Likewise, the penalty mask of the (tf)th frame It is expressed as in mathematical formula 9.

[0092] [Mathematical Formula 9]

[0093]

[0094] Consequently, the response map of the (tf)th frame Is ⊙ Low noise is removed. The disturbance recognition method can effectively and accurately identify and suppress disturbances in the response.

[0095] [Mathematical Formula 10]

[0096]

[0097] Figure 3 shows that the response map of the DSRVMRT tracker remains stable despite disturbances caused by similar objects, enabling accurate target location. The prediction results of the DSRVMRT tracker closely match the ground-truth labels of the tracked target, providing compelling evidence for the effectiveness of the disturbance recognition model.

[0098] 2-5) Reflecting Interval Response Discrepancy Factors

[0099] When VOT is applied to fast-moving object detection or aerial platforms, it presents a problem in that it is difficult to maintain consistent tracking performance using only two consecutive frames. On the other hand, relying on hundreds of historical training samples may be impractical due to memory space limitations.

[0100] To solve these problems, the present invention proposes a new tracking approach aimed at suppressing interval response mismatch.

[0101] In particular, the filter training process for the t-th frame is performed by reflecting the response maps of the t-1-th frame from the tf-th frame, so that the learned filter adapts better to object variations and makes more accurate predictions in the next frame (where f is a natural number greater than or equal to 2).

[0102] The response map extraction method of the present invention requires minimal memory but can significantly improve the robustness of object tracking as shown in FIG. 2 by incorporating a hysteresis interval response that effectively suppresses aberrations.

[0103] 3) The final objective function of DSRVMRT

[0104] Therefore, the final objective function of DSRVMRT is expressed as Equation 11.

[0105] [Mathematical Formula 11]

[0106]

[0107] is the weight distribution function.

[0108] Step S2: Multiple Feature Extraction

[0109] In the present invention, multiple features are extracted from an input frame, and the multiple features may consist of at least one manual feature among HOG (Histogram of Oriented Gradients), ColorName (CN), Intensity (IC), and Saliency, and deep features extracted using ResNet-50, etc.

[0110] At least one of HOG (Histogram of Oriented Gradients), ColorName (CN), and Intensity (IC)

[0111] Step S3: Target Tracking

[0112] Multiple response maps corresponding to deep features and manual features are generated using DSRVMRT and fused to obtain a final response map, and the target location is identified based on the final response map.

[0113] 1) Conversion to the frequency domain

[0114] In the present invention, auxiliary variables By applying it to Equation 10, Equation 11 is transformed into Equation 12.

[0115] [Mathematical Formula 12]

[0116]

[0117] And to interpret Equation 12, Parseval's theorem is used to transform the loss function into the frequency domain. The d-th channel response of the t-th frame After formulating it as such, apply it to Equation 11 to transform Equation 11 into Equation 13.

[0118] [Mathematical Formula 13]

[0119]

[0120] 2) Objective Function Analysis

[0121] In the present invention, the calculation speed is accelerated using the ADMM method, and Equation 13 is interpreted using an augmented Lagrangian function.

[0122] [Mathematical Formula 14]

[0123]

[0124] Here, m represents the Lagrangian multiplier and μ represents the penalty factor, respectively. For simplification and Defines.

[0125] and Assuming that, mathematical formula 15 is obtained.

[0126] [Mathematical Formula 15]

[0127]

[0128] In this state, sub-problems , , By interpreting the Lagrange term through the ADMM method It can be minimized.

[0129] Sub-problems analysis

[0130] It can be interpreted according to mathematical formula 16.

[0131] [Mathematical Formula 16]

[0132]

[0133] Sub-problems is time-consuming to calculate It includes. Accordingly, the present invention uses ADMM to calculate the optimization problem.

[0134] Here, Is Dependent on, And, is a conjugate operation.

[0135] Mathematical formula 16 can be converted into N independent sub-representations.

[0136] [Mathematical Formula 17]

[0137]

[0138] Here, And, Is The DFT of (i.e., ) and, The answer is calculated by mathematical formula 18.

[0139] [Mathematical Formula 18]

[0140]

[0141] Since the matrix inversion in Equation 18 is computationally expensive, the Sherman-Morrison formula (i.e., Optimization and acceleration of calculations are pursued using ). In this case, And, am.

[0142] As a result, mathematical formula 18 becomes identical to mathematical formula 19.

[0143] [Mathematical Formula 19]

[0144]

[0145] Here, am.

[0146] Sub-problems analysis

[0147] Sub-problems By calculating the partial derivative with respect to , the optimal solution as in Equation 20 is obtained.

[0148] [Mathematical Formula 20]

[0149]

[0150] Simplify mathematical formula 20 as mathematical formula 21.

[0151] [Mathematical Formula 21]

[0152]

[0153] Sub-problems analysis

[0154] Sub-problems By calculating the partial derivative of, the optimal solution as in Equation 22 is obtained.

[0155] [Mathematical Formula 22]

[0156]

[0157] Simplify mathematical formula 22 as mathematical formula 23.

[0158] [Mathematical Formula 23]

[0159]

[0160] Here, am.

[0161] During iterative optimization, the filter tends to focus on the reliable regions of the tracked object, while imposing a significant penalty on cluttered regions. However, to enable the filter to focus on reliable regions for better convergence, Use additional ADMM solutions when optimizing.

[0162] This method can yield better results, but it requires more computational resources and is slower compared to baseline trackers.

[0163] Step S4: Update Step

[0164] 1) Lagrange Update

[0165] The Lagrangian can be updated according to mathematical formulas 24 and 25.

[0166] [Mathematical Formula 24]

[0167]

[0168] [Mathematical Formula 25]

[0169]

[0170] Here, superscripts i and i + 1 represent repetitions, and β represents the step length, respectively.

[0171] 2) Exterior model update

[0172] In the DSRVMRT algorithm, an adaptive update strategy is used to handle inconsistent modifications to object appearances. Model is updated according to mathematical formula 26.

[0173] [Mathematical Formula 26]

[0174]

[0175] Here, is the learning rate, and t and (t+1) represent the t-th and (t+1)-th frames, respectively.

[0176] FIG. 4 is a diagram illustrating the object tracking results according to one embodiment of the present invention.

[0177] In Figure 4, red represents the object tracking result using the DSRVMRT of the present invention, and green represents the object tracking result using the conventional BACF.

[0178] By comparison, it can be seen that when object occlusion and deformation phenomena occur in basketball game footage, the DSRVMRT of the present invention, unlike BACF, successfully detects and suppresses disturbance factors caused by occlusion-related responses to perform accurate object repositioning operations.

[0179] FIG. 5 is a diagram illustrating a moving object tracking device using a multiple normalization tracking filter that recognizes disturbance suppression response variation in an image surveillance system according to an embodiment of the present invention.

[0180] As illustrated in FIG. 3, the device (100) of the present invention may include a DSRVMRT configuration unit (110) that configures a disruptor-suppressed response variation-aware multiregularized tracking filter (DSRVMRT) by adding a salience recognition normalization term, a response variation recognition term, a channel reliability recognition normalization term, a disturbance suppression correlation filter term, and an interval response mismatch element to the objective function of BACF; a multi-feature extraction unit (120) that extracts multiple features including deep features and at least one manual feature from an input frame; a target tracking unit (130) that generates a response map corresponding to the multiple features using the DSRVMRT and determines a target location based on the response map; and an update unit (140) that updates the DSRVMRT and the appearance model according to the target location and resets the search area of ​​the next frame.

[0181] A moving object tracking device configured in this manner may be implemented as a separate hardware device or as an embedded system in which the hardware or software is embedded as part of other hardware or software, but is not limited thereto.

[0182] Although preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above. It is understood that various modifications can be made by those skilled in the art without departing from the essence of the invention as claimed in the claims, and such modifications should not be understood individually from the technical spirit or perspective of the present invention.

Claims

Claim 1 A method for tracking a moving object using a disruptor-suppressed response variation-aware multiregularized tracking filter executed in a processor of a video surveillance system, comprising: a DSRVMRT configuration step of configuring a disruptor-suppressed response variation-aware multiregularized tracking filter (DSRVMRT) by adding a salience-aware normalization term, a response variation-aware term, a channel reliability-aware normalization term, a disruptor-suppressed correlation filter term, and an interval response mismatch element to the objective function of a BACF; a multi-feature extraction step of extracting multiple features including deep features and at least one manual feature from an input frame; a target tracking step of generating a response map corresponding to the multiple features using the DSRVMRT and determining a target location based on the response map; and an update step of updating the DSRVMRT and the appearance model according to the target location and resetting the search area of ​​the next frame. Claim 2 In paragraph 1, the above DSRVMRT is " It has an objective function of ", and the above R SAR is the salience recognition normalization term, the above R RVAR The response variation recognition term, the above R CAR is the channel reliability recognition normalization term, the above R DSR A moving object tracking method using a perturbator suppression response variation recognition multiple normalization tracking filter, characterized in that represents a perturbation suppression correlation filter term, x represents a feature map, y represents a Gaussian shape label, P represents a binary matrix, d represents a channel index, D represents the total number of channels, t represents a time index, and h represents a correlation filter. Claim 3 In paragraph 2, the above protrusion recognition normalization term is " It is expressed as ", and the above SR is a predefined protrusion normalizer, the above A moving object tracking method using a multiple normalization tracking filter that recognizes disturbance suppression response variation, characterized in that is a normalization parameter and ⊙ is a correlation operator. Claim 4 In paragraph 2, the above response variation recognition term is " It is expressed as ", where the above λ represents the normalization parameter, and the above is two response maps and A moving object tracking method using a multiple normalization tracking filter that recognizes disturbance suppression response variation characterized by the difference between them. Claim 5 In paragraph 2, the channel reliability recognition normalization term is " It is expressed as ", and the above is a diagonal matrix composed of all D-channel weights, the above is a preset constant, the above A moving object tracking method using a multiple normalization tracking filter that recognizes disturbance suppression response variation, characterized by each representing an initial weight distribution. Claim 6 In paragraph 2, the above disturbance suppression correlation filter term is " It is expressed as ", and the above is the penalty mask of the (tf)th frame, above is the response map of the (tf)-th frame, the above ⊙ is the correlation operator, the above is a time interval weighting coefficient that adjusts the reflection strength of the (tf)th frame response, the above A moving object tracking method using a multiple normalization tracking filter that recognizes disturbance suppression response variation, characterized by each representing a weight distribution function. Claim 7 A moving object tracking method using a perturbator suppression response variation recognition multiple normalization tracking filter, wherein, in claim 6, the DSRVMRT configuration step trains a filter for the t-th frame by reflecting response maps from the tf-th to t-1-th frames to suppress interval response mismatch. Claim 8 A method for tracking a moving object using a multi-normalized tracking filter that recognizes disturbance suppression response variation, wherein, in claim 1, the at least one manual feature includes at least one of HOG (Histogram of Oriented Gradients), ColorName (CN), Intensity (IC), and Saliency. Claim 9 A moving object tracking device using a disruptor-suppressed response variation-aware multiregularized tracking filter, comprising: a DSRVMRT component that configures a disruptor-suppressed response variation-aware multiregularized tracking filter (DSRVMRT) by adding a salience-aware normalization term, a response variation awareness term, a channel reliability awareness normalization term, a disturbance-suppressed correlation filter term, and an interval response mismatch element to the objective function of a BACF; a multi-feature extraction unit that extracts multiple features including deep features and at least one manual feature from an input frame; a target tracking unit that generates a response map corresponding to the multiple features using the DSRVMRT and determines a target location based on the response map; and an update unit that updates the DSRVMRT and the appearance model according to the target location and resets the search area of ​​the next frame.

Images