CN115631216A - Holder target tracking system and method based on multi-feature filter fusion - Google Patents
Holder target tracking system and method based on multi-feature filter fusion Download PDFInfo
- Publication number
- CN115631216A CN115631216A CN202211644874.1A CN202211644874A CN115631216A CN 115631216 A CN115631216 A CN 115631216A CN 202211644874 A CN202211644874 A CN 202211644874A CN 115631216 A CN115631216 A CN 115631216A
- Authority
- CN
- China
- Prior art keywords
- target
- filter
- response
- image
- tracking
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 54
- 230000004927 fusion Effects 0.000 title claims abstract description 35
- 230000004044 response Effects 0.000 claims abstract description 89
- 230000008859 change Effects 0.000 claims abstract description 23
- 238000010586 diagram Methods 0.000 claims abstract description 19
- 230000008569 process Effects 0.000 claims abstract description 8
- 238000001914 filtration Methods 0.000 claims description 18
- 238000004364 calculation method Methods 0.000 claims description 10
- 230000000875 corresponding effect Effects 0.000 claims description 8
- 230000003044 adaptive effect Effects 0.000 claims description 7
- 230000001419 dependent effect Effects 0.000 claims description 7
- 230000009471 action Effects 0.000 claims description 3
- 239000002131 composite material Substances 0.000 claims description 2
- 230000002596 correlated effect Effects 0.000 claims description 2
- 238000012795 verification Methods 0.000 claims 1
- 238000004422 calculation algorithm Methods 0.000 description 5
- CURLTUGMZLYLDI-UHFFFAOYSA-N Carbon dioxide Chemical compound O=C=O CURLTUGMZLYLDI-UHFFFAOYSA-N 0.000 description 4
- 229910002092 carbon dioxide Inorganic materials 0.000 description 2
- 239000001569 carbon dioxide Substances 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000033764 rhythmic process Effects 0.000 description 2
- 230000000903 blocking effect Effects 0.000 description 1
- 230000001276 controlling effect Effects 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/277—Analysis of motion involving stochastic approaches, e.g. using Kalman filters
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
- G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention discloses a pan-tilt target tracking system and a pan-tilt target tracking method, wherein tracking filters are respectively constructed according to different characteristics, weight is adaptively distributed to each filter, weighting fusion is carried out to obtain target position information, then the characteristic filter with the largest weight is used for predicting target dimensions in different dimension spaces, so that accurate dimension information of a target is obtained, the problem of dimension change in the target motion process is solved, and the information of the shielded or tracking loss state of the target is timely output by designing a template filter and a track filter. In the invention, all filters are not updated in fixed steps, and the updating rate of the filters is determined by setting the peak value and the peak-to-side lobe ratio of the fused response diagram. The multi-filter fusion tracking system and the multi-filter fusion tracking method can realize the target tracking of the holder, increase the accuracy and the robustness of the target tracking, and simultaneously can output the state information of the target in time, including correct tracking, shielding and loss.
Description
Technical Field
The invention relates to a holder target tracking system and a method, in particular to a holder target tracking system and a method based on multi-feature filter fusion; belonging to the technical field of image target tracking.
Background
The holder target tracking is a key technology in holder application, an existing target tracking algorithm is mainly a tracking algorithm based on relevant filtering, and although the processing requirement of a holder chip is met, the problems that tracking is easy to lose and drift, target shielding cannot be processed and the like still exist. For example, most kernel-dependent filtering tracking methods use a single manual feature, which is weak in feature representation capability; or simply adding and summing the correlation representations of various manual features to achieve the purpose of feature fusion, and the representation capability of various features cannot be fully exerted, so that the target tracking effect is not robust and is easy to lose.
With the development of the technology, a part of correlation filtering methods provide a new technology for respectively training correlation filters for channels with different manual characteristics, and weights are adaptively distributed to each filter, so that although the characteristic representation capability is enhanced, the calculation amount is multiplied with the increase of characteristic channels (for example, 32 channels are provided for histogram gradient characteristics, and 32 filters are trained), so that the real-time effect cannot be achieved in a low-end pan-tilt chip. In addition, a deep learning training characteristic is adopted, and a related filter is combined to track a target so as to achieve the purpose of improving the tracking precision, such as a SiamFC series method; or training the target tracking model end to end by using the latest deep learning technology transform algorithm framework. However, the methods have the problems of large model calculation amount, low running speed and inapplicability to common holder chips.
In addition, most target tracking methods based on correlation filtering adopt a filter to predict on multiple scales when processing target scale changes, and select the scale with the largest response peak value to update the target state, or adopt a DSST filtering method to train the filter alone to predict the scale changes. The method scales the length and the width of the target in the same proportion, and the variation trend of the length and the width of the target is different and is a common phenomenon along with the appearance variation and not strictly following the rule in the moving process of the target. Therefore, such a tracking method cannot accurately predict the target scale, so that background noise is continuously introduced into the model or local information of the target is too focused, the filter model is deteriorated, and finally tracking failure is caused.
In addition to the above drawbacks, the current methods for tracking related filtering targets mostly update the model at a fixed learning rate, so that noises such as background are inevitably introduced continuously, and the distinguishing capability of the model is continuously damaged. In addition, each related filtering target tracking method mostly expresses the target state by the maximum value of the final response map, and if the maximum value is lower than a certain threshold value, the target shielding or the tracking failure is judged. Due to the continuous decline of the distinguishing capability of the tracking model, the tracking state of the target cannot be described by simply depending on the extreme value of the response graph, and meanwhile, the problem that the target is subjected to transient shielding cannot be effectively processed.
For the above reasons, it is necessary to provide a pan-tilt target tracking algorithm to optimize and solve the above problems and drawbacks.
Disclosure of Invention
The invention aims to provide a holder target tracking system and method based on multi-feature filter fusion, aiming at the problems in the prior art, the holder target tracking is realized through multi-feature filter fusion decision, the accuracy and the robustness of target tracking are improved, and meanwhile, the state information of a target can be timely output.
In order to achieve the above object, the present invention adopts the following technical solutions:
the invention firstly discloses a holder target tracking system based on multi-feature filter fusion, which comprises:
the correlation response graph fusion module: training a plurality of feature correlation filters by utilizing a kernel correlation filter based on a previous frame of image and target information, setting self-adaptive weights, fusing to generate a final correlation response diagram, and predicting the position of a target central point of a current frame;
a target scale prediction module: predicting target scale information of the current frame in a scale change space according to the feature correlation filter with the maximum weight;
the correlation response graph checking module: checking the fused correlation response graph, calculating the peak value and peak side lobe ratio, selecting whether a template filter is required to be introduced to judge the target state, and simultaneously selecting whether all correlation filters are updated;
a template filter module: when the correlation response graph checking module judges that the target position is predicted unreliably, inputting a current predicted target result into the template filter module, judging the target state, if the target is shielded, updating the target position by using the track prediction module, and if the target is shielded for a long time, losing the target and stopping tracking;
a trajectory prediction module: when the target is shielded, a track predictor is trained by utilizing the information of the center points of the previous multiple frames of targets, the center point of the target of the current frame is predicted, and the position information of the target is updated;
a model updating module: based on the result of the correlation response diagram checking module, analyzing the peak value and the peak value sidelobe ratio, and defining whether the model is updated or not and the updating frequency; if so, all correlation filters are updated.
The invention also discloses a holder target tracking method based on the multi-feature filter fusion of the tracking system, which comprises the following steps:
s1, extracting a plurality of manual features based on an image of an initial frame, a target and background information around the target given by a man-made or detector, training a correlation filter with corresponding features based on a kernel correlation filter target tracking method (kcf), respectively calculating feature maps corresponding to the current frame image, and obtaining a final response map by utilizing self-adaptive weight fusionObtaining the position of the center point of the target;
S2, utilizing the correlation filter with the maximum weight in the step S1 to predict in a scale set space which meets the relatively independent change trend of the length and the width of the target to obtain the scale of the new change of the target;
S3, verifying the fused correlation filtering response graph in the step S1 to obtain the peak value sidelobe ratio of the response graphAnalyzing the peak condition and the value distribution condition of the response diagram,higher values indicate more reliable prediction results of the multi-feature filter;
s4, extracting target features based on the image and the target information of the initial frame, training a template filter, after the peak value sidelobe ratio in the step S3 is evaluated, inputting the target information predicted in the steps S1 and S2 into the template filter to calculate the similarity between a prediction result and the template image so as to predict the shielding state of the target, and if the target is continuously shielded for a long time, judging that the target is lost;
s5, training a track predictor by utilizing the position information of a plurality of frames of targets before the current frame, and predicting the position information of the targets when the targets are shielded; in the step, the situation that the output target in the step S4 is shielded is processed through a track prediction module, and a track filter is realized based on linear Kalman filtering;
and S6, combining the response diagram peak value in the step S1 and the check module result in the step S3, adaptively defining whether each filter is updated or not and updating frequency, and respectively updating the plurality of filters in the step S1 and updating the template filter in the step S4.
It should be noted that, in the above method of the present invention, all filters are not updated with a fixed step size, but the rate of updating the filters is determined by setting the peak value and the peak-to-side lobe ratio of the fused response map in combination, which solves the following problems well: in the prior art, the problem of tracking failure caused by continuous introduction of background noise is solved by updating a fixed step size model in a related filtering target tracking algorithm.
Preferably, the aforementioned various manual features include: grayscale features, histogram gradient features (HOG features), color-name space features (CN features), and color features (RGB features).
Preferably, in the foregoing step S1, a final response map obtained by adaptive weight fusion is represented as:
, wherein ,representing image blocksFirst, theThe weight values of the individual feature dependent filters,for image blocksFirst, theA response map of the individual features is generated,represents the number of correlation filters;
,representing image blocksFirst, theThe characteristics of the composite material are that,is shown asA feature-dependent filter for the output of the image sensor,representing a convolution calculation.
More preferably, in the foregoing step S1, the weight value of the adaptive weightThe ratio of the peak value of each characteristic filter response map to the peak sum of all response maps is as follows:, wherein ,is shown asThe peak of the response map of the individual feature filters,is shown asThe peak of the response map of the individual feature filter;
Peak value of fused response mapCorresponding coordinate pointIs the position of the target center point, wherein 。
More preferably, in step S2, the prediction of the target scale is performed by selecting the feature filter with the largest weightThe method is realized by the following specific steps:
first, the length and width of the target are respectively determinedIntroducing a scale change poolForming a final set of scale-change spacesWherein the subscript、The length and the width of the target enclosing frame are respectively taken as objects of the scale change action, and subscripts 0, 1 and 2 respectively represent different scale change rates;
then, using the scale space setAnd predicting the position of the center point of the targetObtaining image regions under different scale spaces, wherein ,(ii) a Respectively extract the firstA characteristicAnd a characteristic filterAnd (3) calculating:
finally, by comparing the scale spaceCalculating the peak value of the response map under the middle 9 scalesScale of maximum correspondence of peak valueI.e. the new changed scale of the target:
A higher value indicates a more reliable prediction result, wherein,represents the final response graph obtained by the adaptive weight fusion in step S1,represents the proportion of the side lobe region centered on the peak value to the size of the whole response map region,representing the mean of the remaining part of the response map after removal of the sidelobe regions,the standard deviation of the remaining part of the response plot after removal of the sidelobe region is shown.
Further preferably, in step S4, based on the target information given by the initial frame, the bounding box of the target is not expanded, the background information is not included, and the image area is obtained and used as the template imageExtracting gray image characteristics, training template filter by using kernel correlation filtering technology(ii) a After target tracking beginsThe position of the target center point obtained by the predictionAnd predicted target scale informationObtaining a target prediction result imageBy passingAnd calculating the similarity between the target prediction result image and the template image to measure the state of the target and judge whether the target is blocked or lost.
Further preferably obtained by the check module in the pair S3When the value is evaluated, if it is less than the threshold valueIf so, indicating that the prediction reliability of the tracking result is low, extracting gray level characteristics from the image of the prediction target area, and performing correlation calculation with the template filter; if the peak value of the response image of the template filter is less than the threshold valueIf yes, the target is shielded, subsequent processing is needed, target state information is output, and the shielding of the target is accumulated and counted; when the target is shielded and the count is larger than the threshold valueIf the target is lost, the tracking is exited; if the peak value of the response image of the template filter is larger than the threshold valueIndicating that the target state is normal and the output destination is normalAnd marking state information, and clearing the shielding count of the target.
Still more preferably, in the step S5, the track predictor module is used to process the situation that the output target is blocked in S4, and the track filter is implemented based on linear kalman filtering. Specifically, the horizontal axis and vertical axis coordinates of the target center point position information are used as two-dimensional input, and the current frame is used as the previous frameAnd (3) training a track filter by using the coordinates of the target central point in the frame range, predicting the position information of the target central point of the current frame, and updating the target central point obtained by calculation in the S1.
It should be further explained that the criteria for model update are: when the filters in step S1 fuse the peaks of the response mapGreater than a threshold valueAnd the peak-to-side lobe ratio calculated by the check module in step S3Greater than a threshold valueUpdating each characteristic correlation filter and the template tracking filter by using the tracking result, or not updating;
the formula for model update is:
, wherein ,indicates the learning rate of each feature filter update,for the previous frameA feature-dependent filter is provided that is,for the updated current frameThe characteristics of the characteristic filter are used as the characteristic filter,retraining based on current frame target informationA feature correlation filter;
, wherein ,represents the learning rate of the template filter update,is a template filter for the previous frame,an updated target filter for the current frame,a template filter retrained based on current frame target information.
The invention has the advantages that:
(1) The pan-tilt tracking system and the pan-tilt tracking method realize pan-tilt target tracking through multi-feature filter fusion decision, solve the problem of common target scale change and the problem of special deformation of target appearance through designing scale spaces which are respectively based on different proportion changes of length and width, increase the accuracy and robustness of target tracking, and simultaneously can output the state information of the target in time, including correct tracking, shielding and loss.
(2) The method can predict the position information and the scale information of the target with given initial information in the subsequent image frame, utilizes the technology of self-adaptive weight fusion of a plurality of filter related to characteristics (at least comprising gray level characteristics, histogram gradient characteristics, RGB characteristics and color naming characteristics), self-adaptively distributes weight to each filter, and carries out weighting fusion to obtain the target position information, thereby fully exerting the characteristic representation capability of different characteristics, improving the accuracy of the tracking method, and controlling the calculated amount of a tracking core module so that the tracking core module can be applied to a holder platform;
(3) In the target tracking process, predicting the target scale in different scale spaces by using a characteristic filter with the largest weight, wherein the scale spaces comprise different changes of the length and the width of the target, so that accurate scale information of the target is obtained, and the problem of scale change in the target motion process is solved; meanwhile, training a template filter which takes a target image as a reference and does not contain background information, predicting the state information of the target by combining a response image weighted by a plurality of characteristic filters and the response image of the template filter, outputting the blocked or tracking lost state information of the target in time, introducing a target track predictor, and predicting the current target position information according to the track of a preorder frame when the target is blocked so as to better process the target blocking;
(4) In addition, all filters are not updated in fixed steps, and the updating rate of the filters is determined by setting the peak value and the peak-to-side lobe ratio of the fused response diagram. Through the organic combination of the check module and other modules, the updating rate and rhythm of each model are effectively controlled, and the robustness of the method is further improved.
Drawings
FIG. 1 is a block diagram of a pan-tilt target tracking system of the present invention;
FIG. 2 is a logic block diagram of a pan-tilt target tracking method of the present invention;
FIG. 3 is a logic diagram of the operation of the template filter of the present invention;
FIG. 4 is a logic diagram of the operation of the trajectory predictor of the present invention.
Detailed Description
The invention is described in detail below with reference to the figures and the embodiments.
Example 1
Referring to fig. 1, the embodiment discloses a pan-tilt-zoom target tracking system based on multi-feature filter fusion, which includes the following six functional modules:
(1) A correlation response graph fusion module: training a plurality of feature correlation filters by utilizing a kernel correlation filter based on a previous frame of image and target information, setting self-adaptive weights, fusing to generate a final correlation response diagram, and predicting the position of a target central point of a current frame;
(2) A target scale prediction module: predicting target scale information of the current frame in a scale change space according to the feature correlation filter with the largest weight;
(3) The correlation response graph checking module: checking the fused correlation response graph, calculating the peak value and peak side lobe ratio, selecting whether a template filter is required to be introduced to judge the target state, and simultaneously selecting whether all correlation filters are updated;
(4) A template filter module: when the correlation response graph checking module judges that the target position is predicted unreliably, inputting a current predicted target result into the template filter module, judging the target state, if the target is shielded, updating the target position by using the track prediction module, and if the target is shielded for a long time, losing the target and stopping tracking;
(5) A trajectory prediction module: when the target is shielded, a track predictor is trained by utilizing the information of the center points of the previous multiple frames of targets, the center point of the target of the current frame is predicted, and the position information of the target is updated;
(6) A model updating module: based on the result of the correlation response diagram checking module, analyzing the peak value and the peak value sidelobe ratio, and defining whether the model is updated or not and the updating frequency; if so, all correlation filters are updated.
The tracking system of the embodiment predicts the position information and the scale information of the target with given initial information in the subsequent image frame through the image data acquired by the pan-tilt. Firstly, tracking filters are respectively constructed by different features, the gray level features, the histogram gradient features, the RGB features and the color naming features are adopted, the weight is adaptively distributed to each filter, and weighting fusion is carried out to obtain target position information, so that the representation capability of the different features is fully utilized, and the calculated amount is controlled within a certain range.
And then, predicting the target scale in different scale spaces by using the characteristic filter with the largest weight, wherein the scale spaces comprise different changes of the length and the width of the target, further obtaining accurate scale information of the target, and solving the problem of scale change in the target motion process. Meanwhile, a template filter which takes the target image as a reference and does not contain background information is trained, the state information of the target is predicted by combining the response graphs weighted by the plurality of characteristic filters and the response graph of the template filter, and the state information that the target is shielded or lost by tracking is output in time. In addition, a target track predictor is introduced, and when the target is occluded, the current target position information is predicted according to the track of the preamble frame, so that the occlusion of the target can be better processed.
Moreover, all filters are not updated in fixed steps, but different condition thresholds are set to combine with peak-to-side lobe ratios of fused correlation filter response diagrams to determine the updating rate of the filters. The invention realizes the target tracking of the holder through the fusion tracking system of the multi-filter, increases the accuracy and the robustness of the target tracking, and can output the state information of the target in time, including correct tracking, shielding and loss.
Example 2
The present embodiment discloses a multi-tracking method implemented based on the tracking system of embodiment 1, and with reference to fig. 2, the method specifically includes the following steps:
s1, extracting a plurality of manual features based on an image of an initial frame, a target and background information around the target, wherein the image, the target and the background information are given by a human or a detector, and the manual features at least comprise: a grayscale feature, a histogram gradient feature (HOG feature), a color namespace feature (color-name, CN feature), and a color feature (RGB feature).
Then, training a correlation filter with corresponding characteristics based on a kernel correlation filter target tracking method (kcf), respectively calculating characteristic graphs corresponding to the current frame image, and obtaining a final response graph by utilizing self-adaptive weight fusionObtaining the position of the center point of the target。
Specifically, the final response graph obtained by the adaptive weight fusion is represented as:
, wherein ,which represents the number of the relevant filters,is an image blockFirst, theA response map of the individual features is generated,,representing image blocksFirst, theThe characteristics of the device are as follows,is shown asThe characteristic of each of the plurality of characteristic-dependent filters,representing a convolution calculation.
Representing image blocksFirst, theThe weighted value of each feature correlation filter is specifically the ratio of the peak value of each feature filter response map to the peak value sum of all response maps, that is:。
wherein ,is shown asThe peak of the response map of the individual feature filters,is shown asThe response pattern peak of the individual feature filter; namely:,;indicating that the maximum value is calculated.
Finally, the peak of the fused response mapCorresponding coordinate pointIs the position of the target center point, wherein 。
S2, predicting in a scale set space meeting the requirement that the target length and width change trend is relatively independent by using the correlation filter with the maximum weight in the step S1 to obtain the scale of the target new change。
The prediction of the target scale is realized by selecting the characteristic filter with the maximum weight valueThe method is realized by the following specific steps:
first, the length and width of the target are respectively determinedIntroducing a scale change poolForm the mostFinal set of scale-change spacesWherein, subscript、The length and the width of the target enclosing frame are respectively taken as objects of the scale change action, and subscripts 0, 1 and 2 respectively represent different scale change rates;
then, using the scale space setAnd the predicted target center point positionObtaining image regions in different scale spaces, wherein ,(ii) a Respectively extract the followingIs a characteristicAnd a characteristic filterAnd (3) calculating:
finally, by comparing the scale spacesCalculating the peak value of the response map under the middle 9 scalesScale of maximum correspondence of peak valueI.e. the new changed scale of the target:
s3, verifying the fused correlation filtering response graph in the step S1 to obtain the peak value sidelobe ratio of the response graphAnalyzing the peak condition and the value distribution condition of the response diagram,higher values indicate more reliable prediction results for the multi-feature filter.
Peak to side lobe ratio, wherein ,represents the final response graph obtained by the adaptive weight fusion in step S1,represents the proportion of the side lobe region centered on the peak value to the size of the whole response map region,representing the mean of the remaining portion of the response after removal of the sidelobe regions,the standard deviation of the remaining part of the response plot after removal of the sidelobe region is shown.
And S4, extracting target characteristics based on the image of the initial frame and the target information, training a template filter, as shown in FIG. 3, after the peak value sidelobe ratio in the step S3 is evaluated, inputting the target information predicted in the steps S1 and S2 into the template filter to calculate the similarity between the prediction result and the template image so as to predict the shielding state of the target, and if the target is continuously shielded for a long time, judging that the target is lost.
When the target information is given based on the initial frame, the surrounding frame of the target is not expanded, the background information is not contained, the image area is obtained and is used as the template imageExtracting gray image characteristics, training template filter by using kernel correlation filtering technology(ii) a After the target tracking is started, the target central point position obtained by the prediction is utilizedAnd predicted target scale informationObtaining a target prediction result imageBy passingCalculating the similarity between the target prediction result image and the template image to measure the state of the targetAnd judging whether the target is blocked or lost.
It is specifically stated that the check module obtains in the pair S3When the value is evaluated, if it is less than the threshold valueIf the prediction reliability of the tracking result is low, extracting gray features from the image of the prediction target area, and performing correlation calculation with the template filter; if the peak value of the response image of the template filter is less than the threshold valueIf yes, the target is shielded, subsequent processing is needed, target state information is output, and the shielding of the target is accumulated and counted; when the target is shielded and the count is larger than the threshold valueIf the target is lost, the tracking is exited; if the peak value of the response image of the template filter is larger than the threshold valueAnd indicating that the target state is normal and outputting target state information, and clearing the target by the shielding count. The threshold value is generally selected based on empirical values of those skilled in the art, and, in particular in the present embodiment,the value of the carbon dioxide is 0.3,the value of the carbon dioxide is 0.4,taking the value of 15.
S5, training a track predictor by utilizing the position information of a plurality of frames of targets before the current frame, and predicting the position information of the targets when the targets are shielded as shown in figure 4; in this step, the situation that the output target of step S4 is occluded is processed by a trajectory prediction module, and a trajectory filter is implemented based on linear kalman filtering.
Specifically, the horizontal axis and vertical axis coordinates of the target center point position information are used as two-dimensional input, and the current frame is used as the previous frameAnd (3) training a track filter by using the coordinates of the target central point in the frame range, predicting the position information of the target central point of the current frame, and updating the target central point obtained by calculation in the S1.
It should be further explained that the criteria for model update are: when the filters in step S1 fuse the peaks of the response mapGreater than a threshold valueAnd the peak-to-side lobe ratio calculated by the check module in step S3Greater than a threshold valueAnd updating each characteristic correlation filter and the template tracking filter by using the tracking result, or not updating. The threshold value here is likewise chosen on the basis of empirical values of a person skilled in the art, and, in particular in the present embodiment,the value is 0.7.
The formula for model update is:
, wherein ,the learning rate of each feature filter update is shown, and in this embodiment, a value of 0.012 is suggested,for the previous frameThe characteristics of the correlated filter are used to determine the characteristic,for the updated current frameThe characteristics of the characteristic filter are used as the characteristic filter,retraining based on current frame target informationA feature correlation filter;
, wherein ,which represents the learning rate of the updating of the template filter, the value of 0.025 is suggested in this embodiment,the template filter for the previous frame is used,the updated target filter for the current frame,a template filter retrained based on current frame target information.
And S6, combining the response diagram peak value in the step S1 and the check module result in the step S3, adaptively defining whether each filter is updated or not and updating frequency, and respectively updating the plurality of filters in the step S1 and updating the template filter in the step S4. That is, in the method of the present invention, all filters are not updated in fixed steps, but the rate of updating the filters is determined by setting the peak and peak-to-side lobe ratio of the fused response map in combination. Through the organic combination of the check module and other modules, the updating rate and rhythm of each model are effectively controlled, and the robustness of the method is further improved.
In summary, the pan-tilt tracking system and the tracking method of the invention realize pan-tilt target tracking through multi-feature filter fusion decision, and through designing the scale space based on different ratio changes of length and width, the problem of common target scale change is solved, the problem of special deformation of target appearance is solved, the accuracy and robustness of target tracking are increased, and meanwhile, the state information of the target including correct tracking, shielding and loss can be timely output, so that the feature representation capability of different features is fully exerted, the accuracy of the tracking method of the invention is improved, and the calculated amount of the tracking core module is controlled, so that the pan-tilt tracking system and the tracking method can be applied to a pan-tilt platform.
The foregoing illustrates and describes the principles, general features, and advantages of the present invention. It should be understood by those skilled in the art that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by using equivalent alternatives or equivalent variations fall within the scope of the present invention.
Claims (10)
1. The utility model provides a cloud platform target tracking system based on multi-feature filter fuses which characterized in that includes:
the correlation response graph fusion module: training a plurality of feature correlation filters by utilizing a kernel correlation filter based on a previous frame of image and target information, setting self-adaptive weights, fusing to generate a final correlation response diagram, and predicting the position of a target central point of a current frame;
a target scale prediction module: predicting target scale information of the current frame in a scale change space according to the feature correlation filter with the maximum weight;
the correlation response graph checking module: checking the fused correlation response graph, calculating the peak value and peak side lobe ratio, selecting whether a template filter is required to be introduced to judge the target state, and simultaneously selecting whether all correlation filters are updated;
a template filter module: when the correlation response graph checking module judges that the target position is predicted unreliably, inputting the current predicted target result into the template filter module, judging the target state, if the target is shielded, updating the target position by using the track prediction module, and if the target is shielded for a long time, losing the target and stopping tracking;
a trajectory prediction module: when the target is shielded, a track predictor is trained by utilizing the information of the center points of the previous multiple frames of targets, the center point of the target of the current frame is predicted, and the position information of the target is updated;
a model updating module: based on the result of the correlation response diagram checking module, analyzing the peak value and the peak value sidelobe ratio, and defining whether the model is updated or not and the updating frequency; if so, all correlation filters are updated.
2. A holder target tracking method based on multi-feature filter fusion is characterized by comprising the following steps:
s1, extracting various manual characteristics based on an image of an initial frame, a target and background information around the target, training relevant filters with different characteristics, calculating characteristic graphs corresponding to the current frame image respectively, and obtaining a final response graph by utilizing self-adaptive weight fusionObtaining the position of a target central point;
s2, predicting in a scale set space meeting the requirement that the target length and width change trend is relatively independent by using the correlation filter with the maximum weight in the step S1 to obtain the scale of the target new change;
S3, verifying the fused correlation filtering response graph in the step S1 to obtain the peak value sidelobe ratio of the response graphAnalyzing the peak condition and the numerical distribution condition of the response diagram;
s4, extracting target features based on the image and the target information of the initial frame, training a template filter, after the peak value sidelobe ratio in the step S3 is evaluated, inputting the target information predicted in the steps S1 and S2 into the template filter, calculating the similarity between a prediction result and the template image to predict the shielding state of the target, and if the target is continuously shielded for a long time, judging that the target is lost;
s5, training a track predictor by utilizing the position information of a plurality of frames of targets before the current frame, and predicting the position information of the targets when the targets are shielded;
and S6, combining the response diagram peak value in the step S1 and the check module result in the step S3, adaptively defining whether each filter is updated or not and updating frequency, and respectively updating the plurality of filters in the step S1 and updating the template filter in the step S4.
3. The pan-tilt target tracking method based on multi-feature filter fusion as claimed in claim 2, wherein in the step S1,,
wherein ,representing image blocksFirst, theThe weight values of the individual feature-dependent filters,is an image blockFirst, theThe response map of the individual features is,represents the number of correlation filters;
4. The pan-tilt target based on multi-feature filter fusion of claim 3The tracking method is characterized in that in the step S1, the weight value of the adaptive weightThe ratio of the peak value of each characteristic filter response map to the peak sum of all response maps is as follows:, wherein ,is shown asThe peak of the response map of the individual feature filters,is shown asThe response pattern peak of the individual feature filter;
5. The pan-tilt target tracking method based on multi-feature filter fusion as claimed in claim 2, wherein in the step S2, the target scale is predicted by selecting the feature filter with the largest weightTo realize the purpose of the method, the device is provided with a plurality of sensors,
first, the length and width of the target are respectively determinedIntroducing a scale change poolForming a final set of scale-change spacesWherein, subscript、Objects which respectively represent the scale change action are the length and the width of a target enclosing frame, and subscripts 0, 1 and 2 respectively represent different scale change rates;
then, using the scale space setAnd the predicted target center point positionObtaining image regions under different scale spaces, wherein ,(ii) a Respectively extract the followingA characteristicAnd a characteristic filterAnd (3) calculating:
finally, by comparing the scale spaceCalculating the peak value of the response map under the middle 9 scalesScale of maximum correspondence of peak valueI.e. the new changed scale of the target:
6. the pan-tilt-zoom target tracking method based on multi-feature filter fusion as claimed in claim 2, wherein in the step S3, the peak-to-side lobe ratio,
wherein ,represents the final response graph obtained by the adaptive weight fusion in step S1,represents the proportion of the side lobe region centered on the peak value to the size of the whole response map region,representing the mean of the remaining part of the response map after removal of the sidelobe regions,showing the standard deviation of the remaining portion of the response plot after removal of the sidelobe regions,higher values indicate more reliable prediction results.
7. The pan-tilt-zoom target tracking method based on multi-feature filter fusion as claimed in claim 2, wherein in step S4, based on the target information given by the initial frame, the image area is obtained and used as the template imageExtracting gray image characteristics, training template filter by using kernel correlation filtering technologyDuring training, background information and Hanning window processing are not introduced; after the target tracking is started, the target central point position obtained by the prediction is utilizedAnd predicted target scale informationObtaining a target prediction result imageBy passingAnd calculating the similarity between the target prediction result image and the template image to measure the state of the target and judge whether the target is blocked or lost.
8. The pan-tilt target tracking method based on multi-feature filter fusion as claimed in claim 6, wherein the target tracking method is obtained by the verification module in S3Evaluating the value if it is less than the threshold valueIf so, indicating that the prediction reliability of the tracking result is low, extracting gray level characteristics from the image of the prediction target area, and performing correlation calculation with the template filter; if the peak value of the response image of the template filter is less than the threshold valueIf the target is shielded, the subsequent processing is needed, the target state information is output, and the shielding of the target is accumulated and counted; when the target is shielded and the count is larger than the threshold valueIf the target is lost, the tracking is exited; if the peak value of the response image of the template filter is larger than the threshold valueAnd indicating that the target state is normal and outputting target state information, and clearing the target by the shielding count.
9. The holder target tracking method based on the multi-feature filter fusion as claimed in claim 2, wherein in the step S5, the track predictor module is used to process the condition that the output target is blocked in S4, and the track filter is implemented based on linear kalman filtering; the horizontal axis and vertical axis coordinates of the position information of the target central point are used as two-dimensional input, and the coordinates before the current frame are utilizedAnd (3) training a track filter by using the coordinates of the target central point in the frame range, predicting the position information of the target central point of the current frame, and updating the target central point obtained by calculation in the S1.
10. A pan-tilt target tracking method based on multi-feature filter fusion according to any one of claims 2 to 9, wherein in the step S6, the criteria for model update are as follows: when each filter fuses the peak value of the response map in step S1Greater than a threshold valueAnd the peak-to-side lobe ratio calculated by the check module in step S3Greater than a threshold valueUpdating each characteristic correlation filter and the template tracking filter by using the tracking result, or not updating;
the formula for model update is:
, wherein ,indicates the learning rate of each feature filter update,for the previous frameThe characteristics of the correlated filter are used to determine the characteristic,for the updated current frameThe characteristics of the characteristic filter are used as the characteristic filter,retraining based on current frame target informationA feature correlation filter;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211644874.1A CN115631216B (en) | 2022-12-21 | 2022-12-21 | Multi-feature filter fusion-based holder target tracking system and method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211644874.1A CN115631216B (en) | 2022-12-21 | 2022-12-21 | Multi-feature filter fusion-based holder target tracking system and method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN115631216A true CN115631216A (en) | 2023-01-20 |
CN115631216B CN115631216B (en) | 2023-05-12 |
Family
ID=84909630
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211644874.1A Active CN115631216B (en) | 2022-12-21 | 2022-12-21 | Multi-feature filter fusion-based holder target tracking system and method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115631216B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116228817A (en) * | 2023-03-10 | 2023-06-06 | 东南大学 | Real-time anti-occlusion anti-jitter single target tracking method based on correlation filtering |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110009060A (en) * | 2019-04-17 | 2019-07-12 | 东北大学 | A kind of robustness long-term follow method based on correlation filtering and target detection |
CN110276785A (en) * | 2019-06-24 | 2019-09-24 | 电子科技大学 | One kind is anti-to block infrared object tracking method |
CN112164093A (en) * | 2020-08-27 | 2021-01-01 | 同济大学 | Automatic person tracking method based on edge features and related filtering |
CN113989331A (en) * | 2021-11-12 | 2022-01-28 | 山西大学 | Long-term target tracking method based on context multi-clue information and adaptive response |
CN114612508A (en) * | 2022-02-28 | 2022-06-10 | 桂林电子科技大学 | Anti-occlusion related filtering target tracking method for multi-feature online learning |
-
2022
- 2022-12-21 CN CN202211644874.1A patent/CN115631216B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110009060A (en) * | 2019-04-17 | 2019-07-12 | 东北大学 | A kind of robustness long-term follow method based on correlation filtering and target detection |
CN110276785A (en) * | 2019-06-24 | 2019-09-24 | 电子科技大学 | One kind is anti-to block infrared object tracking method |
CN112164093A (en) * | 2020-08-27 | 2021-01-01 | 同济大学 | Automatic person tracking method based on edge features and related filtering |
CN113989331A (en) * | 2021-11-12 | 2022-01-28 | 山西大学 | Long-term target tracking method based on context multi-clue information and adaptive response |
CN114612508A (en) * | 2022-02-28 | 2022-06-10 | 桂林电子科技大学 | Anti-occlusion related filtering target tracking method for multi-feature online learning |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116228817A (en) * | 2023-03-10 | 2023-06-06 | 东南大学 | Real-time anti-occlusion anti-jitter single target tracking method based on correlation filtering |
CN116228817B (en) * | 2023-03-10 | 2023-10-03 | 东南大学 | Real-time anti-occlusion anti-jitter single target tracking method based on correlation filtering |
Also Published As
Publication number | Publication date |
---|---|
CN115631216B (en) | 2023-05-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107808122B (en) | Target tracking method and device | |
CN108416250B (en) | People counting method and device | |
CN103530893B (en) | Based on the foreground detection method of background subtraction and movable information under camera shake scene | |
CN102147851B (en) | Device and method for judging specific object in multi-angles | |
CN101339655B (en) | Visual sense tracking method based on target characteristic and bayesian filtering | |
CN108549841A (en) | A kind of recognition methods of the Falls Among Old People behavior based on deep learning | |
CN108470354A (en) | Video target tracking method, device and realization device | |
CN108846826A (en) | Object detecting method, device, image processing equipment and storage medium | |
CN103927762B (en) | Target vehicle automatic tracking method and device | |
CN107633226A (en) | A kind of human action Tracking Recognition method and system | |
CN110555870B (en) | DCF tracking confidence evaluation and classifier updating method based on neural network | |
CN110263660A (en) | A kind of traffic target detection recognition method of adaptive scene changes | |
CN104866868A (en) | Metal coin identification method based on deep neural network and apparatus thereof | |
CN111767847B (en) | Pedestrian multi-target tracking method integrating target detection and association | |
CN108446694A (en) | A kind of object detection method and device | |
CN107808138A (en) | A kind of communication signal recognition method based on FasterR CNN | |
CN108681689B (en) | Frame rate enhanced gait recognition method and device based on generation of confrontation network | |
CN112818871B (en) | Target detection method of full fusion neural network based on half-packet convolution | |
CN115631216A (en) | Holder target tracking system and method based on multi-feature filter fusion | |
CN115439654B (en) | Method and system for finely dividing weakly supervised farmland plots under dynamic constraint | |
CN110706208A (en) | Infrared dim target detection method based on tensor mean square minimum error | |
CN115641471A (en) | Countermeasure sample generation method and system based on generation of countermeasure network | |
CN105184815B (en) | Assemble a crowd event detecting method and system | |
CN112529941B (en) | Multi-target tracking method and system based on depth trajectory prediction | |
CN105551063A (en) | Method and device for tracking moving object in video |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
TA01 | Transfer of patent application right |
Effective date of registration: 20230419 Address after: No. 518 Zhongshan East Road, Qinhuai District, Nanjing City, Jiangsu Province, 210000 Applicant after: Jincheng Group Co.,Ltd. Applicant after: AVIC Jincheng Unmanned System Co.,Ltd. Address before: 210000 no.216, Longpan Middle Road, Qinhuai District, Nanjing City, Jiangsu Province Applicant before: AVIC Jincheng Unmanned System Co.,Ltd. |
|
TA01 | Transfer of patent application right | ||
GR01 | Patent grant | ||
GR01 | Patent grant |