CN108648213A - A kind of implementation method of KCF track algorithms on TMS320C6657 - Google Patents
A kind of implementation method of KCF track algorithms on TMS320C6657 Download PDFInfo
- Publication number
- CN108648213A CN108648213A CN201810220756.5A CN201810220756A CN108648213A CN 108648213 A CN108648213 A CN 108648213A CN 201810220756 A CN201810220756 A CN 201810220756A CN 108648213 A CN108648213 A CN 108648213A
- Authority
- CN
- China
- Prior art keywords
- kcf
- implementation method
- algorithms
- track algorithms
- kcf track
- 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.)
- Pending
Links
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
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- 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
- G06T2207/20048—Transform domain processing
- G06T2207/20056—Discrete and fast Fourier transform, [DFT, FFT]
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to target following technical field, a kind of implementation method of KCF track algorithms on TMS320C6657 is disclosed, based on the hardware platform of TMS320C6657 processors, to two-dimensional fast fourier transform FFT2D algorithm optimizations;Language level code optimization, including loop unrolling are carried out to KCF algorithms, using Inline Function, optimize program structure.The result of actual measurement shows that the image of 320 × 240 pixels, when tracking target is 64 × 32 pixel, wave door size is 128 × 64.The tracking time of each frame is about 15310000 cpu cycles, if DSP is operated under the frequency of 1GHz, taking for each frame is no more than 15.5 milliseconds, and about each second handles 64 frame images, meets requirement of real-time.
Description
Technical field
The invention belongs to a kind of reality of target following technical field more particularly to KCF track algorithms on TMS320C6657
Existing method.
Background technology
Currently, the prior art commonly used in the trade is such:Computer vision is one and lives in frontier of knowledge, very
The research field of jump, it is different from the research to mankind or animal vision in biology, but is known by means of image procossing, pattern
Not, the knowledge such as mathematics and physics come establish model and processing data.To allow computer to have intelligent understanding to extraneous object,
It can assist it that object with the real world is identified, be like that computer has " vision ", widen the view and see the world.
Computer vision replaces human eye to be observed and detected using machine, can not only improve precision and detection efficiency, Er Qieke
To be difficult to reach applied to some human eyes or human eye is difficult to the occasion met the requirements.The recognition and tracking of target is always to calculate
One research hotspot of machine visual field.Identification refers mainly to computer according to certain feature and experience accumulation, for receiving
Certain information (be typically image), analyze content therein, its attribute differentiated;And it is typically on having sequential to track
On the basis of hereafter, continuous tracing and positioning is carried out for the target in computer " visual field ".In traditional target tracking domain
In, it usually needs the image of target is pre-saved as template.Into when line trace, chosen using some search strategies in the visual field
It may be the region undetermined of target, compare the similarity of template and region to be searched according to mathematics or physical method later, finally
The size of similarity is tracked to determine the position of target, completes tracking.Typical search strategy has global search, slightly matches refinement
Matching etc., common similarity calculating method include Euclidean distance, histogram COS distance, score distance, cut that distance, this is more by person of outstanding talent
Husband's distance, histograms of oriented gradients, fuzzy characteristics comparison etc., further include color histogram etc. for colored image.But
The feature that traditional tracking uses is all relatively simple, and tracking result is barely satisfactory, and needs the mould that target is obtained ahead of time
Plate image, is difficult to realize in some cases.On this basis, numerous scholars have investigated the tracking of on-line study formula,
This kind of method can be directed to arbitrary target and carry out real-time tracking.Operator only needs the initial end in tracking sequence artificial
Specified target (it includes mesh target area that one is generally referred on PC), track algorithm will be according to selected target
It practises, the model obtained according to study is into line trace in new video frame, and the result that algorithm obtains tracking is as sample, into one
Step is learnt, and is updated to model, is so moved in circles, is continuously tracked to video sequence.Because this one kind is calculated
Method can adaptively learning objective model, therefore the various interference for occurring in sequence have stronger adaptability, KCF
Algorithm is exactly a wherein very classical track algorithm.KCF track algorithms by learn regularization least square grader come
Differentiate target, and introduce circular matrix and cyclic shift is carried out to training sample, can approximation regard as, pass through that learn regularization minimum
Two multiply grader to differentiate target, and introduce circular matrix to training sample carry out cyclic shift, can approximation regard as to target
Intensive sampling then utilize circular matrix by correlation filter and kernel function to efficiently obtaining a large amount of training sample
It connects, to propose core correlation filter (KCF).In recent years, with high speed digital signal processor (digital
Signal processer, DSP) development, have many high performance DSP that can quickly carry out the mathematical operation of large amount of complex,
Larger memory space can also be improved simultaneously, meets the real-time demand in terms of data processing, with the TMS320 series of TI companies
For, the chips such as C6416, C665x, C667x in C6000 series are obtained all in real-time target following and are widely answered
With.But most of track algorithm in such platform is all the fairly simple traditional algorithm of such as MAD, NCC, tracking effect
Fruit is barely satisfactory.Since data volume is big in object tracking process for the new algorithms such as KCF, data correlation is high, hinders it in reality
Application in the engineering of border.Therefore how KCF algorithms to be optimized using the characteristics and advantages of DSP platform, becomes realization and calculates
The key of method tracking system.TMS320C6657 is that the having based on KeyStone frameworks of TI companies publication pinpoints and floating-point meter
The double-nuclear DSP of calculation ability.Chip interior integrates two C66x series kernels, and each core highest may operate at 1.25GHz clock frequencies
Rate, monokaryon performance are four times of TMS320C64x+ series DSP processing capacities, and monoblock chip possesses 80GMACS or 40GFLOPS fortune
Calculation ability.C66x kernels are integrated with 90 completely new instructions for floating-point and towards vector math calculation process, are very suitable for counting
The functions such as word signal processing and figure acquisition.In addition the C6000 fixed points before C66x kernels and TI and Floating-point DSP kernel are to offspring
Code is compatible, it is ensured that software portability shortens software development cycle.TMS320C6657 multi-core DSPs are integrated with abundant on piece
Storage resource, single kernel possess the level-one program cache area of 32KB sizes and the level one data buffer area of 32KB sizes, and
It can be configured to the special region of memory of the 1024KB sizes of RAM or caching.Meanwhile two kernels share the interior of 1024KB sizes
Deposit space.Further include 32 DDR3 that can be operated in 1333MHz frequencies in addition, in order to quickly access chip external memory
External memory interface.For different application scenarios, a large amount of Peripheral Interface resources of integrated chip:The four roads channels SRIO, it is single
Channel highest transmission speed is up to 5Gbps;One gigabit Ethernet mouth;Two generation PCIe interfaces;HyperLink high-speed interfaces;It is logical
With parallel serial ports (uPP) etc..KCF algorithm introductions:Among the research in vision tracking field in recent years, one of maximum breakthrough
It is exactly widely to introduce the method learnt based on discrimination model.The problem of vision tracks can naturally be converted into one
The problem of a on-line study.A given initial pictures block for including tracking target, the purpose of tracking is then to train one
A identification and classification device distinguishes tracking target and background.To in the training process of identification and classification device, it is easy to will pay close attention to
Point is placed on interested tracking target --- for the positive sample of classifier training.But the core principles of method of discrimination are then
Same or more focus are placed on the background around tracking target --- for the negative sample of classifier training.Most often
The method of acquisition negative sample is that the different location around tracking target chooses a certain number of image blocks, then is carried out to it
Scaling in varying degrees and transformation.During acquiring negative sample, a very challenging factor is then
It is --- the negative sample that can be obtained from an image is almost inexhaustible.But since vision is tracked to real-time
Requirement, present Vision Tracking can only find between negative sample quantity as much as possible and relatively low operand
One equalization point.Common practice is a certain number of negative samples of selection random in each frame for classifier training, without
It is to take principle exhausted.KCF track algorithms are broadly divided into detection, training and update three phases.During sample training,
A large amount of training sample is obtained by carrying out cyclic shift to selected target, classification is then trained by the sample of acquisition
Device.During target detection, the related coefficient between area to be tested and actual tracking target, choosing are calculated using kernel function
Take related coefficient is maximum to wait for favored area as new target.Meanwhile using two-dimensional fast fourier transform (FFT2D) by time domain
It is mapped to frequency domain, reduces the operation in training and detection process.Finally when fast target detects, quick Fu of two dimension is utilized again
In leaf inverse transformation (IFFT2D) frequency domain is mapped back into time domain, obtain the position that target in present frame occurs.The algorithm is revolved in target
Turn, partial occlusion etc. has good tracking effect.Currently, all concentrating on anti-serious screening for most of research of KCF
Gear, anti-target loss, anti-dimensional variation etc., it is intended that stable long time-tracking method is developed into, still, for KCF
Application of the method in Practical Project is but seldom related to.Although the KCF algorithm speeds of service are very fast, it is needed during tracking
Carry out a large amount of FFT2D operations, for realize platform computing capability have very high requirement, if platform computing capability compared with
Difference can then extend the time of tracking consumption, lose the advantage of KCF track algorithms;During tracking simultaneously, need largely to assist
Matrix is also larger for the demand of memory for calculating.In conclusion problem of the existing technology is:Although KCF is flat in PC
The effect of platform is fine, but is applied in Practical Project not yet.A large amount of floating point arithmetic involved in KCF algorithms,
The calculating speed of platform is required very high;The operation involved in algorithm is mostly based on complex field simultaneously, therefore for memory
Consumption be significantly larger than traditional algorithm.The different module such as tracking, detection, study all respectively needs some complementary squares in algorithm
Gust, what is preserved in these matrixes is all complex number type floating number, and committed memory is 8 times of same size gray level image, is accounted for memory
With quite greatly, the memory of embedded platform is generally smaller, it is difficult to which directly transplanting is realized.And in KCF algorithms include a large amount of
Exponent arithmetic, matrix operation, FFT operations, calculate considerably complicated, for the more demanding of platform, realize that difficulty is bigger.
In conclusion problem of the existing technology is:KCF algorithms are applied in Practical Project not yet, and
Its calculating is considerably complicated, for the more demanding of platform, realizes that difficulty is bigger.
Solve the difficulty and meaning of above-mentioned technical problem:KCF algorithms are during tracking, it is necessary first to selected target, so
Needing the size according to institute's selected target to choose pending image block afterwards, (usual image block size is the 2.5 of selected target
Times), the size for calculating required companion matrix in the algorithm is all related to the size of image block.It is largely transported in KCF algorithms
At last based on floating number, but the operation of floating number or the operation of floating number are not supported in some Embedded platforms
Efficiency will be far below the operation efficiency of fixed-point number.The also most important, a large amount of operations carried out in the algorithm are all bases
In complex field, therefore the data stored in matrix need to include real part and imaginary part, are typically embedded into formula platform and utilize 8-bit type numbers
According to image is preserved, such as char types, but if preserving complex number type floating number, the memory consumed is 8 times of original image size,
The size of image can be handled by greatly limiting.Meanwhile in calculating process including a large amount of exponent arithmetic, matrix operation, FFT
Operation, if needed to calculate the Gaussian kernel correlation between characteristics of image and trace model when tracking, detection, when more new model, needs
There are dot product, the point division operation of matrix, it is extremely complex and time-consuming, for embedded platform, a large amount of such operation meeting pair
CPU causes prodigious load.If do not optimized to algorithm, the consumption of memory conference cause tracking during memory overflow
Go out, tracking is gently then caused to fail, it is heavy then can cause program crashing or even chip is burnt.
For KCF as one of the algorithm that public attention is most contained in the world in recent years, the performance in terms of tracking is very good.
But it is limited to the limitation of PC platforms, it can not be used in many occasions, such as the terminal guidance from motion tracking, guided missile of unmanned plane
Etc..In fact, these application scenarios are very big for the demand of outstanding track algorithm, but this kind of application scenarios
In, track algorithm is conventionally based on embedded platform.If can solve the problems, such as that KCF algorithms are transplanted in embedded platform, just
KCF algorithms can be applied in a large amount of simple device, such as intelligent video camera head, unmanned plane, even military aspect.At these
, can be on the basis of target following in application scenarios, the function of Improving Equipment improves its applicability, convenient in target following
On the basis of further extended function.
Invention content
In view of the problems of the existing technology, the reality the present invention provides a kind of KCF track algorithms on TMS320C6657
Existing method.
The invention is realized in this way a kind of implementation method of KCF track algorithms on TMS320C6657, the KCF with
Implementation method of the track algorithm on TMS320C6657 include:Based on the hardware platform of TMS320C6657 processors, realize that KCF is calculated
Method.In order to ensure the speed of service of algorithm, without using piece external storage, all data are distributed and are all stored in piece in calculating process
Middle completion, in particular it is necessary to which the picture frame of processing is stored in internuclear shared drive, such selection can both save CPU
Core enjoys memory certainly, while also allowing for the cooperation operation of subsequent development multinuclear, and image data can directly be read in shared drive,
Avoid address conversion operation.According to KCF algorithm flows, training, detection-phase are required for calculating the Gauss nuclear phase between characteristics of image
Guan Xing realizes the quick meter to Gaussian kernel correlation so first according to kernel function and the property of circular matrix in DSP platform
It calculates.On this basis, the modules such as training and the detection in KCF algorithms are completed, while also having the basic operations largely used ---
FFT2D operations.Algorithm operates on one of TMS320C6657 CPU cores, when handling new frame image, is read in MSM
Required image data is taken, so the memory size that algorithm can use is 1MB.Due to needing to consume prodigious memory in algorithm
As auxiliary space when calculating, therefore it is susceptible to memory overflow problem, on this basis, needs further to optimize,
Ensure feasibility and real-time, including to two-dimensional fast fourier transform FFT2D algorithm optimizations;Language level is carried out to KCF algorithms
Code optimization, including loop unrolling, use Inline Function;Optimize program structure, optimizes the distribution etc. of memory space.
Further, implementation method of the KCF track algorithms on TMS320C6657 optimizes FFT2D, for defeated
The image to be calculated entered carries out FFT transform to each row, then carries out FFT transform to each row, by being obtained after converting twice
Be exactly FFT2D result.
Further, memory space is smaller, needs auxiliary vector t, and data line is copied in vectorial t, calculates FFT knots
Fruit is written original position, recycles this process and terminate until all rows all calculate, and completes row transformation;The data of each row are answered successively
It makes in vectorial t, calculates FFT result, be as a result written in original matrix, recycle the operation until completing all row, obtain being FFT2D
Result.
Further, implementation method of the KCF track algorithms on TMS320C6657 calculates floating-point using cossp functions
Several cosine value, for calculating Hamming window;Gaussian matrix is calculated using sqrtsp, divsp and expsp;Use DSPF_sp_
Maxidx, DSPF_sp_maxval come the maximum value in quick obtaining matrix and subscript.
Further, image to be tracked is stored in by implementation method of the KCF track algorithms on TMS320C6657
In the MSM of DSP, auxiliary data needed for track algorithm is distributed in enjoying certainly for CPU core in memory.
Further, the Compiler Optimization of implementation method of the KCF track algorithms on TMS320C6657 is provided by TI
Programming tool CCS complete.
Further, implementation method of the KCF track algorithms on TMS320C6657 replaces integer using shifting function
Multiplication and division operation;Cycle is optimized, larger interior cycle is used.
In conclusion advantages of the present invention and good effect are:The present invention proposes a kind of reality by DSP embedded platforms
When Target Tracking System implementation, based on the hardware platform of TMS320C6657 processors, to two-dimensional fast fourier transform
(FFT2D) algorithm optimization is carried out, language level code optimization, including loop unrolling secondly are carried out to KCF algorithms, use inline letter
Number, optimization program structure etc., the exquisite place of KCF algorithms is, is used in each frame to identification and classification device thousands of negative
Sample is trained, to ensure precision when tracking.At the same time, which is become using circular matrix by discrete fourier again
The characteristic under frequency domain after (DFT) is changed, avoids and directly traverses each negative sample, to ensure that algorithm has high reality
Shi Xing.The result of actual measurement shows that the image of 320 × 240 pixels, when tracking target is 64 × 32 pixel, wave door size is 128 ×
64.The tracking time of each frame is about 15310000 cpu cycles, if DSP is operated under the frequency of 1GHz, each frame
Take be no more than 15.5 milliseconds, about each second handle 64 frame images, meet requirement of real-time.
Description of the drawings
Fig. 1 is implementation method flow chart of the KCF track algorithms provided in an embodiment of the present invention on TMS320C6657.
Fig. 2 is the target location schematic diagram provided in an embodiment of the present invention exported according to algorithm.
Fig. 3 is the effect application of implementation method of the KCF track algorithms provided in an embodiment of the present invention on TMS320C6657
Schematic diagram.
Specific implementation mode
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
Artificial intelligence, image procossing, pattern-recognition, target detection, machine learning etc. have been merged in the target following of the present invention
Advanced technology in multiple fields.It obtains mesh by detection, extraction, identification to target object in image sequence and study
The relevant informations such as target shape, position, speed, track, then by the analysis and processing to relevant information, realize and target is moved
The understanding of behavior is finally completed the tasks such as tracking, prediction.
As shown in Figure 1, implementation method of the KCF track algorithms provided in an embodiment of the present invention on TMS320C6657 includes
Following steps:
S101:Based on the hardware platform of TMS320C6657 processors, two-dimensional fast fourier transform (FFT2D) is carried out
Algorithm optimization;
S102:Language level code optimization, including loop unrolling are carried out to KCF algorithms, using Inline Function, optimize program knot
Structure.
The application principle of the present invention is further described below in conjunction with the accompanying drawings.
1, the realization and optimization of DSP platform:
(1) there is higher effect than DFT2D using FFT2D come to replace DFT2D, FFT2D be two-dimensional Fast Fourier Transform (FFT)
Rate.It is not only complicated but if FFT2D operations are realized in pure hand weaving, but also take huge.FFT2D operations and IFFT2D operations exist
It is repeatedly called in the present invention, is the bottleneck place of real-time, if using hand-code, it is impossible to reach requirement of real-time.It is first
First FFT2D is optimized.DSP platform has very good embedded support for mathematic(al) manipulations such as FFT, but only one-dimensional
FFT function APIs for using, according to the property of FFT2D, two-dimensional FFT result can be calculated using one-dimensional FFT.For defeated
The image to be calculated entered first carries out FFT transform to each row, and referred to as row variation carries out FFT transform to each row again later,
Referred to as rank transformation, by convert twice obtain later be exactly FFT2D result.FFT operations for each row or column, can
It is calculated with calling the DSPF_sp_fftSPxSP functions in dsplib, this is the library function that TI officials provide, for FFT operations
The optimization of hardware view has been carried out, the operational capability of DSP is made full use of, computational efficiency can be greatly speeded up.Similarly, IFFT2D
Same mode may be used to be calculated.In this calculating process, row transformation and the sequencing of rank transformation do not influence most
Whole result.If memory space is smaller, an auxiliary vector t is only needed.Data line is copied into vectorial t successively first
In, FFT result is calculated, original position is written, recycles this process and terminates until all rows all calculate, completes row transformation at this time.Then
The data of each row are copied in vectorial t successively, FFT result is calculated, writes the result into original matrix, cycle is until completing institute
The operation for having row, what is obtained at this time is exactly the result of FFT2D;
(2) use Inline Function, Inline Function that can be mapped directly into the compilation of C6000, efficiency of code execution and remittance
It compiles quite, far above common pure hand weaving code.TI companies provide abundant auxiliary library function, mathlib, dsplib,
Imglib etc., which provide a large amount of common Inline Function API, for realizing the operation of fixed point and floating number.Occur in program
A large amount of matrix dot product, point division operation and the vectorial modulus value of calculating etc. take longer operation, can be used in C6657 inline
Function optimizes.Cossp functions are such as used to calculate the cosine value of floating number, for calculating Hamming window;Using sqrtsp,
Divsp and expsp etc. calculates Gaussian matrix;Carry out quick obtaining matrix using DSPF_sp_maxidx, DSPF_sp_maxval
In maximum value and subscript etc..Such as the calculating of Gaussian function, if using simple standard c functions execute once-through operation when
Between about 7000us, reasonable employment Inline Function can be reduced to 2000us, and efficiency greatly promotes.
(3) data store optimization, when PC platforms run track algorithm, the distribution and recycling of memory are all by operating system
Agency, but the memory in DSP needs developer to distribute, the problems such as otherwise spilling it is easy to appear memory.It is tracked in KCF
In algorithm, calculating, detection process, training process of FFT2D etc. are required for using companion matrix, if being not added with optimization, DSP
Memory use far from enough, can then reduce computational efficiency using piece external storage.By the compression to companion matrix, the piece of C6657
Upper resource has met the EMS memory occupation of program.In the implementation of this paper, image to be tracked is stored in the MSM of DSP
In (Multicoreshared memory, internuclear shared drive), the auxiliary data needed for track algorithm is interior in enjoying certainly for CPU core
Deposit middle distribution.That is, the double-core of C6657 can respectively run a trace routine, can track in piece image simultaneously
Two targets, can further develop multiple target tracking.
(4) Compiler Optimization, the programming tool CCS provided by TI provide powerful compiling optimization come what is completed in CCS
It instructs to complete the function of optimization, by reasonably selecting compiling option, the operational efficiency of code can be promoted.Compiler it is excellent
It is as shown in table 1 to change option:
1 Compiler Optimization option of table
(5) some other optimizations of program level, such as replace the multiplication and division operation of integer using shifting function;To recycle into
Row optimization, improves the execution efficiency of code using larger interior cycle as possible;Reasonable employment keyword and global variable are reduced
The code of redundancy improves execution efficiency.
The application effect of the present invention is explained in detail with reference to experiment.
1, experimental result is demonstrated:
Under CCS environment, makes to show a C language KCF track algorithms, and optimize, be transplanted to TMS320C6657
Development board, to resolution ratio be 320 × 240 image sequence into line trace, target sizes are 64 × 32, and wave door size is 2 times of mesh
Mark size.
Track algorithm only operates in monokaryon, and DSP reads in video sequence in MSM successively, after reading, CPU operation with
Track algorithm exports the position the result is that target.In first frame, algorithm will extract feature according to the initial position of target, carry out
Training, will subsequently be continuously tracked target.The tracking situation of object observing and processing time confirm the feasibility of system
And time performance.It according to the target location that algorithm exports, is marked in artwork, it is observed that just whether the result of tracking
Really.As shown in Figure 2.Four width figures are the tracking situations of different frame in video sequence, it can be seen that in each frame, algorithm is obtained
Position and observation obtain the result is that coincideing, even if in 164 frames, billboard is slight to vehicle to be blocked, but still
It accurately tracks.It can thus be seen that algorithm is successful for the tracking of target, as shown in Figure 3.
In the 165 to 173rd frame, black vehicle is seriously blocked by billboard, it can be seen that in this process, algorithm is
It realizes and accurately tracks, and after black reappears, still tenacious tracking.This is that traditional NCC algorithms cannot achieve
's.
It in summary it can be seen, the algorithm keeps track effect that the present invention realizes is stable, and compared with traditional NCC algorithms,
Situations such as blocking, is more robust.Before carrying out storage optimization, there is memory spilling at runtime in DSP6657, leads to program
It interrupts, can not continue to track.After carrying out storage optimization, when not carrying out FFT2D optimizations and compiling optimization, the run time of each frame
About 243 milliseconds, the speed of service is very slow.After the method proposed in the present invention optimizes, when the operation of each frame
Between about 15.5 milliseconds, efficiency improves about 16 times, has reached real-time requirement.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention
All any modification, equivalent and improvement etc., should all be included in the protection scope of the present invention made by within refreshing and principle.
Claims (7)
1. a kind of implementation method of KCF track algorithms on TMS320C6657, which is characterized in that the KCF track algorithms exist
Implementation method on TMS320C6657 includes:Based on the hardware platform of TMS320C6657 processors, to two-dimentional fast Fourier
Convert FFT2D algorithm optimizations;Language level code optimization, including loop unrolling are carried out to KCF algorithms, use Inline Function, optimization
Program structure.
2. implementation method of the KCF track algorithms as described in claim 1 on TMS320C6657, which is characterized in that described
Implementation method of the KCF track algorithms on TMS320C6657 optimizes FFT2D, for the image to be calculated of input, to every
A line carry out FFT transform, then to each row carry out FFT transform, by convert twice obtain later be exactly FFT2D result.
3. implementation method of the KCF track algorithms as claimed in claim 2 on TMS320C6657, which is characterized in that storage is empty
Between it is smaller, need auxiliary vector t, data line copied in vectorial t, calculate FFT result, be written original position, recycle this mistake
Journey terminates until all rows all calculate, and completes row transformation;The data of each row are copied in vectorial t successively, calculate FFT knots
As a result fruit is written in original matrix, recycle the operation until completing all row, obtain be FFT2D result.
4. implementation method of the KCF track algorithms as described in claim 1 on TMS320C6657, which is characterized in that described
Implementation method of the KCF track algorithms on TMS320C6657 calculates the cosine value of floating number using cossp functions, for calculating
Hamming window;Gaussian matrix is calculated using sqrtsp, divsp and expsp;Use DSPF_sp_maxidx, DSPF_sp_maxval
Come the maximum value in quick obtaining matrix and subscript.
5. implementation method of the KCF track algorithms as described in claim 1 on TMS320C6657, which is characterized in that described
Image to be tracked is stored in the MSM of DSP by implementation method of the KCF track algorithms on TMS320C6657, track algorithm institute
The auxiliary data needed is distributed in enjoying certainly for CPU core in memory.
6. implementation method of the KCF track algorithms as described in claim 1 on TMS320C6657, which is characterized in that described
The Compiler Optimization of implementation method of the KCF track algorithms on TMS320C6657 is completed by the programming tool CCS that TI is provided.
7. implementation method of the KCF track algorithms as described in claim 1 on TMS320C6657, which is characterized in that described
Implementation method of the KCF track algorithms on TMS320C6657 replaces the multiplication and division operation of integer using shifting function;To recycle into
Row optimization, uses larger interior cycle.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810220756.5A CN108648213A (en) | 2018-03-16 | 2018-03-16 | A kind of implementation method of KCF track algorithms on TMS320C6657 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810220756.5A CN108648213A (en) | 2018-03-16 | 2018-03-16 | A kind of implementation method of KCF track algorithms on TMS320C6657 |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108648213A true CN108648213A (en) | 2018-10-12 |
Family
ID=63744246
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810220756.5A Pending CN108648213A (en) | 2018-03-16 | 2018-03-16 | A kind of implementation method of KCF track algorithms on TMS320C6657 |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108648213A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109978779A (en) * | 2019-03-12 | 2019-07-05 | 东南大学 | A kind of multiple target tracking device based on coring correlation filtering method |
CN111415370A (en) * | 2020-04-13 | 2020-07-14 | 中山大学 | Embedded infrared complex scene target real-time tracking method and system |
CN112183493A (en) * | 2020-11-05 | 2021-01-05 | 北京澎思科技有限公司 | Target tracking method, device and computer readable storage medium |
CN114546413A (en) * | 2022-02-21 | 2022-05-27 | 深圳市佳贤通信设备有限公司 | Method for automatic deployment and automatic optimization of monitoring function time consumption |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170018090A1 (en) * | 2015-07-15 | 2017-01-19 | Thomson Licensing | Method and apparatus for object tracking in image sequences |
CN107610151A (en) * | 2017-08-24 | 2018-01-19 | 上海汇纳信息科技股份有限公司 | Pedestrian track line handling method/system, computer-readable recording medium and equipment |
-
2018
- 2018-03-16 CN CN201810220756.5A patent/CN108648213A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170018090A1 (en) * | 2015-07-15 | 2017-01-19 | Thomson Licensing | Method and apparatus for object tracking in image sequences |
CN107610151A (en) * | 2017-08-24 | 2018-01-19 | 上海汇纳信息科技股份有限公司 | Pedestrian track line handling method/system, computer-readable recording medium and equipment |
Non-Patent Citations (3)
Title |
---|
徐松等: "TMS320C6678多核DSP导航制导软件框架设计", 《工业控制计算机》 * |
杜文彬等: "基于DSP6678的KCF算法实现及优化系统设计", 《国外电子测量技术》 * |
王永俊等: "基于TMS320C6678核相关滤波器跟踪算法实现及改进", 《电子技术应用》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109978779A (en) * | 2019-03-12 | 2019-07-05 | 东南大学 | A kind of multiple target tracking device based on coring correlation filtering method |
CN111415370A (en) * | 2020-04-13 | 2020-07-14 | 中山大学 | Embedded infrared complex scene target real-time tracking method and system |
CN112183493A (en) * | 2020-11-05 | 2021-01-05 | 北京澎思科技有限公司 | Target tracking method, device and computer readable storage medium |
CN114546413A (en) * | 2022-02-21 | 2022-05-27 | 深圳市佳贤通信设备有限公司 | Method for automatic deployment and automatic optimization of monitoring function time consumption |
CN114546413B (en) * | 2022-02-21 | 2024-06-04 | 深圳市佳贤通信科技股份有限公司 | Method for time-consuming automatic deployment and automatic optimization of monitoring function |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Yi et al. | ASSD: Attentive single shot multibox detector | |
Xu et al. | Light-YOLOv3: fast method for detecting green mangoes in complex scenes using picking robots | |
CN108805796A (en) | Dynamic accuracy management for integer deep learning primitive | |
CN108648213A (en) | A kind of implementation method of KCF track algorithms on TMS320C6657 | |
CN110046537A (en) | The system and method for carrying out dynamic face analysis using recurrent neural network | |
CN104680558B (en) | The Struck method for tracking target accelerated using GPU hardware | |
Liu et al. | Coastline extraction method based on convolutional neural networks—A case study of Jiaozhou Bay in Qingdao, China | |
Bacea et al. | Single stage architecture for improved accuracy real-time object detection on mobile devices | |
Rahman et al. | Parallel implementation of a spatio-temporal visual saliency model | |
Kang et al. | RCS-YOLO: A fast and high-accuracy object detector for brain tumor detection | |
Zhang et al. | Saliency hierarchy modeling via generative kernels for salient object detection | |
Wang et al. | Multi-tailed vision transformer for efficient inference | |
Song et al. | A parallel Canny edge detection algorithm based on OpenCL acceleration | |
Sariturk et al. | Comparison of residual and dense neural network approaches for building extraction from high-resolution aerial images | |
Ban et al. | Real-Time object detection based on convolutional block attention module | |
Wang et al. | MGA-YOLOv4: a multi-scale pedestrian detection method based on mask-guided attention | |
Zhu et al. | Scene text detection with selected anchors | |
Cong-An et al. | Feature aligned ship detection based on improved RPDet in SAR images | |
CN105869105A (en) | Method for accelerating GPU directed at A+ super-resolution technology | |
Tuli et al. | A novel face detector based on YOLOv3 | |
Yong et al. | Efficient parallel recursive Gaussian SIFT algorithm based on multi-core DSP | |
Sun et al. | Distilling Siamese Trackers with Attention Mask | |
Liu et al. | Design and implementation of scale adaptive kernel correlation filtering algorithm based on hls | |
Korytkowski et al. | Fast computing framework for convolutional neural networks | |
CN116883726B (en) | Hyperspectral image classification method and system based on multi-branch and improved Dense2Net |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181012 |
|
RJ01 | Rejection of invention patent application after publication |