CN109559329A - A kind of particle filter tracking method based on depth denoising autocoder - Google Patents
A kind of particle filter tracking method based on depth denoising autocoder Download PDFInfo
- Publication number
- CN109559329A CN109559329A CN201811433093.1A CN201811433093A CN109559329A CN 109559329 A CN109559329 A CN 109559329A CN 201811433093 A CN201811433093 A CN 201811433093A CN 109559329 A CN109559329 A CN 109559329A
- Authority
- CN
- China
- Prior art keywords
- encoder
- particle
- weight
- training
- automatic
- 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
- 239000002245 particle Substances 0.000 title claims abstract description 71
- 238000000034 method Methods 0.000 title claims abstract description 56
- 230000008569 process Effects 0.000 claims abstract description 26
- 238000005070 sampling Methods 0.000 claims abstract description 7
- 238000012549 training Methods 0.000 claims description 43
- 238000013528 artificial neural network Methods 0.000 claims description 36
- 230000006870 function Effects 0.000 claims description 34
- 230000009467 reduction Effects 0.000 claims description 24
- 230000004913 activation Effects 0.000 claims description 22
- 239000011159 matrix material Substances 0.000 claims description 6
- 230000007704 transition Effects 0.000 claims description 6
- 238000013519 translation Methods 0.000 claims description 6
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000000295 complement effect Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 230000000644 propagated effect Effects 0.000 claims description 3
- 230000001902 propagating effect Effects 0.000 claims description 3
- 230000004580 weight loss Effects 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 7
- 230000000007 visual effect Effects 0.000 description 5
- 230000008859 change Effects 0.000 description 4
- 238000005286 illumination Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 210000003710 cerebral cortex Anatomy 0.000 description 1
- 238000012885 constant function Methods 0.000 description 1
- 230000006735 deficit Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000000513 principal component analysis Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 239000013585 weight reducing agent Substances 0.000 description 1
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
-
- 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
-
- 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
- G06T2207/20084—Artificial neural networks [ANN]
-
- 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
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Multimedia (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to computer vision analysis technical fields, more particularly to a kind of particle filter tracking method based on depth denoising autocoder, including initial phase, using the manual spotting position of video first frame, it is tracked in process and tracking processing in first frame, it needs respectively in target background and prospect a certain number of positive negative samples selected around, initialize trained network model, second step, carry out importance sampling, third step, calculate observation probability, 4th step, update weight, 5th step, judge and selects the maximum particle of weight in weight information, think the particle be we next the target to be tracked, tracking new samples are updated for next frame, process of the circulation second step to the 5th step, until video playing finishes, this method can efficiently differentiate target signature and background, improve track algorithm Precision.
Description
Technical Field
The invention belongs to the technical field of computer vision analysis, and particularly relates to a particle filter tracking method based on a depth denoising automatic encoder.
Background
Visual target tracking is an important research direction for computer vision and visual analysis. Typical visual analysis requires consistent and stable tracking of the object of interest. For monocular visual target tracking, numerous scholars propose theories and algorithms worthy of reference. In practical applications, the problem still faces a huge challenge due to the influence of factors such as complex background, target occlusion, rapid target motion, illumination change and the like.
The deep neural network has strong learning ability in the aspects of target detection and target classification. The deep learning architecture is more suitable for learning classification features than specific targets. In addition, the deep neural network algorithm usually requires a long iterative training process to converge, and the real-time requirement of online learning is difficult to meet. Therefore, it is difficult to extend the current deep learning network architecture to the field of target tracking.
Disclosure of Invention
The invention provides a particle filter tracking method based on a depth denoising automatic encoder, aiming at solving the problems of interference such as complex background, light change, target shielding and the like in the target tracking process and the problem of poor anti-jamming capability of the existing tracking algorithm. The technical problem to be solved by the invention is realized by the following technical scheme:
a particle filter tracking method based on a depth denoising automatic encoder comprises the following steps: step 1: training a deep network model, carrying out unsupervised greedy training on each layer of network layer by layer, adding noise into training data to obtain more stable feature expression, and carrying out supervised learning on the features through a classification neural network to further optimize parameters of the network;
step 2: manually calibrating the target position by adopting a first frame of a video, selecting positive and negative training samples from the sequence, and initializing the deep network model in the step 1;
and step 3: sampling particle sets by adopting importance, then propagating each particle forwards through a trained network model, and calculating the confidence coefficient of each particle in the online tracking process through a classification neural network;
and 4, step 4: calculating the observation probability of the particles according to the confidence coefficient of the particles in the step 3;
and 5: and (5) updating the weight of the particles according to the observation probability in the step (4) to determine the target position, updating and tracking a new sample for the next frame, and circulating the processes from the step (3) to the step (5) until the video is played.
Further, the depth network model in the step 1 is overlapped by an automatic noise reduction coder, and the output of the next layer is used as the input of the upper layer; the automatic noise reduction encoder comprises an encoder, a decoder and an implicit layer, wherein the decoder needs to predict original undamaged data according to noise characteristics and finally outputs the closest original input, Gaussian noise is usually used as an attenuation vector, and the expression of the Gaussian noise is as follows:
where x is the original input without noise interference,is data contaminated by noise, and σ represents the degree of regularization of the auto-encoder.
Further, the training process in step 1 is as follows: assume a training sample set x ∈ R for unlabeled classesdMapping the input x to the hidden layer by the activation function f to obtain z ∈ Rd
z∈fθ(x)＝σ(Wx+b) (1)
Where θ is { W, b }, W is a weight matrix, b is a coding layer vector, σ is a non-linear activation function, and the decoder remaps the input coded representation to form reconstructed y
y∈fθ′(h)＝σ(W′h+b′) (2)
Where θ '{ W', b '}, W' is the transpose of the weight matrix W, and σ is the activation function of decoding; the automatic noise reduction encoder enables y to be approximately equal to x through the process;
hypothesis training set { (x)(1)，y(1))，...，(x(m)，y(m)) Contains m training samples, x represents a single sample feature, y represents the input to which the sample corresponds, and a single sample (x, y) is taken to define its cost function;
wherein h isW，b(x) Is corresponding to a networkThe output value of sample x, and thus the cost function of the training set of m samples, is:
λ is the weight loss coefficient, controlling the relative importance of the two parts; the process of training an automatic noise reduction encoder is to adjust the minimum reconstruction error J (W, b) of the parameters { theta, theta' } in the training sample set, where J (W, b) is a convex function, usually optimized by an iterative method.
Further, the classification neural network comprises an automatic noise reduction encoder coding part and k classification layers connected with sparse constraints.
Further, the classification neural network learning method in the step 1 is as follows, and Z is an activation function of a hidden layer of the self-encoder. In the forward propagation phase, the activation function Z is:
z＝WTx+b (6)
wherein x is an input vector; w is a weight; b is a bias (bias).
The first K maxima of the activation function are maintained and all others are set to zero:
wherein, (gamma)cIs complementary to z, (Γ)c＝sup pk(z). The sparsity z is used to calculate the network reconstruction error:
where x is the training sample set, W represents the weight, b' represents the transpose of the bias (bias), the weight is back-propagated by the first K maxima of the activation function output to reconstruct the error iteration adjustment.
Further, the confidence coefficient algorithm in step 3 is as follows: let oiIs corresponding to class kiThe output of the neural network, the expectation of the output value is the posterior probability.
E{oi}＝P(ki|x) (9)
Where x is the network input. Typically, the corresponding class of the largest output is taken as the decision, so the confidence can be obtained from the posterior probability of the neural network, and the largest output of the classified neural network is taken as the confidence:
c(x)＝E{max oi} (10)
further, the importance sampling method in step 3 is as follows:
when a new frame image arrives, q(s) is distributed according to the importancet|st-1，y1：t) And a motion model, from the set of particles at time t-1Obtaining n particles at time tWherein the importance weight corresponding to the set of particlesSum of (S)tIs 1; target state stS is represented by six affine parameters, horizontal translation, vertical translation, scaling, width/height ratio, rotation and skewt(tx, ty, sxy, ra, ar, sa); distribution per dimension of state transitionsIn the motion model, the model is an independent zero-mean normal distribution model.
Further, the observation probability calculation method in step 4 is as follows:
each particle is propagated forward through the classification neural network to obtain its confidenceAnd will maximize confidenceComparing with the set threshold value tau ifReselecting the positive and negative training samples, and initializing a classification neural network; if it is notThe probability of observation of the particle is calculated as follows:
wherein y istIt means that the sample at time t corresponds to the input,refers to the ith particle at time t.
Further, the method for updating the weight of the particle in step 5 includes:
wherein,each dimension representing a state transition in the importance sample is distributed,i.e. the calculated probability distribution of the particles, the distribution q(s) in generalt|st-1，y1：t) Using a first order Markov process q(s)t|st-1) I.e. the state transition is independent of the model observations, then the weights are updated as:
whereinIndicating the weight at the time immediately before the update,representing the observation probability of the particles obtained by the previous step, wherein for each frame, the particle with the maximum weight is a tracking result; updating a positive sample for each tracking frame, and then tracking the next positive sample; determining a state corresponding to the largest particle as a frame target position outside the current vehicle
The invention has the beneficial effects that:
the automatic noise reduction encoder obtains distributed characteristic representation of high-dimensional complex input through unsupervised greedy training layer by layer and parameter optimization multi-layer network structure, and only needs to adjust network parameters for different tasks; the method can effectively distinguish the target characteristics from the background through the automatic noise reduction encoder for deep noise reduction; and a classification neural network is introduced, so that the classification capability of the network is improved, the precision of a tracking algorithm is improved, and finally, particle filtering is used for tracking the target.
Drawings
FIG. 1 is a schematic diagram of an automatic noise reduction encoder.
Fig. 2 is a schematic diagram of a classification neural network structure.
Fig. 3 is a diagram illustrating the tracking result of the indoor occlusion phenomenon.
Fig. 4 is a diagram illustrating the tracking result of the outdoor occlusion phenomenon.
Fig. 5 is a diagram illustrating the tracking result of the illumination change target.
Fig. 6 is a schematic diagram of a target fuzzy target tracking result.
Detailed Description
The present invention will be described in further detail with reference to specific examples, but the embodiments of the present invention are not limited thereto.
A particle filter tracking method based on a depth denoising automatic encoder comprises the following steps:
step 1: training a deep network model, carrying out unsupervised greedy training on each layer of network layer by layer, adding noise into training data to obtain more stable feature expression, and carrying out supervised learning on the features through a classification neural network to further optimize parameters of the network;
step 2: manually calibrating the target position by adopting a first frame of a video, selecting positive and negative training samples from the sequence, and initializing the deep network model in the step 1;
and step 3: sampling particle sets by adopting importance, then propagating each particle forwards through a trained network model, and calculating the confidence coefficient of each particle in the online tracking process through a classification neural network;
and 4, step 4: calculating the observation probability of the particles according to the confidence coefficient of the particles in the step 3;
and 5: and (5) updating the weight of the particles according to the observation probability in the step (4) to determine the target position, updating and tracking a new sample for the next frame, and circulating the processes from the step (3) to the step (5) until the video is played.
As shown in fig. 1, in step 1, the depth network model is overlapped by an automatic noise reduction encoder, the depth automatic encoder is a typical unsupervised learning network, which is a depth network model, overlapped by a self-encoder, and uses the output of the next layer as the input of the upper layer, the essence of the automatic encoder is to learn the same function, i.e. the input of the network is equal to the reconstructed output, and the training and parameter optimization process is to realize output and reproduction input; the automatic noise reduction encoder comprises an encoder, a decoder and an implicit layer; the auto noise reduction encoder accepts the corrupted data as input and predicts the original uncorrupted data as output by training. The purpose of the noise reduction automatic encoder is to allow the use of a very large encoder and simultaneously prevent useless constant functions between the encoder and the decoder, and based on a statistical theory, the core idea of the automatic noise reduction encoder is to disturb original input and noise according to a certain rule, so that the original input is damaged, damaged data is input into a network, and the representation of a hidden layer is obtained. The decoder needs to predict the original uncorrupted data according to the noise characteristics, and finally outputs the closest original input, which is the effect of removing interference, and gaussian noise is usually used as an attenuation vector, and its expression is:
where x is the original input without noise interference,which is noise contaminated data, and sigma represents the degree of regularization of the auto-encoder, for the problem of generation of impairment data, the binomial random numbers are not only simple but also easy to calculate, i.e., binomial distribution random numbers are generated using the same input shape and then multiplied by the input. We use the squared error function as the reconstruction error and train it in exactly the same way as other feed forward networks.
The training process in step 1 is as follows: assume a training sample set x ∈ R for unlabeled classesdMapping the input x to the hidden layer by the activation function f to obtain z ∈ Rd
z∈fθ(x)＝σ(Wx+b) (1)
Where θ is { W, b }, W is a weight matrix, b is a coding layer vector, σ is a non-linear activation function, and the decoder remaps the input coded representation to form reconstructed y
y∈fθ′(h)＝σ(W′h+b′) (2)
Where θ '{ W', b '}, W' is the transpose of the weight matrix W, and σ is the activation function of decoding; the automatic noise reduction encoder enables y to be approximately equal to x through the process;
hypothesis training set { (x)(1)，y(1)，...，(x(m)，y(m)) Contains m training samples, x represents a single sample feature, y represents the input to which the sample corresponds, and a single sample (x, y) is taken to define its cost function;
wherein h isW，b(x) Is the output value of sample x corresponding to the network, so the cost function of the training set of m samples is:
it can be seen that the first part of the equation is the mean variance term of the cost function. The second part is a weight attenuation term which can prevent the weight from changing too much so as to prevent overfitting, and lambda is a weight reduction coefficient which controls the relative importance of the two parts; the process of training an automatic noise reduction encoder is to adjust the minimum reconstruction error J (W, b) of the parameters { theta, theta' } in the training sample set, where J (W, b) is a convex function, usually optimized by an iterative method.
The classification neural network comprises an automatic noise reduction encoder coding part and classification layers connected with k sparse constraints.
The purpose of constructing a classification neural network is to calculate the confidence of each particle in the online tracking process. The automatic noise reduction encoder consists of an encoding part of the automatic noise reduction encoder and k sparse constraint connected classification layers, and a schematic diagram of a classification neural network structure is shown in FIG. 2; the invariant characteristic of the target can be effectively learned by introducing k sparse constraint, the linear discrimination capability of the classification neural network is improved, and the overfitting problem is solved to a certain extent. Neuroscience research shows that the response of visual signals in cerebral cortex is sparse, so that the expression of original signals can be more meaningful by introducing sparse restriction in a deep neural network, particularly for a classification task, the idea is verified in principal component analysis and sparse coding, K sparse restriction keeps K maximum activation functions of hidden layers, and the rest are set to be zero.
The classified neural network learning method in the step 1 is as follows, and Z is an activation function of a hidden layer of a self-encoder. In the forward propagation phase, the activation function Z is:
z＝WTx+b (6)
wherein x is an input vector; w is a weight; b is a bias (bias).
The first K maxima of the activation function are maintained and all others are set to zero:
wherein, (gamma)cIs complementary to z, (Γ)c＝sup pk(z). The sparsity z is used to calculate the network reconstruction error:
where x is the training sample set, W represents the weight, b' represents the transpose of the bias (bias), the weight is back-propagated by the first K maxima of the activation function output to reconstruct the error iteration adjustment. The confidence level of the classification neural network output is the confidence level, which reflects its decision confidence at some point in the feature vector space.
The confidence coefficient algorithm in the step 3 is as follows: let oiIs corresponding to class kiThe output of the neural network, the expectation of the output value is the posterior probability.
E{oi}＝P(ki|x) (9)
Where x is the network input. Typically, the corresponding class of the largest output is taken as the decision, so the confidence can be obtained from the posterior probability of the neural network, and the largest output of the classified neural network is taken as the confidence:
c(x)＝E{max oi} (10)
the importance sampling method in step 3 is as follows:
when a new frame image arrives, q(s) is distributed according to the importancet|st-1，y1：t) And a motion model, from the set of particles at time t-1Obtaining n particles at time tWherein the importance weight corresponding to the set of particlesSum of (S)tIs 1; target state stS is represented by six affine parameters, horizontal translation, vertical translation, scaling, width/height ratio, rotation and skewt(tx, ty, sxy, ra, ar, sa); distribution per dimension of state transitionsIs an independent zero-mean normal score in the motion modelAnd (4) cloth modeling.
The observation probability calculation method in step 4 is as follows:
each particle is propagated forward through the classification neural network to obtain its confidenceAnd will maximize confidenceComparing with the set threshold value tau ifReselecting the positive and negative training samples, and initializing a classification neural network; if it is notThe probability of observation of the particle is calculated as follows:
wherein y istIt means that the sample at time t corresponds to the input,refers to the ith particle at time t.
The method for updating the weight of the particles in the step 5 comprises the following steps:
wherein,each dimension representing a state transition in the importance sample is distributed,i.e. the calculated probability distribution of the particles, the distribution q(s) in generalt|st-1，y1：t) Using a first order Markov process q(s)t|st-1) I.e. the state transition is independent of the model observations, then the weights are updated as:
whereinIndicating the weight at the time immediately before the update,representing the observation probability of the particles obtained by the previous step, wherein for each frame, the particle with the maximum weight is a tracking result; updating a positive sample for each tracking frame, and then tracking the next positive sample; the state corresponding to the largest particle is determined as the frame target position outside the current vehicle.
The experimental operating environment is as follows: 3.8GHz, quad-core AMD processor, 8GB memory. Video sequences in various environments are used for verification, including illumination change, target occlusion and target rapid motion.
Fig. 3 and 4 show occlusion of an object. The shielding phenomenon is the phenomenon that a target is partially shielded due to the complexity of the surrounding environment and the interference of other surrounding objects, and the target cannot be lost in the whole tracking process of the tracker; outdoor photography often produces strong light changes. When the light changes greatly, the performance of target tracking is affected, and as can be seen from fig. 5, when a target enters a tunnel, huge illumination changes exist in an image, but from the tracking result, the algorithm accurately completes the tracking task.
The problem of target blurring appears in fig. 6, the target blurring is caused by too fast speed of the target in the moving process or unstable photography, the image of the target in the image is not clear, the target is blurred, the tracking effect is affected, the tracking algorithm in the text accurately completes tracking, and no target is lost.
The foregoing is a more detailed description of the invention in connection with specific preferred embodiments and it is not intended that the invention be limited to these specific details. For those skilled in the art to which the invention pertains, several simple deductions or substitutions can be made without departing from the spirit of the invention, and all shall be considered as belonging to the protection scope of the invention.
Claims (9)
1. A particle filter tracking method based on a depth denoising automatic encoder is characterized in that:
step 1: training a deep network model, carrying out unsupervised greedy training on each layer of network layer by layer, adding noise into training data to obtain more stable feature expression, and carrying out supervised learning on the features through a classification neural network to further optimize parameters of the network;
step 2: manually calibrating the target position by adopting a first frame of a video, selecting positive and negative training samples from the sequence, and initializing the deep network model in the step 1;
and step 3: sampling particle sets by adopting importance, then propagating each particle forwards through a trained network model, and calculating the confidence coefficient of each particle in the online tracking process through a classification neural network;
and 4, step 4: calculating the observation probability of the particles according to the confidence coefficient of the particles in the step 3;
and 5: and (5) updating the weight of the particles according to the observation probability in the step (4) to determine the target position, updating and tracking a new sample for the next frame, and circulating the processes from the step (3) to the step (5) until the video is played.
2. The particle filter tracking method based on the automatic deep denoising encoder as claimed in claim 1, wherein: in the step 1, the deep network models are overlapped by an automatic noise reduction coder, and the output of the next layer is used as the input of the upper layer; the automatic noise reduction encoder comprises an encoder, a decoder and an implicit layer, wherein the decoder needs to predict original undamaged data according to noise characteristics and finally outputs the closest original input, Gaussian noise is usually used as an attenuation vector, and the expression of the Gaussian noise is as follows:
where x is the original input without noise interference,is data contaminated by noise, and σ represents the degree of regularization of the auto-encoder.
3. The particle filter tracking method based on the automatic deep denoising encoder as claimed in claim 1, wherein: the training process in the step 1 is as follows: assume a training sample set x ∈ R for unlabeled classesdMapping the input x to the hidden layer by the activation function f to obtain z ∈ Rd
z∈fθ(x)＝σ(Wx+b) (1)
Where θ is { W, b }, W is a weight matrix, b is a coding layer vector, σ is a non-linear activation function, and the decoder remaps the input coded representation to form reconstructed y
y∈fθ′(h)＝σ(W′h+b′) (2)
Where θ '{ W', b '}, W' is the transpose of the weight matrix W, and σ is the activation function of decoding; the automatic noise reduction encoder enables y to be approximately equal to x through the process;
hypothesis training set { (x)(1)，y(1))，...，(x(m)，y(m)) Contains m training samples, x represents a single sample feature, y represents the input to which the sample corresponds, and a single sample (x, y) is taken to define its cost function;
wherein h isW，b(x) Is the output value of sample x corresponding to the network, so the cost function of the training set of m samples is:
λ is the weight loss coefficient, controlling the relative importance of the two parts; the process of training an automatic noise reduction encoder is to adjust the minimum reconstruction error J (W, b) of the parameters { theta, theta' } in the training sample set, where J (W, b) is a convex function, usually optimized by an iterative method.
4. The particle filter tracking method based on the automatic deep denoising encoder as claimed in claim 1, wherein: the classification neural network comprises an automatic noise reduction encoder coding part and k classification layers connected with sparse constraints.
5. The particle filter tracking method based on the automatic deep denoising encoder as claimed in claim 1, wherein: the classified neural network learning method in the step 1 is as follows, and Z is an activation function of a hidden layer of a self-encoder. In the forward propagation phase, the activation function Z is:
z＝WTx+b (6)
wherein x is an input vector; w is a weight; b is a bias (bias).
The first K maxima of the activation function are maintained and all others are set to zero:
wherein, (gamma)cIs complementary to z, (Γ)c＝sup pk(z). The sparsity z is used to calculate the network reconstruction error:
where x is the training sample set, W represents the weight, b' represents the transpose of the bias (bias), the weight is back-propagated by the first K maxima of the activation function output to reconstruct the error iteration adjustment.
6. The particle filter tracking method based on the automatic deep denoising encoder as claimed in claim 1, wherein: the confidence coefficient algorithm in the step 3 is as follows: let oiIs corresponding to class kiThe output of the neural network, the expectation of the output value is the posterior probability.
E{oi}＝P(ki|x) (9)
Where x is the network input. Typically, the corresponding class of the largest output is taken as the decision, so the confidence can be obtained from the posterior probability of the neural network, and the largest output of the classified neural network is taken as the confidence:
c(x)＝E{maxoi} (10)
7. the particle filter tracking method based on the automatic deep denoising encoder as claimed in claim 1, wherein: the importance sampling method in the step 3 is as follows:
when a new frame image arrives, q(s) is distributed according to the importancet|st-1，y1∶t) And a motion model, from the set of particles at time t-1Obtaining n particles at time tWherein the importance weight corresponding to the set of particlesSum of (S)tIs 1; target state StS is represented by six affine parameters, horizontal translation, vertical translation, scaling, width/height ratio, rotation and skewt(tx, ty, sxy, ra, ar, sa); distribution per dimension of state transitionsIn the motion model, the model is an independent zero-mean normal distribution model.
8. The particle filter tracking method based on the automatic deep denoising encoder as claimed in claim 1, wherein: the observation probability calculation method in the step 4 is as follows:
each particle is propagated forward through the classification neural network to obtain its confidenceAnd will maximize confidenceComparing with the set threshold value tau ifReselecting the positive and negative training samples, and initializing a classification neural network; if it is notThe probability of observation of the particle is calculated as follows:
wherein y istIt means that the sample at time t corresponds to the input,refers to the ith particle at time t.
9. The particle filter tracking method based on the automatic deep denoising encoder as claimed in claim 1, wherein: the method for updating the weight of the particles in the step 5 comprises the following steps:
wherein,each dimension representing a state transition in the importance sample is distributed,i.e. the calculated probability distribution of the particles, the distribution q(s) in generalt|st-1，y1：t) Using a first order Markov process q(s)t|st-1) I.e. the state transition is independent of the model observations, then the weights are updated as:
whereinIndicating the weight at the time immediately before the update,representing the observation probability of the particles obtained by the previous step, wherein for each frame, the particle with the maximum weight is a tracking result; updating a positive sample for each tracking frame, and then tracking the next positive sample; the state corresponding to the largest particle is determined as the frame target position outside the current vehicle.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811433093.1A CN109559329B (en) | 2018-11-28 | 2018-11-28 | Particle filter tracking method based on depth denoising automatic encoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811433093.1A CN109559329B (en) | 2018-11-28 | 2018-11-28 | Particle filter tracking method based on depth denoising automatic encoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109559329A true CN109559329A (en) | 2019-04-02 |
| CN109559329B CN109559329B (en) | 2023-04-07 |
Family
ID=65867657
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811433093.1A Active CN109559329B (en) | 2018-11-28 | 2018-11-28 | Particle filter tracking method based on depth denoising automatic encoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109559329B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110473557A (en) * | 2019-08-22 | 2019-11-19 | 杭州派尼澳电子科技有限公司 | A kind of voice signal decoding method based on depth self-encoding encoder |
| CN110825123A (en) * | 2019-10-21 | 2020-02-21 | 哈尔滨理工大学 | Control system and method for automatic following loading vehicle based on motion algorithm |
| CN110889459A (en) * | 2019-12-06 | 2020-03-17 | 北京深境智能科技有限公司 | Learning method based on edge and Fisher criterion |
| CN111552322A (en) * | 2020-04-29 | 2020-08-18 | 东南大学 | Unmanned aerial vehicle tracking method based on LSTM-particle filter coupling model |
| CN111563423A (en) * | 2020-04-17 | 2020-08-21 | 西北工业大学 | Unmanned aerial vehicle image target detection method and system based on depth denoising automatic encoder |
| CN111735458A (en) * | 2020-08-04 | 2020-10-02 | 西南石油大学 | Navigation and positioning method of petrochemical inspection robot based on GPS, 5G and vision |
| CN111931368A (en) * | 2020-08-03 | 2020-11-13 | 哈尔滨工程大学 | UUV target state estimation method based on GRU particle filter |
| CN111950503A (en) * | 2020-06-16 | 2020-11-17 | 中国科学院地质与地球物理研究所 | Aviation transient electromagnetic data processing method and device and computing equipment |
| CN112396635A (en) * | 2020-11-30 | 2021-02-23 | 深圳职业技术学院 | Multi-target detection method based on multiple devices in complex environment |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2002171436A (en) * | 2000-12-04 | 2002-06-14 | Sony Corp | Image processing apparatus |
| CN103211563A (en) * | 2005-05-16 | 2013-07-24 | 直观外科手术操作公司 | Methods and system for performing 3-D tool tracking by fusion of sensor and/or camera derived data during minimally invasive robotic surgery |
| CN105654509A (en) * | 2015-12-25 | 2016-06-08 | 燕山大学 | Motion tracking method based on composite deep neural network |
| CN105894008A (en) * | 2015-01-16 | 2016-08-24 | 广西卡斯特动漫有限公司 | Target motion track method through combination of feature point matching and deep nerve network detection |
| CN106127804A (en) * | 2016-06-17 | 2016-11-16 | 淮阴工学院 | The method for tracking target of RGB D data cross-module formula feature learning based on sparse depth denoising own coding device |
| CN106203350A (en) * | 2016-07-12 | 2016-12-07 | 北京邮电大学 | A kind of moving target is across yardstick tracking and device |
| CN107403222A (en) * | 2017-07-19 | 2017-11-28 | 燕山大学 | A kind of motion tracking method based on auxiliary more new model and validity check |
| CN108460760A (en) * | 2018-03-06 | 2018-08-28 | 陕西师范大学 | A kind of Bridge Crack image discriminating restorative procedure fighting network based on production |
-
2018
- 2018-11-28 CN CN201811433093.1A patent/CN109559329B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2002171436A (en) * | 2000-12-04 | 2002-06-14 | Sony Corp | Image processing apparatus |
| CN103211563A (en) * | 2005-05-16 | 2013-07-24 | 直观外科手术操作公司 | Methods and system for performing 3-D tool tracking by fusion of sensor and/or camera derived data during minimally invasive robotic surgery |
| CN105894008A (en) * | 2015-01-16 | 2016-08-24 | 广西卡斯特动漫有限公司 | Target motion track method through combination of feature point matching and deep nerve network detection |
| CN105654509A (en) * | 2015-12-25 | 2016-06-08 | 燕山大学 | Motion tracking method based on composite deep neural network |
| CN106127804A (en) * | 2016-06-17 | 2016-11-16 | 淮阴工学院 | The method for tracking target of RGB D data cross-module formula feature learning based on sparse depth denoising own coding device |
| CN106203350A (en) * | 2016-07-12 | 2016-12-07 | 北京邮电大学 | A kind of moving target is across yardstick tracking and device |
| CN107403222A (en) * | 2017-07-19 | 2017-11-28 | 燕山大学 | A kind of motion tracking method based on auxiliary more new model and validity check |
| CN108460760A (en) * | 2018-03-06 | 2018-08-28 | 陕西师范大学 | A kind of Bridge Crack image discriminating restorative procedure fighting network based on production |
Non-Patent Citations (5)
| Title |
|---|
| 季顺平 等: "基于HSV颜色特征和贡献度重构的行人跟踪", 《激光与光电子学进展》 * |
| 杨红红等: "基于稀疏约束深度学习的交通目标跟踪", 《中国公路学报》 * |
| 焦婷 等: "存在运动目标时的图像镶嵌方法研究", 《计算机应用研究》 * |
| 程帅 等: "多示例深度学习目标跟踪", 《电子与信息学报》 * |
| 邓俊锋: "基于稀疏自动编码器和边缘降噪自动编码器的深度学习算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110473557A (en) * | 2019-08-22 | 2019-11-19 | 杭州派尼澳电子科技有限公司 | A kind of voice signal decoding method based on depth self-encoding encoder |
| CN110473557B (en) * | 2019-08-22 | 2021-05-28 | 浙江树人学院（浙江树人大学） | Speech signal coding and decoding method based on depth self-encoder |
| CN110825123A (en) * | 2019-10-21 | 2020-02-21 | 哈尔滨理工大学 | Control system and method for automatic following loading vehicle based on motion algorithm |
| CN110889459A (en) * | 2019-12-06 | 2020-03-17 | 北京深境智能科技有限公司 | Learning method based on edge and Fisher criterion |
| CN110889459B (en) * | 2019-12-06 | 2023-04-28 | 北京深境智能科技有限公司 | Learning method based on edge and Fisher criteria |
| CN111563423A (en) * | 2020-04-17 | 2020-08-21 | 西北工业大学 | Unmanned aerial vehicle image target detection method and system based on depth denoising automatic encoder |
| CN111552322A (en) * | 2020-04-29 | 2020-08-18 | 东南大学 | Unmanned aerial vehicle tracking method based on LSTM-particle filter coupling model |
| CN111950503A (en) * | 2020-06-16 | 2020-11-17 | 中国科学院地质与地球物理研究所 | Aviation transient electromagnetic data processing method and device and computing equipment |
| CN111950503B (en) * | 2020-06-16 | 2024-01-30 | 中国科学院地质与地球物理研究所 | Aviation transient electromagnetic data processing method and device and computing equipment |
| CN111931368A (en) * | 2020-08-03 | 2020-11-13 | 哈尔滨工程大学 | UUV target state estimation method based on GRU particle filter |
| CN111735458A (en) * | 2020-08-04 | 2020-10-02 | 西南石油大学 | Navigation and positioning method of petrochemical inspection robot based on GPS, 5G and vision |
| CN112396635A (en) * | 2020-11-30 | 2021-02-23 | 深圳职业技术学院 | Multi-target detection method based on multiple devices in complex environment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109559329B (en) | 2023-04-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109559329B (en) | Particle filter tracking method based on depth denoising automatic encoder | |
| CN108133188B (en) | Behavior identification method based on motion history image and convolutional neural network | |
| Mane et al. | Moving object detection and tracking using convolutional neural networks | |
| Ma et al. | Variational implicit processes | |
| CN106127804B (en) | The method for tracking target of RGB-D data cross-module formula feature learnings based on sparse depth denoising self-encoding encoder | |
| US20220019805A1 (en) | Video watermark identification method and apparatus, device, and storage medium | |
| Jena et al. | Augmenting gail with bc for sample efficient imitation learning | |
| Civicioglu | Using uncorrupted neighborhoods of the pixels for impulsive noise suppression with ANFIS | |
| CN110047096B (en) | A kind of multi-object tracking method and system based on depth conditions random field models | |
| CN113793359B (en) | Target tracking method integrating twin network and related filtering | |
| CN114861838B (en) | Intelligent classification method for pulsatile neural brains based on neuron complex dynamics | |
| CN107945210A (en) | Target tracking algorism based on deep learning and environment self-adaption | |
| Wiggers et al. | Predictive sampling with forecasting autoregressive models | |
| CN115731396A (en) | Continuous learning method based on Bayesian variation inference | |
| CN118295029B (en) | Seismic data denoising method integrating self-attention and Mamba architecture | |
| CN112766339A (en) | Trajectory recognition model training method and trajectory recognition method | |
| Subudhi et al. | Kernelized fuzzy modal variation for local change detection from video scenes | |
| EP4392935A1 (en) | Robustifying nerf model novel view synthesis to sparse data | |
| CN112837342B (en) | Target tracking method, terminal equipment and storage medium | |
| Huang | Auto-attentional mechanism in multi-domain convolutional neural networks for improving object tracking | |
| Kalirajan et al. | Deep learning for moving object detection and tracking | |
| Lu et al. | Cdvae: Co-embedding deep variational auto encoder for conditional variational generation | |
| CN113256685B (en) | Target tracking method and system based on convolutional neural network dictionary pair learning | |
| Mehrkanoon et al. | Learning from partially labeled data. | |
| Klaiber et al. | A Systematic Literature Review on SOTA Machine learning-supported Computer Vision Approaches to Image Enhancement |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |