CN105654509A - Motion tracking method based on composite deep neural network - Google Patents

Abstract

The invention discloses a motion tracking method based on a composite deep neutral network. The motion tracking method comprises the steps of performing offline training on a weight by means of a large number of samples; using a neural network of which the number of nodes reduce gradually; adding a logistic classifier at the top of the network which is obtained through offline training so that a target and a background can be distinguished, and setting the parameter of an adjusting network; adjusting an observation model in a tracking process; at the first frame, performing adaptive adjustment on the first frame, adjusting the network by means of the target and a background sample so that the network can identify the target; combining the observation model with a dynamic model in which a particle filtering algorithm is used; acquiring particles around the target of a previous frame in a new frame in particle filtering, transmitting the acquired particles to the observation model, determining the confidence of the particles by the observation model, and determining the particle with highest confidence as the target.

Description

Based on the motion tracking method of compound deep neural network

Technical field

The present invention relates to the motion tracking method of a kind of object, in particular to a kind of motion tracking method based on compound deep neural network, it is a kind of tracking method utilizing deep neural network detection target.

Background technology

Motion tracking owing to being blocked, intensity of illumination change, rotations, face inner rotary, background clutter etc. affect and are considered as the task that has challenge in face, are also the important component parts of computer vision. Motion tracking can be widely used and multiple field, such as video monitoring, intelligent transportation, Product checking, unusual checking. It is suggested despite a large amount of model, but several crucial problems are not still solved fully.

A motion tracking system is generally made up of two models: observation model and dynamicmodel. Observation model is the model describing target object. Dynamicmodel is used for determining the conversion of object state and state. In tracking system, observation model is very important, and model accuracy is on tracking effect impact greatly. Accurate observation model can promote follows the tracks of success ratio (successrate), reduces central position error (center-of-locationerror).

Degree of depth study tracker (deeplearningtracker) uses deep neural network as observation model in the track. Owing to using complete vector base in deep neural network, the calculated amount in regulate process is huge. Therefore the speed of degree of depth study tracker often is difficult to accept. In tracking process, degree of depth study tracker uses BP algorithm to regulate. Due to excessively dark Web vector graphic BP algorithm training can produce diffusion problem (diffusion is when with BP algorithm regulating networks, loss partially leading of function significantly reduce along with backpropagation. When the network number of plies is very slow compared with the network weight of Shen Shi lower floor adjustment speed), the weights of network lower floor often can not get sufficient adjustment.

Summary of the invention

The present invention is directed to the deficiencies in the prior art, it is proposed to a kind of motion tracking method based on compound deep neural network, the method can reduce calculated amount, accelerates tracking speed, ensures to follow the tracks of success ratio, reduces central position error.Thus make motion tracking more accurate, more fast.

In order to solve the technical problem of above-mentioned existence, the present invention is achieved by the following technical solutions: a kind of motion tracking method based on compound deep neural network, and the method content comprises and comprising the following steps:

Step 1 off-line training:

Owing to different objects followed the tracks of by needs, need, before using observation model, the weights that different target can be adapted to fast, it is thus desirable to use great amount of samples that its weights are carried out off-line training; In training and tracking process, a large amount of calculated amount is consumed, so removing the structure that complete base vector carrys out simplified network, it may also be useful to the neural network that node quantity is successively decreased owing to crossing complete base vector;

Step 2 model initialization:

Logistic sorter is added so that it is object and background can be distinguished, and the parameter of regulating networks is set with off-line training gained network top;

Step 3 Automatic adjusument:

Often causing following the tracks of when object outward appearance changes in tracking process and drift about, therefore, observation model needs to regulate in tracking process; When the first frame, the first frame is carried out Automatic adjusument, with target and background sample, network is regulated, enable network identify target;

Step 4 target tracking:

Observation model and dynamicmodel being combined, dynamicmodel uses particle filter algorithm; Particle filter gathers particle in a new frame centered by previous frame target, gives observation model by the particle collected, and observation model judges the degree of confidence of particle, and the maximum particle of degree of confidence is target.

Wherein, in step 3, described first frame being carried out Automatic adjusument, its regulate process step is as follows:

1) input: the positive negative sample of collection and label S={ (x₁,y₁),......,(x_n,y_n), wherein y_i={ 0,1} is corresponding positive sample and negative sample respectively;

2) parameter setting: the minimum threshold alpha of reconstructed error of make an uproar automatic coding and BP algorithm falls in setting_DAnd ��_B, maximum training number of times ��_DAnd ��_B, study rate ��_DAnd ��_B, momentum m_DAnd m_BWith weight penalty coefficient w_DAnd w_B, the Noise figure �� of automatic coding of making an uproar falls;

3) compound training: perform DCTtrain (NN, X, Y, opts) process, it may also be useful to fall make an uproar automatic coding training lowest layer network and the network using the training of BP algorithm more high-rise; As reconstructed error E < �� or training number of times T >=�� time, training terminates;

Automatic adjusument process in tracking process only performs 3).

The present invention and prior art:

Owing to adopting technique scheme, a kind of motion tracking method based on compound deep neural network provided by the invention, compared with prior art has such advantage:

(1) two kinds of network combined methods are used to eliminate the diffusion problem in network self-adapting to a certain extent so that bottom-layer network obtains more effective training.

(2) simplified network structure, it may also be useful to the neural network that node quantity is successively decreased replaced complete base vector so that network training and tracking speed are accelerated greatly. The application of composite network can extract more effective feature quasi-complement removed complete base vector after the problem that affected of precision so that the present invention has higher real-time and robustness.

Accompanying drawing explanation

Fig. 1 is the schema using the motion tracking method based on compound deep neural network of the present invention;

Fig. 2 is Automatic adjusument schema;

Fig. 3 is compound training schema in Automatic adjusument process;

Automatic coding training schema of making an uproar falls in Fig. 4 in Automatic adjusument process;

Fig. 5 is BP algorithm flow figure in Automatic adjusument process.

Embodiment

The inventive method is further described by the motion tracking process below in conjunction with video sequence " woman ":

The present invention uses the motion tracking method based on compound deep neural network, and as shown in Figure 1, its particular content comprises following step:

Step 1 off-line training:

Owing to different objects followed the tracks of by needs, need, before using observation model, the weights that different target can be adapted to fast, it is thus desirable to use great amount of samples that its weights are carried out off-line training; In training and tracking process, a large amount of calculated amount is consumed, so removing the structure that complete base vector carrys out simplified network, it may also be useful to the neural network that node quantity is successively decreased owing to crossing complete base vector;

Step 2 model initialization:

Logistic sorter is added so that it is object and background can be distinguished, and the parameter of regulating networks is set with off-line training gained network top;

Step 3 Automatic adjusument:

Often causing following the tracks of when object outward appearance changes in tracking process and drift about, therefore, observation model needs to regulate in tracking process; When the first frame, the first frame is carried out Automatic adjusument, with target and background sample, network is regulated, enable network identify target;

Step 4 target tracking:

Observation model and dynamicmodel being combined, dynamicmodel uses particle filter algorithm; Particle filter gathers particle in a new frame centered by previous frame target, gives observation model by the particle collected, and observation model judges the degree of confidence of particle, and the maximum particle of degree of confidence is target.

Wherein, in step 3, described first frame being carried out Automatic adjusument, as shown in Figure 2, its regulate process step is as follows for its schema:

1) input: the positive negative sample of collection and label S={ (x₁,y₁),......,(x_n,y_n), wherein y_i={ 0,1} is corresponding positive sample and negative sample respectively;

2) parameter setting: the minimum threshold alpha of reconstructed error of make an uproar automatic coding and BP algorithm falls in setting_DAnd ��_B, maximum training number of times ��_DAnd ��_B, study rate ��_DAnd ��_B, momentum m_DAnd m_BWith weight penalty coefficient w_DAnd w_B, the Noise figure �� of automatic coding of making an uproar falls;

3) compound training: perform DCTtrain (NN, X, Y, opts) process, namely uses and falls make an uproar automatic coding training lowest layer network and the network using the training of BP algorithm more high-rise; As reconstructed error E < �� or training number of times T >=�� time, training terminates;

Automatic adjusument process in tracking process only performs 3).

Above step 1) step 3) described in be Automatic adjusument training overall process. In step 3) compound train in, described execution DCTtrain (NN, X, Y, opts) process, will illustrate that the concrete implementation step of DCTtrain (NN, X, Y, opts) is as follows here in detail, as shown in Figure 3:

(1) input: input variable is DCTtrain (NN, X, Y, opts), shown in, wherein NN represents the neural network after off-line training, and X represents target and background sample set, Y represents the label of sample, its order and sample one_to_one corresponding, and opts represents the parameter of setting;

(2) in lowest layer use, automatic coding training of making an uproar is fallen, as shown in Figure 4:

1. parameter setting: network parameter is arranged by the parameter in opts

2. network training:

A. according to Noise figure contaminated sample

B. reconstruct image z is obtained by contaminated sample by network forward direction:

$y=f_{\mathrm{\&theta;}}\left(\tilde{x}\right)$ Z=g_��'(y)

And obtain reconstructed error:

$\underset{W,W^{\&prime;},b,b^{\&prime;}}{min}\underset{i=1}{\overset{k}{\&Sigma;}}\left|\right|x_i-z_i|{|}_2^2+\mathrm{\&lambda;}\left(\right|\left|W\right|{|}_F^2+\left|\right|W^{\&prime;}\left|{|}_F^2\right);$

C. weights are upgraded by reconstructed error through backpropagation;

D. above process is repeated until reconstructed error is less than threshold alpha_DOr reach maximum training number of times ��_D;

E. by network, sample being obtained the first hidden layer and export a, it is using as more high-rise input;

(3) in the training of more high-rise use BP algorithm, as shown in Figure 5:

1. parameter setting: network parameter is arranged by the parameter in opts;

2. network training: use the training of BP algorithm;

3. repetition training process is until reconstructed error is less than threshold alpha_BOr reach maximum training number of times ��_B��

The content not being described in detail in specification sheets of the present invention belongs to the known prior art of professional and technical personnel in the field.

Although disclosing most preferred embodiment and the accompanying drawing of the present invention for the purpose of illustration, but it will be appreciated by those skilled in the art that: without departing from the spirit and scope of the invention and the appended claims, various replacement, change and amendment are all possible. Therefore the present invention should not be limited to the content disclosed in most preferred embodiment and accompanying drawing.

Claims (3)

1. the motion tracking method based on compound deep neural network, it is characterised in that: the method content comprises and comprising the following steps:

Step 1 off-line training:

Owing to different objects followed the tracks of by needs, need, before using observation model, the weights that different target can be adapted to fast, it is thus desirable to use great amount of samples that its weights are carried out off-line training; In training and tracking process, a large amount of calculated amount is consumed, so removing the structure that complete base vector carrys out simplified network, it may also be useful to the neural network that node quantity is successively decreased owing to crossing complete base vector;

Step 2 model initialization:

Logistic sorter is added so that it is object and background can be distinguished, and the parameter of regulating networks is set with off-line training gained network top;

Step 3 Automatic adjusument:

Often causing following the tracks of when object outward appearance changes in tracking process and drift about, therefore, observation model needs to regulate in tracking process; When the first frame, the first frame is carried out Automatic adjusument, with target and background sample, network is regulated, enable network identify target;

Step 4 target tracking:

Observation model and dynamicmodel being combined, dynamicmodel uses particle filter algorithm; Particle filter gathers particle in a new frame centered by previous frame target, gives observation model by the particle collected, and observation model judges the degree of confidence of particle, and the maximum particle of degree of confidence is target.

2. a kind of motion tracking method based on compound deep neural network according to claim 1, it is characterised in that: in step 3, described first frame being carried out Automatic adjusument, its regulate process step is as follows:

1) input: the positive negative sample of collection and label S={ (x₁,y₁),......,(x_n,y_n), wherein y_i={ 0,1} is corresponding positive sample and negative sample respectively;

2) parameter setting: the minimum threshold alpha of reconstructed error of make an uproar automatic coding and BP algorithm falls in setting_DAnd ��_B, maximum training number of times ��_DAnd ��_B, study rate ��_DAnd ��_B, momentum m_DAnd m_BWith weight penalty coefficient w_DAnd w_B, the Noise figure �� of automatic coding of making an uproar falls;

3) compound training: perform DCTtrain (NN, X, Y, opts) process, it may also be useful to fall make an uproar automatic coding training lowest layer network and the network using the training of BP algorithm more high-rise; As reconstructed error E < �� or training number of times T >=�� time, training terminates;

Automatic adjusument process in tracking process only performs 3).

3. a kind of motion tracking method based on compound deep neural network according to claim 1, it is characterized in that: in step 3) compound train in, described execution DCTtrain (NN, X, Y, opts) process, DCTtrain (NN will be described here in detail, X, Y, opts) concrete implementation step as follows:

(1) input: input variable is DCTtrain (NN, X, Y, opts), shown in, wherein NN represents the neural network after off-line training, and X represents target and background sample set, Y represents the label of sample, its order and sample one_to_one corresponding, and opts represents the parameter of setting;

(2) in lowest layer use, automatic coding training of making an uproar is fallen:

1. parameter setting: network parameter is arranged by the parameter in opts

2. network training:

A. according to Noise figure contaminated sample x��x��;

B. reconstruct image z is obtained by contaminated sample by network forward direction:

$\begin{array}{cc}y=f_{\mathrm{\&theta;}}\left(\tilde{x}\right)& z=g_{{\mathrm{\&theta;}}^{\&prime;}}\left(y\right)\end{array}$

And obtain reconstructed error:

$\underset{W,W^{\&prime;},b,b^{\&prime;}}{min}\underset{i=1}{\overset{k}{\&Sigma;}}\left|\right|x_i-z_i|{|}_2^2+\mathrm{\&lambda;}\left(\right|\left|W\right|{|}_F^2+\left|\right|W^{\&prime;}\left|{|}_F^2\right);$

C. weights are upgraded by reconstructed error through backpropagation;

D. above process is repeated until reconstructed error is less than threshold alpha_DOr reach maximum training number of times ��_D;

E. by network, sample being obtained the first hidden layer and export a, it is using as more high-rise input;

(3) train at more high-rise use BP algorithm:

1. parameter setting: network parameter is arranged by the parameter in opts;

2. network training: use the training of BP algorithm;

3. repetition training process is until reconstructed error is less than threshold alpha_BOr reach maximum training number of times ��_B��