CN108596146A - Road multi-target classification method - Google Patents

Abstract

The invention discloses a road multi-objective classification method, which learns a mixed Bayesian network structure by importing training samples and applying a constraint-based NPC algorithm; then performs parameter learning on discrete variables and continuous variables in the network structure to obtain a network The distribution of each node in the network, and then the parameters are combined, and finally the test sample is used for the inference of the Bayesian network and the classification of the road target. On the one hand, this method abandons the need for high-resolution and close-range images, and greatly reduces the amount of calculation by using simple low-level features of road targets. On the other hand, the construction of the hybrid Bayesian network structure avoids all variables in the traditional Bayesian network classifier as discrete variables, which is easy to cause the loss of target information, and at the same time leads to a search space in the processing and analysis of road multi-target data. and a sharp increase in the amount of computation. The present invention adopts a Bayesian network in which continuous nodes and discrete nodes coexist, which is more practical.

Description

Road multi-objective classification method

technical field

The invention relates to an intelligent video monitoring technology, in particular to a road multi-target classification method.

Background technique

The research on smart cars is in the ascendant. Environmental recognition is the basic module of smart cars and the prerequisite for autonomous driving of smart cars. A basic task of environment recognition is the detection and identification of obstacles. Existing research focuses on the detection of a single type of obstacle, but less research on the detection and recognition of multiple types of obstacles. For a smart car driving in an urban road environment, it is not only necessary to detect the targets ahead, but also to identify their types.

The current general classifiers include decision tree, neural network, support vector machine, Adaboost, etc. The classification decision-making basis comes from the learning of sample data, so it needs to be supported by a large number of class samples. Considering the diversity of target types, the complexity of the environment and the uncertainty of the shape, the classification rules obtained only by learning from samples are greatly affected by the sample size and spatial distribution. The Bayesian network has a strong ability to express uncertain knowledge. In the process of classification decision-making, it can make full use of prior knowledge and statistical learning information to make the inference rules more flexible and effective. Or an effective classifier can still be built without sample data. However, the traditional Bayesian network classifier regards all variables as discrete variables, but the discretization of variables will inevitably lead to information loss, and in the processing and analysis of road multi-objective data, the discretization of continuous variables Optimization will lead to a sharp increase in the search space and the amount of computation. Therefore, the road multi-objective classification problem is still a difficult problem in the intelligent vehicle system.

Contents of the invention

In order to solve the above problems, the present invention provides a road multi-objective classification method of a mixed Bayesian network with both discrete nodes and continuous nodes.

To achieve the above object, the present invention adopts the following technical solutions:

A road multi-objective classification method, the method is: applying the constraint-based NPC algorithm to the training samples to learn the mixed Bayesian network structure, and performing parameter learning on the discrete variables and continuous variables in the network structure to obtain each The distribution of nodes, the two types of parameters are combined, and finally the test samples are used for Bayesian network inference and road multi-object classification.

Specifically include the following steps:

1) Establish a data set, including a training data set for training the classification algorithm model and a test data set for testing the classification algorithm;

2) Set the classification category of the target to be identified as m categories, extract n classification features of the target, and divide the classification features into two categories: discrete variables and continuous variables;

A _mn is a two-dimensional matrix containing all category information for each feature:

Where a _i,j is the value of the jth feature of the i-th class target, and {A ₁ ,A ₂ ,…,A _n } are taken as the values of the corresponding nodes in the hybrid Bayesian network;

3) Import the training data set, apply the constraint-based structure learning method NPC algorithm to learn the structure of the hybrid Bayesian network; apply the Bayesian network parameter learning method to obtain the distribution P(X _i |C) of each node, Parameter learning is performed separately for discrete variables and continuous variables;

Combine the two types of parameters obtained above to obtain a hybrid Bayesian network;

4) Import the test data set into the obtained hybrid Bayesian network to classify the urban road targets.

In the above technical solution, further, the classification features described in step 2) are the occlusion, length (length), width (width), height (height) and observation angle (alpha) of the target. Among them, the occlusion of the target is a discrete variable, and the length, width, height and observation angle of the target are continuous variables.

Further, in the step 3), the relationship between the nodes is determined through dependency analysis, if there is a dependency relationship between the nodes, the undirected edge is retained, otherwise the edge between the nodes is removed; specifically, through chi-square statistics The framework of the Bayesian network is determined by doing hypothesis tests to determine the presence or absence of edges, and then the direction of edges is determined from the segmentation sets generated in independent tests; it can be used for any connections that seem counterintuitive or when resolving ambiguous regions. User interaction takes place, and the user can use this opportunity to determine the directionality of the undirected link and resolve its ambiguous regions.

Furthermore, for discrete variables, the conditional probability table CPT is represented by a multidimensional matrix, which is: E(θ _i,j,k |D,B _S ,ξ)

＝(N _i,j,k +1)(N _i,j +r _i -N _i,j,k -1)/(N _i,j +r _i ) ² (N _i,j +r _i +1 )

Among them, θ _{i , j, k represent} the conditional probability table; ξ represents several hypotheses; D is the data set; B _S is the network structure; _j represents the jth parent node of the variable X _i , v _i,k is the value of the variable X _i , then N _i,j,k is the value of the variable X _i in the data set D is v _i,k and the parent node is The number of occurrences of samples of ω _i,j , the calculation formula of N _i,j is

For each feature of a discrete variable, the result of parameter learning is a two-dimensional matrix of m×k size: M ₁ =CPT _m×k ;

For continuous variables, the conditional probability distribution is:

Among them, C represents the category label, and the continuous variable conforms to the univariate normal distribution, that is, X _i ~ N(μ _i , σ _i ), the two important parameters of the normal distribution, the mean value μ _i and the variance σ _i , are directly calculated from the training samples Obtained; when the sample data is incomplete, the EM algorithm can be used to solve the learning problem of the mixed Bayesian network parameters;

For each feature of a continuous variable, the result of parameter learning is a two-dimensional matrix M ₂ =CPD _m×2 =<μ _i , σ _i >, i∈{1,2,...,m}.

Merge the two types of parameters M ₁ and M ₂ to obtain the parameter network θ=<M ₁ , M ₂ >; where M ₁ is the two-dimensional matrix of the m×k conditional probability table CPT _m×k ; M ₂ is The conditional probability distribution CPD of size m×2 is an _m×2 two-dimensional matrix.

The beneficial effects of the present invention are:

According to the mixed Bayesian network-based road multi-objective classification method involved in the present invention, the structure of the mixed Bayesian network is learned through the NPC algorithm based on constraints, and then the parameters of the discrete variables and continuous variables are respectively learned to obtain each in the network. The distribution of a node, and then the parameters are combined, and finally the test sample is used in the network to classify the urban road target. On the one hand, it abandons the need for high-resolution and close-range images, and by using simple low-level features of road targets, such as height, width, and viewing angle, it greatly reduces the amount of calculation and can run in real time. On the other hand, the construction of the mixed Bayesian network structure avoids all variables being regarded as discrete variables in the traditional Bayesian network classifier, which will cause the loss of target information, and cause problems in the processing and analysis of road multi-target data. The search space and computational complexity increase dramatically. The Bayesian network in which continuous nodes and discrete nodes coexist is more realistic.

Description of drawings

Fig. 1 is a logical schematic diagram of an embodiment of the present invention;

FIG. 2 is a schematic diagram of a hybrid Bayesian network structure according to an embodiment of the present invention.

Detailed ways

In order to further understand the present invention, the preferred embodiments of the present invention are described below in conjunction with examples, but it should be understood that these descriptions are only to further illustrate the features and advantages of the present invention, rather than limiting the claims of the present invention.

The present embodiment provides a kind of road multi-target classification method, as shown in accompanying drawing 1, by importing training sample application based on constraint NPC algorithm, mixed Bayesian network structure is learned; Then discrete variable and continuous variable in network structure Variables are parameter learned separately to obtain the distribution of each node in the network, and then the parameters are combined. Finally, the test samples are used for Bayesian network inference and road targets are divided into eight categories. The specific implementation steps are as follows:

1) Use KITTI's 3D object detection benchmark dataset to test this classification method. The data set is all real road scenes, and the data collection scenes are rich. Eight types of obstacles are marked in each frame point cloud by professional annotators: Pedestrian, Car, Van, Truck, Cyclist, Person_sitting, Tram, and Misc. And there are various degrees of occlusion and truncation. After finishing, the whole data set is divided into two parts—one training data set and one testing data set. Among them, the training data set is used for the training of the algorithm model, including a total of about 20,000 target obstacles; the test data set is used for the test of the classification algorithm, including a total of about 20,570 target obstacles.

2) Two sets of schemes are designed for target identification, as shown in Table 1. Scheme 1 subdivides the target into 8 categories, namely pedestrians, sitting people, cars, trucks, vans, cyclists, trams and hybrid vehicles. Plan 2 roughly divides the target into three categories, that is, the sitting person in Plan 1 is merged into the pedestrian category, and the truck and truck are merged into the car category, and trams and hybrid vehicles are ignored.

Table 1 Identification target scheme

3) The experiment is carried out under the Hugin expert platform, and 20,000 road target information data in the training data set are selected as the training samples of the Bayesian classifier. Including 14035 cars (Car), 2263 pedestrians (Pedestrian), 1461 trucks (Van), 528 trucks (Truck), 834 cyclists (Cyclist), 117 sitting people (Person_sitting), 265 trams (Tram) and 497 mixed cars (Misc).

3) The structure of the Bayesian network is crucial to the accuracy of the model. Learning the optimal structure of a Bayesian network takes exponential time because the number of large possible structures for a given set of nodes is superexponential in the number of nodes. We use the constraint-based structural learning method NPC algorithm to learn the network structure, set the prior probability distribution of all variables to be randomly generated, and use the chi-square distribution to construct the statistics of the conditional independent test, and set the significance level to 0.05. The number of EM parameter learning iterations is set to 10, and the Bayesian Information Criterion (BIC) function is used as the scoring function. The generated hybrid Bayesian network structure is shown in Figure 2: it is a hybrid Bayesian network model composed of five layers and six nodes. Each node in the figure is described in detail as follows: Assume that the variable C is a discrete variable, and its values are all possible target types, represented by the root node. The target types to be identified are shown in Table 1, and the observed variables are lidar and visual sensors The characteristics of the observed moving target are represented by sub-nodes. They are discrete variable occlude (whether the target is occluded), continuous variables length (target length), width (target width), height (target height) and alpha (target observation angle, range: -π～π).

3) The steps of hybrid Bayesian network parameter learning are as follows:

(1) Obtain the feature subset X={X1, X2,...,Xn} of the road target from the training sample set D={D ₁ , D ₂ ,...,D _m };

(2) Extract target features: A _mn is a two-dimensional matrix containing all category information of each feature:

Where a _i,j is the value of the jth feature of the i-th type of target, and generally {A ₁ ,A ₂ ,…,A _n } is used as the value of the corresponding node in the network;

(3) Obtain the category labels C={C ₁ ,C ₂ ,...,C _m } of various samples;

(4) Confirm the variable type of each node, that is, divide X into two subsets of discrete node set X _discrete and continuous node set X _continual ;

(5) Discretize X _discrete by the method of discrete Bayesian network;

(6) Carry out parameter learning for discrete nodes and continuous nodes respectively. For each feature in X _discrete , the result of parameter learning is a two-dimensional matrix M ₁ =CPT _mk (conditional probability table) of size m×k. For each feature in X _continual , the result of parameter learning is a m×2 two-dimensional matrix M ₂ =CPD=<μ _i ,σ _i >,i∈{1,2,…,m} (conditional probability distributed).

(7) Combine parameters M ₁ and M ₂ to obtain a parameter network θ=<M ₁ , M ₂ >.

4) Statistically obtain the mean and variance of the eigenvalues of the continuous variables and the conditional probability table of the discrete variables in the training samples.

5) Input the test sample into the trained network structure to classify the test target.

The test data set is composed of 20,570 test targets, including 8 targets in Scheme 1 in Table 1, consisting of 14,707 cars, 2,224 pedestrians, 1,453 trucks, 566 trucks, 793 cyclists, and 105 sitting people, 246 trams and 476 mixed cars. The resulting confusion matrix is shown in Table 2, which shows how well the observed state matches the prediction with an error rate of 4.08%. The first line shows that 2213 of the 2224 pedestrian samples were correctly classified, 6 were misclassified as mixed vehicles, and 5 were misclassified as sitting people. The accuracy of the classification target is compared with other methods as shown in Table 3. It can be seen that this method has a good classification effect on pedestrians, cars, trams and cyclists, but the classification accuracy on vans, trucks and hybrid vehicles is relatively low. However, when this method is applied to the three targets in Scheme 2 of Table 1, the resulting confusion matrix is shown in Table 4, with an error rate of 0.08%, and all 16,727 samples of the car class are correctly classified. In this way, the classification features are obvious, so the classification accuracy is significantly improved.

Table 2 Confusion matrix of eight types of targets

Table 3 Comparison of accuracy rate (Precision) with other methods

For other methods involved in this table, please refer to the follow-up literature for details.

Table 4 Confusion matrix of three types of targets

[1] Shen Zhixi, Huang Xiyue, etc. Multi-type obstacle recognition of intelligent vehicles based on Boosting [J]. Computer Engineering. 2009,35(14):241-242

[2] Silviu Bota, Sergiu Nedevschi. Matthias Konig. A Framework for Object Detection, Tracking and Classification in Urban Traffic Scenarios Using Stereovision [J]. IEEE International Conference on Intelligent Computer Communication & Processing, 2009: 153-156.

[3]Chug-Wei Liang and Chia-Feng Juang. Moving Object Classification Using a Combination of Static Appearance Features and Spatial and Temporal Entropy Values of Optical Flows[J].IEEE TRANSACTIONS OF INTELLIGENTTRANSPORTATION SYSTEMS,2015,16(6):3453-3464 .

[4] Mehran Kafai and Bir Bhanu. Dynamic Bayesian Networks for Vehicle Classification in Video [C]. IEEE transactions on industrial informatics, 2012, VOL.8, NO.1:100-109.

The descriptions of the above embodiments are only used to help understand the method and core idea of the present invention. It should be pointed out that for those skilled in the art, without departing from the principle of the present invention, some improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims (7)

1. a kind of road Multi-Target Classification Method, which is characterized in that this method is：Training sample is applied into the NPC based on constraint Algorithm learns Hybrid Bayesian Network structure, in network structure discrete variable and continuous variable carry out parameter respectively Study obtains the distribution of each node in network, and two class parameters are merged, and test sample is finally used for Bayesian network The reasoning of network simultaneously classifies road multiple target.

2. road Multi-Target Classification Method according to claim 1, which is characterized in that specifically comprise the following steps：

1) data set is established, includes the training dataset for training sorting algorithm model and the survey for testing sorting algorithm Try data set；

2) target classification classification to be identified is set as m classes, extracts n characteristic of division of target, characteristic of division is divided into discrete Two class of variable and continuous variable；

A_mnIt is the two-dimensional matrix for including each feature all categories information：

Wherein a_i,jFor the value of i-th j-th of feature of class target, by { A₁,A₂,…,A_nBe used as in Hybrid Bayesian Network and accordingly save The value of point；

3) training dataset is imported, using the Structure learning method NPC algorithms based on constraint to the structure of Hybrid Bayesian Network Learnt；Distribution P (the X of each node are obtained using Bayesian network parameters learning method_i| C), for discrete variable Parameter learning is carried out respectively with continuous variable；

Two class parameters of above-mentioned acquisition are merged, Hybrid Bayesian Network is obtained；

4) test data set is imported in the Hybrid Bayesian Network obtained, is classified to urban road target.

3. road Multi-Target Classification Method according to claim 1, which is characterized in that point described in the step 2) Category feature is circumstance of occlusion (occlude), length (length), width (width), height (height) and the view angle of target Spend (alpha) five features, wherein the circumstance of occlusion of target is discrete variable, length, width, height and the view angle of target Degree is continuous variable.

4. road Multi-Target Classification Method according to claim 1, which is characterized in that by card side in the step 3) Statistic does hypothesis testing to determine the presence or absence on side, so that it is determined that the frame of Bayesian network, then further according to independent inspection The segmentation collection for testing middle generation determines the direction on side；It determines the directionality of undirected connection in addition, can be interacted by user and solves it Fuzzy region.

5. road Multi-Target Classification Method according to claim 1, which is characterized in that for discrete in the step 3) Variable, conditional probability table CPT is indicated using multi-dimensional matrix, is：

E(θ_i,j,k|D,B_S, ξ) and=(N_i,j,k+1)(N_i,j+r_i-N_i,j,k-1)/(N_i,j+r_i)²(N_i,j+r_i+1)

Wherein, θ_i,j,kIndicate conditional probability table；ξ indicates several hypothesis；D is data set；B_SFor network structure；r_iFor Discrete Stochastic Variable X_iAll possible value number, if using ω_i,jIndicate variable X_iJ-th of father node, v_i,kFor variable X_iValue, Then N_i,j,kFor variable X in data set D_iValue is v_i,kFather node is ω simultaneously_i,jSample occur number, N_i,jCalculating it is public Formula is

For each feature of discrete variable, parameter learning the result is that the two-dimensional matrix of m × k size：M₁= CPT_m×k；

6. road Multi-Target Classification Method according to claim 1, which is characterized in that for continuous in the step 3) Variable, conditional probability distribution are：

Wherein, C represents class label, and continuous variable meets unitary normal distribution, i.e. X_i~N (μ_i,σ_i), normal distribution it is equal Value μ_iAnd variances sigma_iTwo important parameters are directly calculated from training sample；When sample data is incomplete, EM can be passed through Algorithm solves the problem concerning study of Hybrid Bayesian Network parameter；

For each feature of continuous variable, parameter learning the result is that the two-dimensional matrix M of a size of m × 2₂=CPD_m×2=< μ_i, σ_i>,i∈{1,2,…,m}。

7. road Multi-Target Classification Method according to claim 1, which is characterized in that join to two classes in the step 3) Number M₁And M₂Merge, obtain parameter network θ=<M₁, M₂>；Wherein, M₁For the conditional probability table CPT of m × k sizes_m×kTwo Tie up matrix；M₂For the conditional probability distribution CPD of the sizes of m × 2_m×2Two-dimensional matrix.