CN111881030B

CN111881030B - Intelligent traffic data test sample generation method based on understandable characteristic variation

Info

Publication number: CN111881030B
Application number: CN202010714316.2A
Authority: CN
Inventors: 王晓冰; 夏志龙; 房春荣
Original assignee: Shenzhen Muzhi Technology Co ltd
Current assignee: Shenzhen Muzhi Technology Co ltd
Priority date: 2020-07-23
Filing date: 2020-07-23
Publication date: 2022-01-28
Anticipated expiration: 2040-07-23
Also published as: CN111881030A

Abstract

The invention provides an intelligent traffic data test sample generation method based on understandable characteristic variation, which mainly comprises two main steps of V2X scene variation rule generation and variation model construction, wherein the two main steps can be specifically divided into intelligent traffic V2X scene data acquisition, semantic variation rule construction, variation model construction and semantic variation realization. By utilizing the method, intelligent traffic V2X test units and individuals can efficiently solve the problem of insufficient semantic diversity of traffic data of intelligent traffic V2X standard application scene test, semantic variation test samples are expanded, and the software reliability of an intelligent traffic system is finally improved.

Description

Intelligent traffic data test sample generation method based on understandable characteristic variation

Technical Field

The invention belongs to the technical field of intelligent software testing, and particularly relates to a test sample generation technical framework based on externally understandable semantic feature variation, which aims at the expansion of a test sample of an external input state space of a complex software system represented by a vehicle system in an intelligent traffic V2X test scene, and maintains the software reliability of an intelligent traffic V2X system by constructing a large number of high-scene feature coverage test cases.

Background

The intelligent transportation V2X software has many differences from the traditional software in the system construction functional logic implementation, the internal algorithm model which is relied on by the intelligent implementation, such as machine learning, intelligent communication algorithm or model, has no boundary concept in the general sense, and in addition, the large black box characteristics exist in the actual operation, and the result is often difficult to detect or carry out scientific interpretation. Therefore, the relevant data-driven test means surrounding the external data of the intelligent traffic V2X software system is particularly effective.

In the intelligent transportation V2X software system, data is a core element for ensuring the normal operation of the system. There have been several studies around testing, and there are some main ideas of the following researches, one is to introduce the related concepts of traditional software testing such as branching/covering, and the other is to explore new covering standards and testing means from the internal structure and logic of intelligent software. In data sample research, there have been many theoretical and technical researches oriented to traffic environment data expansion, which are used for providing large-scale test data for simulating a driving system of V2X. However, these data augmentation methods are often based on fixed internal abstract logic, such as some random noise addition, and lack in-depth research into traffic scene semantic features, i.e., augmenting data around intelligible semantic features in augmentation, such as weather elements in traffic environments in a V2X scene. Such studies have their necessity, but are still in an early stage at present.

Most of intelligent V2X software has independent traffic semantic application scenes. The concept of semantic features is different from the logical features in the data, and means that the system has understandable and expressible features outside, such as ambient brightness, climate condition style and data density. Semantic mutation is a new technical concept for applying the transformation rule of understandable characteristics to actual data. Typical examples are V2X perception related function modules, camera data testing requires samples covering various common semantic features including weather, illumination, visibility and shooting angle, and corresponding test sample state space coverage is difficult to achieve in a short time by means of random or other mathematical and data science data amplification methods. For semantic rules which can be understood by human beings, how to capture the corresponding transformation relation in high-dimensional image data and how to fuse in a rule acquisition part by using some existing technologies are also hot topics in the industry. Currently, most of the technical fusion is still in an auxiliary research or exploration stage, and a complete technical support and a complete technical framework are lacked for semantic testing.

Disclosure of Invention

The invention aims to provide an intelligent traffic data test sample generation method based on understandable characteristic variation, which integrally comprises two main steps of V2X scene variation rule generation and variation model construction. By using the invention, intelligent traffic V2X test units and individuals can efficiently solve the following important problems:

(1) the V2X standard application scene tests the traffic data semantic diversity is insufficient, and the technical support problem of semantic variation test samples is expanded;

(2) the method comprises the following steps of (1) obtaining and mathematizing semantic features of traffic data in a specific test scene;

(3) quality measurement and analysis problems of data semantic diversity samples.

In order to realize the technology of the invention, the method mainly comprises the following main steps:

1V 2X scene semantic rule extraction. In order to obtain a test sample based on V2X scene semantic feature variation, original data samples are analyzed and expanded by combining crowdsourcing and expert knowledge, semantic features and variation rules under an intelligent traffic application scene are extracted, and the extracted semantic features and the variation rules are converted into a data transformation method which can be realized by mathematics or a computer, so that data operation and use in practical application are facilitated.

Comprises the following 2 sub-steps:

1.1 raw data acquisition. And collecting an initial sample data set, wherein the original data mainly come from different V2X standard scenes, and according to an opening source data set of a specific field under the scene, an intelligent transportation terminal enterprise unit simulation test data set, an actual V2X test sample and network resources.

1.2 mutation rule extraction. Variant data expansion and variant rules the data in the original dataset is too cluttered and therefore it is first necessary to improve the data quality, including data evaluation and expansion. And then labeling different semantic features and variant data. And finally, constructing a transformation relation, acquiring formal expression of mathematics or data science under different semantic feature variations, and constructing a data transformation method executable by the specific traffic software data. The invention introduces a plurality of processing strategies to carry out multi-dimensional processing. And according to the data scale, reasonable combination of two strategies of expert guidance, tool direct processing and crowdsourcing processing is used.

Aiming at the small-scale data application scenarios of the intelligent traffic V2X, such as the test of a specific sensor functional module, the invention introduces expert knowledge, open source tools or related tools in the professional fieldOther software performs data quality control, mutation type analysis, and final mutation rule construction. And for large-scale application data scenes, a crowdsourcing technology is utilized, two forms of an expert tool and a crowdsourcing task are utilized to allocate to complete corresponding tasks, according to the professional degree of the data semantic features, the tasks are reasonably distributed on the crowdsourcing task according to the difficulty of marking specific scene features, and the proportion of participation of experts and crowdsourcing is controlled. Through continuous iteration of the expansion, labeling and extraction of variation rules of the variation data, the sub-step 1.2 finally obtains a processed sample set

Obtaining a sample variation rule set corresponding to a specific semantic scene

The size of n depends on the thickness of the granularity of the semantic features used.

1.3 semantic rule datamation. The obtained variation relation and variation rule are digitalized by professional knowledge or professional tools

；

Different data relationships, using different data representation forms. Suppose that

For the luminance transformation of an image, the data transformation represented on the picture matrix can then be expressed as the matrix pixel values multiplied by a given constant β, according to experience in the field relevant to image processing.

For vehicle radar sensor data, the related data transformation comprises point cloud data point value distribution adjustment, and typical data is reflected by a radar angle and a three-dimensional tracking construction modeAnd measuring the road state of the object entity. If it is

The blurring effect of weather, which is an image, can be simply expressed as pixel block convolution in a picture matrix, and the basic range of a matrix operator can be obtained by using PS (picture) display pixel values and by means of a mathematical method. If it is

In a complex data transformation form, r cannot be simply represented by an expression T in order to construct a proper semantic transformation rule.

2 constructing a variation model. Integrating transformation rules into a sample generator

In (1). And constructing a model construction for realizing data variation by utilizing a generation countermeasure network or a fuzzy method based on the mutation amplification data set constructed in the steps. The GAN and the Fuzz are respectively suitable for different variation scenes, the variant data efficiency of the GAN and the Fuzz is higher under a specific variation scene, and the variant data efficiency of the GAN and the Fuzz can be more widely used for constructing different types of test samples. For the common multi-feature multi-label problem of complex latitude data in the intelligent traffic V2X scene, the variation type is further expanded by using a method of synthesizing a plurality of variation modes, and a control function and a constraint rule related to variation are introduced for guiding and controlling the data quality before and after variation and the scale of a related data set.

Further, the specific operation is divided into 2 sub-steps:

2.1 original sample and transformation sample dataset construction. By pairs

By iterative expansion, new data sets are constructed from the original data and the variant samples, respectively

From the original seed data

And variant samples

And (4) forming.

2.2 constructing a variation model. The application of a plurality of modules enables simple variation means to directly integrate single variation in the model without using complex model generation such as light and dark, contrast, angle adjustment and simple tone adjustment of audio, and complex modules such as related road image background style, related characteristics of the whole traffic scene and V2X audio signals utilize relatively complex variation technology such as utilizing a style conversion neural network to realize whole style conversion, utilizing a GAN technology to realize specific characteristic disturbance and utilizing Glow to realize traffic pedestrian behavior characteristic conversion. In addition, the intelligent traffic V2X has multiple variations combined with variations to meet multiple environment variation characteristics in the application scene of the intelligent traffic V2X. In order to control the variation effect not to deviate too far from the display and simultaneously solve the contradiction problem of excessive combination in an ultra-large-scale state space, a metric function of threshold control is introduced for optimizing and balancing the variation combination degree and the sample number:

ffor measuring the degree of variation between a particular two samples,

the difference degree between the original sample data and the variation sample data under a certain variation rule is described;

Ea correlation function for the desired number of generated variant test samples,

it is used to indicate the total number of samples for which a mutation is made,

then represents the sample set under the combined variation, the parameter in the threshold expression

The larger the value of (a) is used for balancing the requirements, the more importance is placed on the generation scale, and the smaller the value is, the more importance is placed on the variation granularity of the transformation sample.

As a bias control item, a reasonable demand offset is set for the optimization function under a specific application scene.

The threshold value is controlled to control the overall generation scale. An appropriate sample size and weight parameter lambda is found during the optimization process.

The invention is characterized in that:

technically, a test method of external semantic features of an intelligent traffic scene in an intelligent V2X software test is provided, and the problems that the traditional data amplification lacks human comprehensible characteristics and the traditional variation method is high in scene coverage cost are solved. Meanwhile, a variation combination generation technology is introduced to improve the coverage rate of the complex scene.

The method is oriented to the standard application scene of the intelligent traffic V2X, has data driving characteristics, and meets the high standard data quality requirements of software and hardware of intelligent traffic perception and decision-making systems in the security critical field.

And a crowdsourcing technology is introduced into data set processing, data labeling and rule extraction, so that the research cost is reduced, the high cost of expert labeling is reduced, the processing speed is increased, and the overall efficiency of the process is finally improved.

Drawings

Fig. 1 is a general framework diagram of the present invention.

FIG. 2 is a flow chart of a semantic feature variation generator.

Fig. 3 is a schematic diagram of transformation rule extraction.

Detailed Description

The embodiments of the present invention are described below with reference to specific examples, and other advantages and effects of the present invention will be readily apparent to those skilled in the art from the disclosure of the present specification.

As shown in fig. 1, the overall schematic diagram of the present invention includes: contains 2 main parts and is continuously optimized based on an iteration flow:

1, extracting external semantic features and transformation relations of the system. In order to capture the transformation rules in a specific intelligent transportation application scenario, related raw data needs to be obtained in the specific intelligent transportation software field for dividing the most basic transformation rules. Such as for a V2X road weather environment, a large amount of image or other sensor data containing weather elements is required. Then, for different transformation rules, it is necessary to mathematically transform them from human comprehensible representational meanings, and extract mathematical expressions which can be implemented by a computer at a data level, such as some specific algebraic conversion forms. In the experiment, the initial sample rule is captured, the sample is expanded by adopting a crowdsourcing technology through experts in the field of intelligent traffic V2X or a specific traffic intelligent terminal development unit, and the rule collection and the feedback induction are integrally expressed as an iterative process for the quality and the scale of the final sample set.

The 2-transformation relation system is internally integrated. Set of categorically distinct data obtained in 1)

Classifying into original data set and variant data sample

、

. Different transformation modules are used in the model for the variation rules of different complexity levels. First consider a number combining mathematics and specialized fieldsThe data sample generation technology adopts algebraic matrix transformation on vectors on the semantic amplification transformation of camera image data, and the generation technology adopts a generation network technology and utilizes distribution functions on the data distribution intensity. For sensor data represented by lattice data, a corresponding transformation mode including a sensor simulation construction model and a mathematical modeling tool is fitted. Under the complex semantic transformation, for the detachable variation rule, a plurality of transformation modes can be used for expanding the diversity of the sample, and a condition discrimination technology for optimizing and controlling the sample scale and the variation degree is introduced for evaluation. And constructing a corresponding data set by the mutation data, and simultaneously using the mutation data to participate in the feedback in the step 1 so as to optimize the semantic mutation rule.

For ease of understanding, the method and internal logic are described in detail using a specific example of intelligent traffic-aware scene sample mutation amplification:

1) and (3) perception data acquisition, wherein the data source comprises an intelligent traffic public data set, real historical data of an intelligent perception terminal and simulation data. Including raw samples with partial variation rules

；

2) Corresponding variation classification, supplement and rule extraction, constructing semantic diversity description and corresponding variation rules in various modes, selecting crowd-sourcing to assign variation sample labeling tasks, and summarizing to form complete and reasonable intelligent traffic vertical domain semantic variation rules after feedback and audit are completed;

3) for general perception data variation types, data rule extraction and mathematics are assisted by intelligent transportation experts (direct/crowdsourcing) or MatLab tools, and general picture variation rules comprise perceived image data brightness, contrast, rain and fog effects, traffic picture shooting angles and picture defiling. For some non-visual perception data, a transformation method is fitted by means of model analysis. For complex transformations, implemented internally in 5) modules, a repository of semantic rule transformations available is created;

4) constructing and integrating the simple transformation proposed by the step 3) into a corresponding variation model. Complex transformation, including data variation related to various semantics, or scene data similar to the overall color tone style of a road, which cannot be directly transformed by mathematics, utilizes style conversion or other effective data generation technologies to complete the related perceptual data variation by means of deep neural network learning ability;

5) adding synthesis transformation, synthesizing by using multiple transformation types to further expand a variation sample scene, and evaluating the balance of control scale and variation degree by using a threshold value and a joint optimization inequality;

6) module integration to form complete variation generator G (R);

7) and (3) actually applying a variation generator to complete the semantic feature variation of the test sample.

Claims

1. An intelligent traffic data test sample generation method based on comprehensible feature variation is characterized in that scene semantic rules of intelligent traffic V2X are extracted, in order to obtain a test sample based on V2X scene semantic feature variation, original data samples are analyzed and expanded in combination with crowdsourcing and expert knowledge, semantic features and variation rules under an intelligent traffic V2X standard application scene are extracted and converted into a data conversion method which can be realized by mathematics or a computer, data operation and use in practical application are facilitated, a variation model is constructed, and conversion rules are integrated into a sample generator

Based on the mutation amplification data set constructed in the above steps, constructing a model construction for realizing data mutation by using a generation countermeasure network or a fuzzy method; digitizing semantic rules, and digitizing the obtained variation relationships and variation rules with the help of professional knowledge or professional tools

，

Different data relationships, using different data representation forms, by comparing the original samples

The new data set is respectively constructed by the original data and the variation sample

From the original seed data

And variant samples

The use of multiple modules, in combination, allows simple mutation strategies to be generated without the need to use complex models, such as light and dark, contrast, angle adjustment, simple pitch adjustment of the audio, direct integration of such single variations within the model, and complex modules, such as the background style of the road image, the relevant characteristics of the whole traffic scene, the audio signal of V2X, then relatively complex variation techniques such as global style transformation using a style transformation neural network, specific feature perturbation using GAN techniques, traffic pedestrian behavior feature transformation using Glow, in addition, the method has a plurality of variations combined with variations to meet a plurality of environment variation characteristics in the application scene of the intelligent traffic V2X, in order to control the variation effect not to deviate too far from the display, meanwhile, the contradiction problem of excessive combination under an ultra-large-scale state space is also solved, and a measurement function controlled by a threshold value is introduced to optimize and balance the variation combination degree and the sample number:

ffor measuring the degree of variation between a particular two samples,

the difference degree between the original sample data and the variation sample data under a certain variation rule is described,

The larger the value of (A) is used for balancing the requirement, the more important the generation scale is, the smaller the value is, the more important the transformation sample variation granularity is,

as a bias control item, a reasonable demand offset is set for the optimization function under a specific application scene,

in order to control the threshold value, the generation scale of the whole is controlled, and a proper sample scale and a weight parameter lambda are searched in the optimization process.

2. The method for generating the intelligent traffic data test sample based on the understandable characteristic variation as claimed in claim 1, wherein an initial sample data set and semantic feature extraction are collected, the initial data mainly comes from different V2X standard scenes, and according to a source data set of a specific field under the scene, an intelligent traffic terminal enterprise unit simulation test data set, an actual V2X test sample and network resources, data quality needs to be improved firstly, including data evaluation and expansion, then different semantic features and variation data are labeled, and finally, a transformation relation is constructed, formal expressions of mathematics or data science under different semantic feature variations are obtained, so that the method is used for constructing a data transformation method executable by specific traffic software data.

3. The method for generating intelligent transportation data test samples based on understandable characteristic variation as claimed in claim 1, wherein a method for constructing variation rules aims at small-scale data application scenarios of intelligent transportation V2X, such as specific sensor function module tests, data quality control, variation type analysis, and final variation rule construction are performed by introducing expert knowledge in professional fields, sourcing tools, or related other software, and for large-scale data application scenarios, corresponding tasks are completed by using a crowdsourcing technique and using an expert tool + crowdsourcing tasks to assign two forms, and according to the professional degree of data semantic characteristics, tasks are reasonably assigned according to the difficulty of specific scene characteristic labeling in a crowdsourcing process, and the balance between experts and crowdsourcing is grasped.