WO2020151059A1 - 一种基于hmm-svm双层改进模型的复杂路况下周边车辆行为识别方法 - Google Patents

一种基于hmm-svm双层改进模型的复杂路况下周边车辆行为识别方法 Download PDF

Info

Publication number
WO2020151059A1
WO2020151059A1 PCT/CN2019/077584 CN2019077584W WO2020151059A1 WO 2020151059 A1 WO2020151059 A1 WO 2020151059A1 CN 2019077584 W CN2019077584 W CN 2019077584W WO 2020151059 A1 WO2020151059 A1 WO 2020151059A1
Authority
WO
WIPO (PCT)
Prior art keywords
hmm
svm
vehicle
model
layer
Prior art date
Application number
PCT/CN2019/077584
Other languages
English (en)
French (fr)
Inventor
蔡英凤
邰康盛
刘擎超
李祎承
王海
陈龙
陈小波
梁军
何友国
Original Assignee
江苏大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 江苏大学 filed Critical 江苏大学
Publication of WO2020151059A1 publication Critical patent/WO2020151059A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/01—Detecting movement of traffic to be counted or controlled
    • G08G1/017—Detecting movement of traffic to be counted or controlled identifying vehicles
    • 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

Definitions

  • the invention belongs to the technical field of vehicle intelligent driving, and in particular relates to a method for recognizing surrounding vehicle behavior under complex road conditions based on an HMM-SVM double-layer improved model.
  • Road scene understanding is one of the important components of vehicle driving assistance systems, and it is also a basic requirement for future vehicle automatic driving.
  • the actual traffic environment is a complex system in which multiple traffic participants interact with each other and dynamically change.
  • smart cars must not only have visual perception, detection and tracking capabilities, but also need to be able to identify and even recognize through vehicle-road status information. Predict the behavior of surrounding vehicles to increase the depth of perception.
  • a high-precision map refers to a high-precision, fine-defined map, covering road network data, lane line data, lane line, road edge types and traffic signs and other data, and its accuracy can reach centimeter level.
  • the high-precision map industry has entered the fast lane of development.
  • Many mainstream electronic map companies such as AutoNavi and Baidu have gradually opened up high-precision map data to the public for free. Widespread application ushered in opportunities.
  • the use of high-precision maps to establish a road joint coordinate system to obtain vehicle state characteristic information in this coordinate system breaks through the limitation of vehicle behavior recognition only on straight roads, and expands the application of surrounding vehicle behavior systems.
  • the current machine learning algorithms related to vehicle behavior recognition can be roughly divided into Bayesian networks, hidden Markov models, support vector machines and correlation vector machines.
  • the Hidden Markov (HMM) model is the most widely used because of its powerful time series modeling ability, but it ignores the influence of negative samples, and only uses the maximum likelihood value for classification, especially when the output multidimensional probability is When the difference between the two is small, it is easy to cause a higher false recognition rate.
  • the support vector machine algorithm (SVM) has significant advantages in two classification problems, which can reflect the difference between categories to a greater extent. It can map the linearly inseparable samples in the low-dimensional space to the high-dimensional space to make it as large as possible The Euclidean distance of, separates similar samples well.
  • this invention proposes a HMM-SVM two-layer improved model that introduces a difference factor , Give full play to the respective advantages of the two classifiers, thereby significantly improving the recognition ability of the surrounding vehicle behavior system.
  • the present invention proposes a method for recognizing surrounding vehicle behavior under complex road conditions based on the HMM-SVM double-layer improved model, which can accurately and accurately determine the behavior of surrounding vehicles under complex road conditions in real time.
  • Recognition provides a reference basis for the decision-making and planning of smart vehicles.
  • a method for identifying surrounding vehicle behaviors under complex road conditions based on the HMM-SVM double-layer improved model includes:
  • the trained HMM model and SVM model are formed into a two-layer model, and a threshold processor is placed between the two-layer models.
  • the threshold processor obtains a large difference between the probability values, the HMM layer recognition result is directly output;
  • extract the corresponding behaviors of the two HMM models with the largest probability value select the corresponding SVM model for two classification, output the SVM layer recognition result, use the NSGA-II algorithm to optimize, obtain the best difference factor ⁇ , and finally get HMM-SVM two-layer improved model;
  • the tracked target vehicle transmits the collected driving information of the vehicle to the vehicle in real time through the Internet of Vehicles.
  • the vehicle uses the trained HMM-SVM double-layer improved model to identify the behavior mode of the tracked vehicle.
  • the typical surrounding vehicle behaviors include car following, left lane change, right lane change, and overtaking.
  • the center line of the lane is the coordinate axis X
  • the line perpendicular to the tangent of the coordinate axis X is the coordinate axis Y.
  • the feature vector in the road joint coordinate system includes the position (Xego, Yego), speed (V Xego , V Yego ), acceleration (a Xego , a Yego ) of the vehicle at time t and the location of surrounding vehicles (Xaro, Yaro), speed (V Xaro , V Yaro ), acceleration (a Xaro , a Yaro ).
  • ⁇ X Xaro-Xego
  • ⁇ Y Yaro-Yego
  • ⁇ V X V Xaro -V Xego
  • ⁇ V Y V Yaro -V Yego
  • ⁇ a X a Xaro -a Xego
  • ⁇ a Y a Yaro -a Yego
  • the SVM model On the basis of the HMM single model to identify the behavior of surrounding vehicles, the SVM model is added to form a two-layer model to identify the behavior of surrounding vehicles, and the excellent time series modeling ability of HMM and the strong two classification ability of SVM are organically combined to give full play to The advantages of the two classifiers;
  • Figure 1 is a schematic diagram of the establishment of a road joint coordinate system
  • Figure 1 (a) is a schematic diagram of the establishment of a straight road in the joint coordinate system
  • Figure 1 (b) is a schematic diagram of the establishment of a curve in the joint coordinate system
  • Figure 2 is a block diagram of a method for identifying behaviors of surrounding vehicles under complex road conditions based on the HMM-SVM two-layer improved model.
  • Step1 State feature acquisition and data processing
  • each experimental vehicle is equipped with an interface to connect to a high-precision map.
  • the high-precision map is different from traditional maps.
  • accuracy is accuracy, which can reach centimeter-level accuracy, which can help experiments.
  • the vehicle achieves high-precision positioning.
  • Each experimental vehicle has an independent ID, which is connected to the Internet of Vehicles through the OBD (On-Board Diagnostics) interface, uploads the data extracted by the vehicle to the cloud, and interacts with other vehicles in real time through the background.
  • OBD On-Board Diagnostics
  • the sampling frequency is set to 50 Hz, that is, the time between two data before and after acquisition is 0.02s.
  • the surrounding vehicles here include not only vehicles in the driver's field of view or vehicle-mounted camera, lidar, millimeter wave radar and other equipment sensing areas, but also vehicles in blind areas, such as vehicles blocked by mountains at the turn of the winding road.
  • Figure 1 (a) is a schematic diagram of the establishment of a straight road in the joint coordinate system
  • Figure 1 (b) is a schematic diagram of the establishment of a curve in the joint coordinate system.
  • Step2 Model training and learning
  • Hidden Markov Model is a probabilistic model about time series, which uses observable variables to study unobservable variables.
  • the HMM model for each vehicle behavior recognition to obtain the initial parameters N, M, A, B, ⁇ ; the characteristics of the same category of surrounding vehicle behavior (car following, left lane change, right lane change, overtaking) obtained in Step1
  • the vector X is used as the input of the HMM model.
  • the HMM corresponding to the behavior category completes the learning of the first layer model.
  • Support Vector Machine (SVM) model is currently the most widely used two-classifier, and it shows many unique advantages in solving small sample, nonlinear and high-dimensional pattern recognition.
  • the basic principle is to find an optimal hyperplane that meets the requirements of data classification, so that the hyperplane has the largest distance from the two types of sample points while ensuring the classification accuracy.
  • the optimal hyperplane requires the maximization of the classification interval, and the distance between two parallel hyperplanes is 2/
  • is required to be minimized, that is, there is a minimization equation when solving: ⁇ ( ⁇ ) 1/2
  • each group includes two behavior categories i, j, one as a positive sample and the other as a negative sample .
  • Step3 Form a two-layer improved model
  • the trained HMM model and SVM model are formed into a two-layer model, and a threshold processor is placed between the two-layer model, and the training sample set corresponding to each typical surrounding vehicle behavior is formed into a total training sample set (B1, B2,..., BN ), the feature vector X of each sample is input to the HMM model as an observation sequence, and the multi-dimensional output probability (P 1 , P 2 ,..., P n ) of each surrounding vehicle behavior is used as the input of the threshold processor.
  • the threshold processor compares the output probabilities, and extracts the two largest probabilities P max-i and P max-ii ; let ⁇ be the difference factor in the threshold processor, when P max-ii /P max-i ⁇
  • P max-ii /P max-i ⁇ the recognition results of the first-level model are directly output, that is, the behavior corresponding to the maximum output probability (P max-i ) of each model.
  • P max-ii /P max-i ⁇ the difference between the output results of the HMM model is small. At this time, if the behavior corresponding to the maximum probability (P max-i ) is directly output, it is easy to produce higher
  • the recognition error rate requires the recognition of the second layer model.
  • the present invention further analyzes the recognition results of the training sample set, and takes the recognition accuracy rate of each behavior of the training sample u 1 , u 2 ,..., u N , the total identification time is T 0 .
  • the NSGA-II algorithm is one of the most popular multi-objective genetic algorithms at present, which has the advantages of fast running speed and good solution set convergence.
  • the tracked target vehicle transmits the collected vehicle driving information to the host vehicle in real time through the Internet of Vehicles.
  • the vehicle combines the features of the two vehicles in the road joint coordinate system to form a new feature vector, and uses the trained HMM-SVM double-layer improvement
  • the model identifies the behavior mode of the tracked vehicle.

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Image Analysis (AREA)
  • Traffic Control Systems (AREA)

Abstract

一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，包括1)离线训练：划分典型周边车辆行为，提取出周边车辆的状态特征信息，通过车联网传至本车，结合本车信息转化为道路联合坐标系下的状态特征信息，生成特征向量X，将X分别输入给HMM和SVM参数学习；2)模型改进：在HMM与SVM中间安置阈值处理器，使用NSGA-II算法优化获得最佳差异因子，得出HMM-SVM双层改进模型；3)在线测试：本车利用HMM-SVM双层改进模型辨别被跟踪车辆所属行为模式。本方法利用高精度地图构建的道路联合坐标系，拓展周边车辆行为系统的应用场合；将HMM优良的时序建模能力和SVM极强的二分类能力有机结合，并对双层模型进行改进，提高车辆行为识别的准确率和识别速度。

Description

一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法 技术领域
本发明属于车辆智能驾驶技术领域，具体涉及一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法。
背景技术
道路场景理解是车辆驾驶辅助系统的重要组成部分之一，同时也是未来车辆自动驾驶的基础要求。现实的交通环境是一个多交通参与者相互影响、动态变化的复杂系统，在这个复杂的系统中智能汽车不仅要具备视觉感知、检测与跟踪能力，还需要能通过车-路状态信息识别甚至是预测周边车辆的行为，从而提高感知深度。
在对周边车辆进行行为识别时，获取准确实用的周边车辆状态特征是必不可少的。现有的车辆行为识别方法在获取车辆状态特征时，将道路默认为直行路段，并不适用于复杂路况(弧形弯、U形弯等)下车辆行为的识别。高精度地图是指高精度、精细化定义的地图，涵盖道路网数据、车道线数据、车道线、道路边缘的类型和交通标志等数据，其精度可达到厘米级。且近年来，随着自动驾驶产业的飞速发展，高精地图产业进入了发展的快车道，许多主流电子地图企业如高德、百度等也已经逐步对外免费开放了高精度地图的数据，给其广泛应用迎来了机遇。利用高精度地图建立道路联合坐标系，获取车辆在该坐标系下的状态特征信息，突破了仅在直行道路对车辆行为识别的限制，拓展了周边车辆行为系统的应用场合。
目前有关车辆行为识别的机器学习算法可大致分为贝叶斯网络、隐马尔科夫模型、支持向量机和相关向量机等。其中，隐马尔科夫(HMM)模型因具有强大的时序建模能力受到最广泛的运用，但其忽略了负样本的影响，且仅用最大似然值进行分类，尤其当输出的多维概率之间差异性较小时，容易造成较高的误识别率。支持向量机算法(SVM)在二分类问题上有显著的优势，能够更大程度地反映了类别间的差异性，能够通过将低维空间线性不可分的样本映射至高维空间中，以尽可能大的欧氏距离将相似的样本很好地分隔开。因此，为充分利用好HMM优良的时序建模能力和SVM极强的二分类能力，同时考虑到识别结果的实时性与可靠性，此发明提出一种HMM-SVM双层改进模型，引入差异因子，发挥好两种分类器各自的优势，从而显著提高周边车辆行为系统的识别能力。
发明内容
本发明针对周边车辆行为识别实时性、可靠性的要求，提出了一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，能够实时准确地对复杂路况下周边车辆的行为作出识别，为智能车辆的决策规划提供参考依据。
本发明的目的可以通过以下技术方案来实现，一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，具体包括：
步骤(1)：离线训练阶段
划分N种典型周边车辆行为，提取出周边车辆的状态特征信息，通过车联网传至本车，将本车及周边车辆的状态特征信息转化均为道路联合坐标系下的状态特征信息，并生成新的特征向量X(△X，△Y，△V X，△V Y，△a X，△a Y)，将同一类别周边车辆行为的特征向量X作为观测序列输入到HMM模型，获取各个车辆行为类别的HMM；周边车辆行为两两为一组，将每组中两种行为类别对应的特征向量输给每个SVM模型进行参数学习，获取每组SVM模型；
步骤(2)：模型改进阶段
将训练好的HMM模型与SVM模型组成双层模型，并在两层模型中间安置阈值处理器，当阈值处理器得出概率值之间的差异性较大时，则直接输出HMM层识别结果；反之，将概率值最大的两个HMM模型对应的行为提取出来，选择相应的SVM模型进行二分类，输出SVM层识别结果，使用NSGA-II算法进行优化，获得最佳差异因子σ，最终得出HMM-SVM双层改进模型；
步骤(3)：在线测试阶段
被跟踪目标车辆将采集到的自车行驶信息通过车联网实时传输给本车，本车利用训练好的HMM-SVM双层改进模型辨别被跟踪车辆所属行为模式。
进一步，所述典型周边车辆行为包括跟驰、左换道、右换道、超车。
进一步，所述道路联合坐标系是以车道的中心线为坐标轴X，与坐标轴X的切线垂直的线为坐标轴Y。
进一步，所述道路联合坐标系下的特征向量包括t时刻本车的位置(Xego，Yego)、速度(V Xego，V Yego)、加速度(a Xego，a Yego)以及周边车辆的位置(Xaro，Yaro)、速度(V Xaro，V Yaro)、加速度(a Xaro，a Yaro)。
更进一步，所述特征向量X中，△X＝Xaro-Xego、△Y＝Yaro-Yego、△V X＝V Xaro-V Xego、△V Y＝V Yaro-V Yego、△a X＝a Xaro-a Xego、△a Y＝a Yaro-a Yego。
进一步，所述NSGA-II算法以识别正确率和识别时间为优化目标，所述优化目标函数 为f＝T 0-u 1-u 2-…u N，其中u 1，u 2，…，u N是训练样本各行为的识别正确率，T 0为识别总时间。
进一步，所述二分类的分类器方程为f(x)＝sign(ω T·X+b)，其中ω是可调的权值向量，b为偏置量。
本发明的有益效果为：
(1)在周车与本车状态特征获取阶段，引进了一种利用高精度地图构建的道路联合坐标系，可以突破仅在直行道路对车辆行为识别的限制，拓展了周边车辆行为系统的应用场合；
(2)在HMM单一模型识别周边车辆行为的基础上，增加SVM模型，组成双层模型来识别周边车辆行为，将HMM优良的时序建模能力和SVM极强的二分类能力有机结合，充分发挥了两种分类器各自的优势；
(3)以识别正确率和识别时间为优化目标建立数学模型，使用NSGA-II算法对阈值处理器进行多目标优化以获得差异因子σ，对HMM-SVM双层模型进行了改进，提高车辆行为识别的准确率和识别速度。
附图说明
图1是道路联合坐标系的建立示意图，图1(a)是直道在联合坐标系的建立示意图，图1(b)是弯道在联合坐标系的建立示意图；
图2是基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法框图。
具体实施方式
下面将结合附图对本发明的技术方案作进一步的说明，但是本发明的保护范围并不限于此。
Step1：状态特征获取与数据处理
首先归纳并划分N种典型的周边车辆行为，分别为跟驰、左换道、右换道、超车等。设定每辆实验车都装备有与高精度地图连接的接口，所述高精度地图除了蕴含大量的道路信息，区别于传统地图的一个重要特征就是精度，可达到厘米级的精度，能够帮助实验车实现高精度定位。各实验车都具有一个独立的ID，通过OBD(On-Board Diagnostics)接口接入车联网，把自车提取的数据上传到云端，通过后台实时与其他车辆互动。考虑到数据传输的实时性与鲁棒性，将采样频率设置为50Hz，即采集前后两次数据之间的时长为0.02s。选定一辆实验车为本车(Ego)，其周围前后左右的车辆即为周边车辆(Around1、2…)。这里的周边车辆不仅包括本车驾驶员视野或车载摄像机、激光雷达、毫米波雷达等设备感知 区域内的车辆，还包括盲区内的车辆，如盘山公路转弯口被山体遮挡的车辆等。
针对周边车辆中要进行识别的目标车辆，本车从车联网中在线下载周边车辆的状态特征信息。利用所接入的高精度地图，参考理论力学里运动质点的自然系表示法，取车道的中心线为坐标轴X，取与坐标轴X的切线垂直的线为坐标轴Y，建立道路联合坐标系，如图1所示，图1(a)是直道在联合坐标系的建立示意图，图1(b)是弯道在联合坐标系的建立示意图。将本车与所下载的周边车辆状态特征信息，转化为道路联合坐标系下的特征信息，即t时刻本车位置(Xego，Yego)、速度(V Xego，V Yego)、加速度(a Xego，a Yego)以及周边车辆位置(Xaro，Yaro)、速度(V Xaro，V Yaro)、加速度(a Xaro，a Yaro)，将上述特征信息转化为新的特征向量X(△X，△Y，△V X，△V Y，△a X，△a Y)，其中，△X＝Xaro-Xego，△Y＝Yaro-Yego，△V X＝V Xaro-V Xego，△V Y＝V Yaro-V Yego，△a X＝a Xaro-a Xego，△a Y＝a Yaro-a Yego。依据上述步骤等量获取各典型周边车辆行为类别对应的特征向量集作为训练样本集，其中根据行为类别，训练样本集可设为B 1，B 2，…，B N。
Step2：模型训练学习
(1)隐马尔科夫模型训练学习
隐马尔科夫模型(Hidden Markov Model,HMM)是关于时序的概率模型，通过可观测到的变量去研究不可观测的变量。可以设隐马可夫模型为一个五元组(Q，V，A，B，π)，其中隐藏状态Q＝{Q 1，Q 2，…，Q N}，N为隐藏状态的数目；可观察状态V＝{V 1，V 2，…，V M}，M为观察状态的数目；隐藏状态转移概率矩阵A＝[a ij] N×N的元素表示HMM模型中各个隐藏状态之间的转移概率，a ij是在t时刻隐藏状态为Q i、在t+1时刻隐藏状态为Q j的概率，a ij＝P(I t+1＝Q j|I t＝Q i)，i＝1，2…，N；j＝1，2…，N，I是长度为T的状态序列，且I＝{I 1,I 2,…,I T}；混淆矩阵B＝[b j(k)] N×M的元素表示HMM模型中各个隐藏状态和观察状态之间的转移概率,b j(k)表示在t时刻，隐藏状态为Q j、观察状态为O t的概率，b j(k)＝P(O t＝V k|I t＝Q j)，k＝1，2…，M；j＝1，2…，N，O是对应的观测序列；初始状态概率矩阵π＝(π i)，其中π i＝P(I 1＝Q i)，i＝1，2，…，N，表示初始时刻t＝1各个隐藏状态Q i的概率。
首先对每一个车辆行为识别HMM模型初始化，获得初始参数N、M、A、B、π；将 Step1中获取的同一类别周边车辆行为(跟驰、左换道、右换道、超车)的特征向量X作为HMM模型的输入，依据模型初始化后的参数，采用Baum-Welch迭代算法调整模型λ＝(A，B，π)的参数，使概率函数最大化，逐步更新模型参数，最终获取各个车辆行为类别对应的HMM，完成第一层模型的学习。
(2)支持向量机模型训练学习
支持向量机(Support Vector Machine，SVM)模型是目前使用最广泛的二分类器，在解决小样本、非线性及高维模式识别中表现出许多特有的优势。其基本原理是寻找一个满足数据分类要求的最优超平面，使得超平面在确保分类精度的情况下，超平面与两类样本点距离最大。该超平面应满足ω T·X+b＝0，其中ω是可调的权值向量，b是偏置量，X为特征向量，T是矩阵的转置符号。最优超平面要求分类间隔最大化，而两平行超平面的距离为2/||ω||，也就是要求||ω||最小化，即在进行求解时有最小化方程：φ(ω)＝1/2||ω||＝1/2(ω，ω)，为使得所有样本在超平面外，上式还应满足约束条件Y iω T·X i+b＞1，Y i∈{-1，1}，i＝1，2，…l，其中Y i表示样本类别，l为样本个数。
将N种典型周边车辆行为两两为一组，共分成h＝N(N-1)/2组，每组包括两种行为类别i、j，一种作为正样本，另一种作为负样本。将Step1中获取正负样本对应的训练样本集B i、B j合并为该组的训练样本集(B i，B j)，用来训练SVM模型。基于MATLAB环境下，使用libsvm支持向量机工具，分别将每组训练样本集(B i，B j)的特征向量X输给每个SVM模型进行参数学习，找出每组的最优超平面结果f k(x)＝ω T·X+b，k＝1，2，…h，从而获取各组进行二分类的SVM模型。
Step3：组成双层改进模型
将训练好的HMM模型与SVM模型组成双层模型，并在两层模型中间安置阈值处理器，将各个典型周边车辆行为对应的训练样本集组成总的训练样本集(B1，B2，…，BN)，将每个样本的特征向量X作为观测序列输入到HMM模型，并把各个周边车辆行为的多维输出概率(P 1，P 2，…，P n)作为阈值处理器的输入。阈值处理器对各输出概率进行比较，提取出最大的两个概率P max-i、P max-ii；设σ为阈值处理器中的差异因子，当P max-ii/P max-i≤σ时，则HMM模型输出结果之间的差异性较大，直接输出第一层模型的识别结果，即各个模型输出概率最大(P max-i)所对应的行为。当P max-ii/P max-i≥σ时，则HMM模型输出结果之间的 差异性较小，这时如果直接输出概率最大(P max-i)所对应的行为，容易产生较高的识别错误率，需要进行第二层模型的识别。将P max-i、P max-ii对应的行为Q i、Q ii提取，将特征向量X输入到行为Q i、Q ii所在的那组SVM模型中，由SVM模型分类器方程f(x)＝sign(ω T·X+b)输出该样本所在的行为类别，将此结果作为最终的识别结果。鉴于阈值处理器中的阈值σ对双层模型识别的实时性与可靠性有重要影响，本发明对训练样本集的识别结果进行进一步分析，取训练样本各行为的识别正确率为u 1，u 2，…，u N，识别总时间为T 0。
为对HMM-SVM双层模型进行改进，以识别正确率和识别时间为优化目标建立数学模型，使用NSGA-II算法对阈值处理器进行多目标优化。所述NSGA-Ⅱ算法是目前最流行的多目标遗传算法之一，具有运行速度快、解集的收敛性好的优点。定义优化目标函数为f＝T 0-u 1-u 2-…u N，针对识别正确率和识别时间的改进来优化差异因子σ，最终获得HMM-SVM双层改进模型。
Step4：在线测试阶段
被跟踪目标车辆将采集到的自车行驶信息通过车联网实时传输给主车，本车结合两车在道路联合坐标系下的特征构成新的特征向量，利用训练好的HMM-SVM双层改进模型辨别被跟踪车辆所属行为模式。
所述实施例为本发明的优选的实施方式，但本发明并不限于上述实施方式，在不背离本发明的实质内容的情况下，本领域技术人员能够做出的任何显而易见的改进、替换均属于本发明的保护范围。

Claims (8)

  1. 一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，其特征在于，具体包括：
    步骤(1)：离线训练阶段
    划分N种典型周边车辆行为，提取出周边车辆的状态特征信息，通过车联网传至本车，将本车及周边车辆的状态特征信息转化均为道路联合坐标系下的状态特征信息，并生成新的特征向量X(△X，△Y，△V X，△V Y，△a X，△a Y)，将同一类别周边车辆行为的特征向量X作为观测序列输入到HMM模型，获取各个车辆行为类别的HMM；周边车辆行为两两为一组，将每组中两种行为类别对应的特征向量输给每个SVM模型进行参数学习，获取每组SVM模型；
    步骤(2)：模型改进阶段
    将训练好的HMM模型与SVM模型组成双层模型，并在两层模型中间安置阈值处理器，当阈值处理器得出概率值之间的差异性较大时，则直接输出HMM层识别结果；反之，将概率值最大的两个HMM模型对应的行为提取出来，选择相应的SVM模型进行二分类，输出SVM层识别结果，使用NSGA-II算法进行优化，获得最佳差异因子σ，最终得出HMM-SVM双层改进模型；
    步骤(3)：在线测试阶段
    被跟踪目标车辆将采集到的自车行驶信息通过车联网实时传输给本车，本车利用训练好的HMM-SVM双层改进模型辨别被跟踪车辆所属行为模式。
  2. 根据权利要求1所述的一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，其特征在于，所述典型周边车辆行为包括跟驰、左换道、右换道、超车。
  3. 根据权利要求1所述的一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，其特征在于，所述道路联合坐标系是以车道的中心线为坐标轴X，与坐标轴X的切线垂直的线为坐标轴Y。
  4. 根据权利要求1或3所述的一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，其特征在于，所述道路联合坐标系下的特征向量包括t时刻本车的位置(Xego，Yego)、速度(V Xego，V Yego)、加速度(a Xego，a Yego)以及周边车辆的位置(Xaro，Yaro)、速度(V Xaro，V Yaro)、加速度(a Xaro，a Yaro)。
  5. 根据权利要求4所述的一种基于HMM-SVM双层改进模型的复杂路况下周边车辆 行为识别方法，其特征在于，所述特征向量X中，△X＝Xaro-Xego、△Y＝Yaro-Yego、△V X＝V Xaro-V Xego、△V Y＝V Yaro-V Yego、△a X＝a Xaro-a Xego、△a Y＝a Yaro-a Yego。
  6. 根据权利要求1所述的一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，其特征在于，所述NSGA-II算法以识别正确率和识别时间为优化目标。
  7. 根据权利要求6所述的一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，其特征在于，所述优化目标函数为f＝T 0-u 1-u 2-…u N，其中u 1，u 2，…，u N是训练样本各行为的识别正确率，T 0为识别总时间。
  8. 根据权利要求1所述的一种基于HMM-SVM双层改进模型的复杂路况下周边车辆行为识别方法，其特征在于，所述二分类的分类器方程为f(x)＝sign(ω T·X+b)，其中ω是可调的权值向量，b为偏置量。
PCT/CN2019/077584 2019-01-22 2019-03-11 一种基于hmm-svm双层改进模型的复杂路况下周边车辆行为识别方法 WO2020151059A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910058190.5 2019-01-22
CN201910058190.5A CN109886304B (zh) 2019-01-22 2019-01-22 一种基于hmm-svm双层改进模型的复杂路况下周边车辆行为识别方法

Publications (1)

Publication Number Publication Date
WO2020151059A1 true WO2020151059A1 (zh) 2020-07-30

Family

ID=66926458

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/077584 WO2020151059A1 (zh) 2019-01-22 2019-03-11 一种基于hmm-svm双层改进模型的复杂路况下周边车辆行为识别方法

Country Status (2)

Country Link
CN (1) CN109886304B (zh)
WO (1) WO2020151059A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112069599A (zh) * 2020-08-28 2020-12-11 北京科技大学 一种履带车辆行驶性能可靠性预测方法
CN114299742A (zh) * 2022-01-20 2022-04-08 福建工程学院 一种高速公路的限速信息动态识别与更新推荐方法

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110288835B (zh) * 2019-06-28 2021-08-17 江苏大学 一种基于运动学预测补偿机制的周边车辆行为实时识别方法
CN110986994B (zh) * 2019-11-14 2021-08-03 苏州智加科技有限公司 一种基于高噪音车辆轨迹数据的换道意图自动标注方法
CN111354193B (zh) * 2020-02-26 2021-09-10 江苏大学 一种基于5g通信的高速公路车辆异常行为预警系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106355948A (zh) * 2015-07-17 2017-01-25 本田技研工业株式会社 转弯预测
CN106971194A (zh) * 2017-02-16 2017-07-21 江苏大学 一种基于改进hmm和svm双层算法的驾驶意图识别方法
CN107967486A (zh) * 2017-11-17 2018-04-27 江苏大学 一种基于v2v通信与hmm-gbdt混合模型的周边车辆行为识别方法
US20180190377A1 (en) * 2016-12-30 2018-07-05 Dirk Schneemann, LLC Modeling and learning character traits and medical condition based on 3d facial features
CN108470460A (zh) * 2018-04-11 2018-08-31 江苏大学 一种基于智能手机与rnn的周边车辆行为识别方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106355948A (zh) * 2015-07-17 2017-01-25 本田技研工业株式会社 转弯预测
US20180190377A1 (en) * 2016-12-30 2018-07-05 Dirk Schneemann, LLC Modeling and learning character traits and medical condition based on 3d facial features
CN106971194A (zh) * 2017-02-16 2017-07-21 江苏大学 一种基于改进hmm和svm双层算法的驾驶意图识别方法
CN107967486A (zh) * 2017-11-17 2018-04-27 江苏大学 一种基于v2v通信与hmm-gbdt混合模型的周边车辆行为识别方法
CN108470460A (zh) * 2018-04-11 2018-08-31 江苏大学 一种基于智能手机与rnn的周边车辆行为识别方法

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112069599A (zh) * 2020-08-28 2020-12-11 北京科技大学 一种履带车辆行驶性能可靠性预测方法
CN112069599B (zh) * 2020-08-28 2023-08-22 北京科技大学 一种履带车辆行驶性能可靠性预测方法
CN114299742A (zh) * 2022-01-20 2022-04-08 福建工程学院 一种高速公路的限速信息动态识别与更新推荐方法
CN114299742B (zh) * 2022-01-20 2023-10-13 福建工程学院 一种高速公路的限速信息动态识别与更新推荐方法

Also Published As

Publication number Publication date
CN109886304A (zh) 2019-06-14
CN109886304B (zh) 2023-09-29

Similar Documents

Publication Publication Date Title
WO2020151059A1 (zh) 一种基于hmm-svm双层改进模型的复杂路况下周边车辆行为识别方法
US10733755B2 (en) Learning geometric differentials for matching 3D models to objects in a 2D image
CN111310583B (zh) 一种基于改进的长短期记忆网络的车辆异常行为识别方法
CN107492251B (zh) 一种基于机器学习与深度学习的驾驶员身份识别与驾驶状态监测方法
Jain et al. Discrete residual flow for probabilistic pedestrian behavior prediction
Gadepally et al. A framework for estimating driver decisions near intersections
Roy et al. Multi-modality sensing and data fusion for multi-vehicle detection
CN111460881A (zh) 基于近邻判别的交通标志对抗样本检测方法和分类装置
Zhang et al. A framework for turning behavior classification at intersections using 3D LIDAR
CN108470460B (zh) 一种基于智能手机与rnn的周边车辆行为识别方法
CN113537411B (zh) 一种基于毫米波雷达的改进模糊聚类方法
Lee et al. Probabilistic inference of traffic participants' lane change intention for enhancing adaptive cruise control
CN114120270A (zh) 一种基于注意力和采样学习的点云目标检测方法
CN106650814B (zh) 一种基于车载单目视觉室外道路自适应分类器生成方法
CN108981728A (zh) 一种智能车辆导航地图建立方法
Gao et al. Deep learning‐based hybrid model for the behaviour prediction of surrounding vehicles over long‐time periods
CN115015908A (zh) 基于图神经网络的雷达目标数据关联方法
CN116168543B (zh) 基于毫米波雷达的车辆轨迹修正方法、装置及存储介质
Liu et al. Time-Series Misalignment Aware DNN Adversarial Attacks for Connected Autonomous Vehicles
Chen et al. Research and application of variant granularity feedback recognition method based on maximal entropy and cloud-membership
US20240124003A1 (en) Method and system for modeling personalized car-following driving styles with model-free inverse reinforcement learning
Jaspers et al. Visual navigation with efficient ConvNet features
Wang et al. A Novel Possibilistic Clustering Algorithm for Measurement Data of Vehicle MMW Radar
Cai et al. CAEV-Deep Learning-Based Hybrid Model for the Behavior Prediction of Surrounding Vehicles Over Long-time Periods
Song et al. Camera Agnostic Two-Head Network for Ego-Lane Inference

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19912103

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19912103

Country of ref document: EP

Kind code of ref document: A1