技术领域Technical Field
本发明属于智能汽车技术领域，具体的说是一种基于智能汽车交通态势风险认知的场景提取方法。The present invention belongs to the technical field of intelligent vehicles, and specifically is a scene extraction method based on intelligent vehicle traffic situation risk recognition.
背景技术Background technique
随着智能汽车技术的不断进步，整个行业正面临着更加复杂的交通环境和多样化的驾驶任务这些新挑战。在智能汽车测试领域，目前采用的基于功能的测试方法，诸如对车辆紧急制动(AEB)、车道保持辅助(ACC)、自适应巡航(LKA)等功能的测试，尽管有效，但此类方法测试的功能较为单一，无法充分验证多项功能综合表现的性能。同时，基于里程的测试方法，不论是在真实交通道路还是虚拟仿真引擎中，绝大多数时间车辆所面临的均为安全场景，而危险场景的出现率极低，这导致测试效率低下，并且存在真实车道测试带给公众财产的安全问题。因此，在面对复杂的交通环境和多样化的驾驶任务时，一种有效的方法是在公开的自然驾驶数据集或数据采集车采集的数据集中，分析车辆的语义信息，筛选出危险场景，并仅在这些危险场景下对智能汽车的算法与硬件进行测试。这种方法能够有效减少测试成本，同时加速智能汽车算法的实际应用，是目前提升测试效率的一个高效策略。在海量自然驾驶数据集中有效地识别危险场景构成了一项关键挑战。当前，处理海量数据集以提取相关信息的主要方法可归类为机器学习、无监督学习及聚类等技术。机器学习方法虽然有其优势，但它依赖于对数据的预先标注。如果训练数据集未能全面涵盖所有潜在的危险测试用例，机器学习方法可能无法生成具有广泛适用性的模型权重。此外，无监督学习和聚类方法在对含有典型危险特征的自然驾驶数据片段进行精确分类方面表现不足，这可能导致包含关键危险场景的数据片段在大量数据中被遗漏，从而难以有效提取。因此，开发能够精确地识别并分类这些关键数据片段，提取危险场景“难”的问题是当前研究领域面临的一项重要任务。With the continuous advancement of smart car technology, the entire industry is facing new challenges such as more complex traffic environments and diversified driving tasks. In the field of smart car testing, the current function-based testing methods, such as testing of vehicle emergency braking (AEB), lane keeping assist (ACC), adaptive cruise control (LKA) and other functions, are effective, but the functions tested by such methods are relatively single and cannot fully verify the performance of the comprehensive performance of multiple functions. At the same time, in the mileage-based testing method, whether on real traffic roads or in virtual simulation engines, most of the time, vehicles are faced with safe scenarios, while the occurrence rate of dangerous scenarios is extremely low, which leads to low testing efficiency and the safety of public property brought by real lane testing. Therefore, in the face of complex traffic environments and diversified driving tasks, an effective method is to analyze the semantic information of the vehicle in the public natural driving data set or the data set collected by the data collection vehicle, filter out dangerous scenarios, and test the algorithms and hardware of smart cars only in these dangerous scenarios. This method can effectively reduce the testing cost and accelerate the practical application of smart car algorithms. It is an efficient strategy to improve testing efficiency. Effectively identifying dangerous scenarios in massive natural driving data sets constitutes a key challenge. Currently, the main methods for processing massive data sets to extract relevant information can be classified into machine learning, unsupervised learning, and clustering techniques. Although machine learning methods have their advantages, they rely on pre-labeling of data. If the training data set fails to comprehensively cover all potential dangerous test cases, machine learning methods may not be able to generate model weights with broad applicability. In addition, unsupervised learning and clustering methods are not good at accurately classifying natural driving data segments containing typical dangerous features, which may cause data segments containing key dangerous scenes to be missed in large amounts of data, making them difficult to extract effectively. Therefore, developing a method that can accurately identify and classify these key data segments and solve the "difficult" problem of extracting dangerous scenes is an important task facing the current research field.
基于以上难点，本发明针对智能汽车测试中生成危险场景的重要挑战，提出了一种创新性的方法。此方法融合了非对称Yukawa交通势能模型与高斯车道线风险模型，旨在精准评估主车在特定时刻遭遇的交通风险。其次，本专利采用了无监督聚类的机器学习技术，基于计算得出的交通风险指数，有效筛选出对智能汽车安全性测试至关重要的关键危险场景。本方法的创新性在于提出了一种改良的汤川势能(Yukawa)模型-非对称Yukawa交通势能，该模型将物理学概念与智能汽车技术相融合，并由此开发出一种创新的智能汽车关键危险测试场景生成方法。Based on the above difficulties, the present invention proposes an innovative method for the important challenge of generating dangerous scenarios in smart car testing. This method combines the asymmetric Yukawa traffic potential model with the Gaussian lane line risk model, aiming to accurately assess the traffic risks encountered by the main vehicle at a specific moment. Secondly, this patent uses unsupervised clustering machine learning technology to effectively screen out key dangerous scenarios that are crucial to the safety testing of smart cars based on the calculated traffic risk index. The innovation of this method lies in the proposal of a modified Yukawa potential model-asymmetric Yukawa traffic potential, which integrates physical concepts with smart car technology, and thereby develops an innovative method for generating key dangerous test scenarios for smart cars.
发明内容Summary of the invention
本发明提供了一种基于智能汽车交通态势风险认知的场景提取方法，该方法适用于测试场景的提取，能够更有效地量化车辆行驶时的危险场景，为智能汽车测试提供了一种更为高效的测试方法，并且能够更好地理解和评估智能汽车在实际道路环境中面临的复杂挑战，进而提高智能汽车的安全性和可靠性，解决了从开源自然驾驶数据集或数据采集车辆自行构建的数据集中提取危险测试场景“难”的问题。The present invention provides a scenario extraction method based on risk recognition of intelligent vehicle traffic situations. The method is suitable for extracting test scenarios, can more effectively quantify dangerous scenarios during vehicle driving, and provides a more efficient testing method for intelligent vehicle testing. It can also better understand and evaluate the complex challenges faced by intelligent vehicles in actual road environments, thereby improving the safety and reliability of intelligent vehicles, and solving the problem of the difficulty of extracting dangerous test scenarios from open source natural driving datasets or datasets constructed by data acquisition vehicles themselves.
本发明技术方案结合附图说明如下：The technical solution of the present invention is described as follows in conjunction with the accompanying drawings:
一种基于势场的智能汽车交通态势风险认知方法，包括以下步骤：A method for intelligent vehicle traffic situation risk recognition based on potential field includes the following steps:
步骤一、输入城市地区的自然驾驶数据集；所述自然驾驶数据集包括车辆轨迹序列数据以及高精地图数据；Step 1: input a natural driving dataset in an urban area; the natural driving dataset includes vehicle trajectory sequence data and high-precision map data;
步骤二、通过车辆轨迹序列数据建立非对称Yukawa交通势能，并计算主车行驶过程中某一时刻所面临的风险；通过高精地图数据建立高斯车道线风险计算模块，并计算动态时空车辆的总风险值；Step 2: Establish asymmetric Yukawa traffic potential energy through vehicle trajectory sequence data, and calculate the risk faced by the main vehicle at a certain moment during driving; establish a Gaussian lane line risk calculation module through high-precision map data, and calculate the total risk value of dynamic space-time vehicles;
步骤三、分析自然驾驶数据的指定片段中，对于主车在行驶过程中某一特定时刻所遭遇的风险进行量化时，采用了两个关键计算模块：一是非对称Yukawa交通势能模块，用于估算该时刻的风险值；二是高斯车道线风险模块，用于负责计算动态时空车辆环境中的风险值；这两个模块的计算结果进行累加，形成建立车辆交通态势风险认知总势场的基础，该累加结果用于最终评估当前自然驾驶数据集片段时段的车辆行车风险；Step 3: Analyze the specified segment of natural driving data. When quantifying the risk encountered by the main vehicle at a specific moment in the driving process, two key calculation modules are used: one is the asymmetric Yukawa traffic potential module, which is used to estimate the risk value at that moment; the other is the Gaussian lane line risk module, which is responsible for calculating the risk value in the dynamic spatiotemporal vehicle environment. The calculation results of these two modules are accumulated to form the basis for establishing the total potential field of vehicle traffic situation risk cognition. The accumulated results are used to finally evaluate the vehicle driving risk in the current natural driving data set segment period;
步骤四、通过无监督聚类机器学习算法，对车辆风险值数据集中的不同轨迹片段所对应的风险值进行分类处理；依托自动驾驶测试场景相关的标准或专家经验设定风险阈值，以便于将无监督聚类的结果进行明确界定，进一步识别和分析出危险场景或非常规场景的特定类别。Step 4: Classify the risk values corresponding to different trajectory segments in the vehicle risk value dataset through an unsupervised clustering machine learning algorithm; set risk thresholds based on standards or expert experience related to autonomous driving test scenarios, so as to clearly define the results of unsupervised clustering and further identify and analyze specific categories of dangerous or unconventional scenarios.
进一步的，所述高精地图数据为包含每条路段的车道线类型和车道分布特征的高精度地图数据信息，以及关键道路属性信息；所述车辆轨迹序列数据通过开源数据集获得。Furthermore, the high-precision map data is high-precision map data information containing lane line types and lane distribution characteristics of each road section, as well as key road attribute information; the vehicle trajectory sequence data is obtained through an open source data set.
进一步的，所述步骤二的具体方法如下：Furthermore, the specific method of step 2 is as follows:
1)采用非对称Yukawa交通势能表征主车与周围物体的相互作用关系，如下所示：1) The asymmetric Yukawa traffic potential is used to characterize the interaction between the main vehicle and surrounding objects, as shown below:
式中，PAYTP为主车i与车辆j即前向车辆与车辆k即后向车辆相加之后的非对称Yukawa交通总势能；A和B为常数，为缩放因子，决定风险函数的基线强度；Δv为车辆之间的纵向速度差；Δv为车辆之间的纵向速度差；Δa为车辆之间的纵向加速度差；α，α′，β，β′，γ，γ′为衰减系数，用于调节距离、速度和加速度对势能的影响；指数ζ为一个非线性因子，当车辆距离非常近时，增加势能，强调高速接近的车辆对主车的影响；Where PAYTP is the total asymmetric Yukawa traffic potential energy after the sum of the main vehicle i and vehicle j, i.e., the forward vehicle, and vehicle k, i.e., the backward vehicle; A and B are constants, and are scaling factors that determine the baseline strength of the risk function; Δv is the longitudinal velocity difference between vehicles; Δv is the longitudinal velocity difference between vehicles; Δa is the longitudinal acceleration difference between vehicles; α, α′, β, β′, γ, γ′ are attenuation coefficients used to adjust the effects of distance, speed, and acceleration on the potential energy; the exponent ζ is a nonlinear factor that increases the potential energy when the vehicles are very close, emphasizing the effect of vehicles approaching at high speed on the main vehicle;
2)利用高斯分布来量化和分析车辆相对于车道中心的横向偏移量Δdl、车辆航向和车道方向的角度差θ与车辆的尺寸宽度w和长度l以及车道线类型type，从而得出在特定的车辆参数和静态道路条件下，车辆行驶的风险概率，如下所示：2) Gaussian distribution is used to quantify and analyze the lateral offset Δdl of the vehicle relative to the lane center, the angular difference θ between the vehicle heading and the lane direction, the vehicle's width w and length l, and the lane line type, so as to obtain the risk probability of vehicle driving under specific vehicle parameters and static road conditions, as shown below:
其中，θ＝φ-λ (6)Where θ = φ - λ (6)
λ＝arctan2(Δlat，Δlon) (7)λ＝arctan2(Δlat，Δlon) (7)
h(θ，w，l)＝ecos(θ)·Vsize (8)h(θ,w,l)＝ecos(θ) ·Vsize (8)
式中，μ为权重，旨在量化不同车道线类型对车辆行为的约束强度；Δdl为量化车辆横向偏移的指标，表示车辆中心点到最近车道线点的距离；标准差σ决定高斯分布宽度的参数，影响关于车辆横向偏移Δdl的风险概率分布的形状；θ为车辆实际行驶方向与车道线方向的角度差；两点的精度差Δlon＝lon2-lon1；两点的纬度差Δlat＝lat2-lat1；两点分别为当前车辆所在车道位置距离最近车道线上的两个连续的点P1(lat1，lon1)以及点P2(lat2，lon2)；为车辆的实际航向；λ为两个连续经纬度坐标P1(lat1，lon1)和P2(lat2，lon2)所定义的微小车道线方向；h(θ，w，l)为车道几何影响因子函数；ecos(θ)为量化车辆航向与车道线方向之间的对齐程度；Vsize为尺寸调控因子；c为常数，为比例因子；b为分母中的常数，为归一化因子，提供Vsize函数的基础值。Wherein, μ is the weight, which aims to quantify the constraint strength of different lane line types on vehicle behavior; Δdl is an indicator for quantifying the lateral offset of the vehicle, which represents the distance from the center point of the vehicle to the nearest lane line point; the standard deviation σ determines the parameter of the Gaussian distribution width, which affects the shape of the risk probability distribution of the vehicle lateral offset Δdl ; θ is the angular difference between the actual driving direction of the vehicle and the direction of the lane line; the accuracy difference between the two points Δlon = lon2 -lon1 ; the latitude difference between the two points Δlat = lat2 -lat1 ; the two points are two consecutive points P1 (lat1 , lon1 ) and point P2 (lat2 , lon2 ) on the lane line nearest to the current vehicle's lane position; is the actual heading of the vehicle; λ is the direction of the tiny lane line defined by two consecutive longitude and latitude coordinates P1 (lat1 , lon1 ) and P2 (lat2 , lon2 ); h(θ, w, l) is the lane geometry influence factor function; ecos(θ) is the degree of alignment between the quantified vehicle heading and the lane line direction; Vsize is the size control factor; c is a constant, which is the proportional factor; b is the constant in the denominator, which is the normalization factor and provides the basic value of the Vsize function.
进一步的，所述权重μ根据车道线的不同类型分为四个类别：实线solid、虚线dashed、双实线double_solid和车道物理边界barrier；权重分配的顺序为：车道物理边界获得最高的权重，其次是双实线，再次是实线，而虚线的权重最低。Furthermore, the weight μ is divided into four categories according to different types of lane lines: solid line, dashed line, double solid line and lane physical boundary barrier; the order of weight allocation is: the lane physical boundary obtains the highest weight, followed by the double solid line, then the solid line, and the dashed line has the lowest weight.
进一步的，所述步骤三的具体方法如下：Furthermore, the specific method of step three is as follows:
式中，通过时刻t车辆的非对称Yukawa总势能与该时刻的高斯车道风险势能，共同构建主导车辆行驶过程中在某一时刻下的综合风险势场模型，从而最终评估当前时刻的车辆行车风险。In the formula, the asymmetric Yukawa total potential energy of the vehicle at time t and the Gaussian lane risk potential energy at that time are used to jointly construct the comprehensive risk potential field model of the dominant vehicle at a certain moment in the driving process, so as to finally evaluate the vehicle driving risk at the current moment.
本发明的有益效果为：The beneficial effects of the present invention are:
1)本发明的基础建立在高精度的道路地图信息之上；该方法的初步步骤涉及对道路环境关键参数的动态监测，具体而言，此阶段包括对车道数量、车道线类型(实线与虚线)、车道边缘(作为不可跨越的物理界限)以及车道左右两侧其他车道的存在情况进行监测。其次，为了深入揭示道路网络结构，每个路段需分配独特的ID标识符，这些ID标识符不仅起到简单路段类型识别作用，ID标识符同时也揭示了路段间的复杂连接关系，如每个路段的左侧、右侧及前方路段。所述的带有详细路段标识符的高精地图道路网络结构，可以寻找具有类似结构的开源自然驾驶数据集，也可以通过实车数据采集车自行搭建，通过这样的数据集可以为解决在复杂城市道路条件下的风险建模提供支撑；1) The foundation of the present invention is built on high-precision road map information; the preliminary steps of the method involve dynamic monitoring of key parameters of the road environment. Specifically, this stage includes monitoring the number of lanes, lane line types (solid lines and dashed lines), lane edges (as physical boundaries that cannot be crossed), and the existence of other lanes on the left and right sides of the lanes. Secondly, in order to deeply reveal the road network structure, each road section needs to be assigned a unique ID identifier. These ID identifiers not only play a simple role in identifying the type of road section, but the ID identifiers also reveal the complex connection relationship between road sections, such as the left, right and front sections of each road section. The high-precision map road network structure with detailed road section identifiers can find open source natural driving data sets with similar structures, or it can be built by itself through real vehicle data collection vehicles. Such data sets can provide support for risk modeling under complex urban road conditions;
2)本发明基于车辆轨迹数据，提出一种“非对称Yukawa交通势能”的模型，该模型能够充分均衡考虑车辆之间的距离、速度、加速度等因素，同时考虑了车辆前方与车辆后方产生的潜在危险，从而实现对车辆间相互作用的风险评估；2) Based on vehicle trajectory data, the present invention proposes a model of "asymmetric Yukawa traffic potential energy", which can fully and balancedly consider factors such as the distance, speed, acceleration, etc. between vehicles, while considering the potential dangers in front of and behind the vehicle, thereby realizing risk assessment of interaction between vehicles;
3)本发明基于车辆轨迹数据以及高精地图数据，引入“高斯车道线风险模型”；该模型通过运用高斯分布来分析车辆相对车道中心的横向偏移、车辆与车道线方向的角度差、车辆尺寸以及车道线类型；该模型不仅考虑了自车的状态，还综合了车道线的类型，与非对称Yukawa交通势能模型相结合，提供一种更贴近实际情况的车辆风险评估方法。3) The present invention introduces a "Gaussian lane risk model" based on vehicle trajectory data and high-precision map data; the model uses Gaussian distribution to analyze the lateral offset of the vehicle relative to the lane center, the angular difference between the vehicle and the lane direction, the vehicle size and the lane type; the model not only takes into account the state of the vehicle, but also integrates the type of lane, and combines with the asymmetric Yukawa traffic potential model to provide a vehicle risk assessment method that is closer to the actual situation.
附图说明BRIEF DESCRIPTION OF THE DRAWINGS
为了更清楚地说明本发明实施例的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，应当理解，以下附图仅示出了本发明的某些实施例，因此不应被看作是对范围的限定，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他相关的附图。In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.
图1为本发明的流程示意图；Fig. 1 is a schematic diagram of the process of the present invention;
图2为高斯车道线风险模型中车辆实际行驶方向与车道线方向的角度差θ示意图；FIG2 is a schematic diagram of the angle difference θ between the actual driving direction of the vehicle and the lane line direction in the Gaussian lane line risk model;
图3为自组织神经网络在智能汽车危险场景生成应用的流程图。FIG3 is a flow chart of the application of self-organizing neural network in the generation of dangerous scenarios for smart cars.
具体实施方式Detailed ways
下面结合附图和实施例对本发明作进一步的详细说明。可以理解的是，此处所描述的具体实施例仅仅用于解释本发明，而非对本发明的限定。另外还需要说明的是，为了便于描述，附图中仅示出了与本发明相关的部分而非全部结构。The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.
参阅图1，本发明提供了一种基于势场的智能汽车交通态势风险认知方法，包括以下步骤：Referring to FIG. 1 , the present invention provides a method for intelligent automobile traffic situation risk recognition based on potential field, comprising the following steps:
步骤一、输入城市地区的自然驾驶数据集；所述自然驾驶数据集包括车辆轨迹序列数据以及高精地图数据，具体如下：Step 1: Input a natural driving dataset in an urban area; the natural driving dataset includes vehicle trajectory sequence data and high-precision map data, as follows:
所述高精地图数据为包含每条路段的车道线类型(区分机动车道与自行车道)和车道分布特征(涵盖左、右车道的存在性)的高精度地图数据信息，以及路段是否邻接交通十字路口等关键道路属性信息。此外，车辆轨迹数据集可通过如中国清华大学智能产业研究院发布的DAIR-V2X车路协同数据集以及美国加州大学发布的interaction数据集等开源数据集进行分析获取。借助这些开源的自然驾驶数据集，本发明旨在分析并提炼车辆行驶中的核心语义信息。这些信息将为提出的非对称Yukawa交通势能模型和高斯车道风险场模型提供关键的参数依据。The high-precision map data is high-precision map data information that includes the lane line type (distinguishing between motor vehicle lanes and bicycle lanes) and lane distribution characteristics (covering the existence of left and right lanes) of each road section, as well as key road attribute information such as whether the road section is adjacent to a traffic intersection. In addition, the vehicle trajectory dataset can be obtained by analyzing open source datasets such as the DAIR-V2X vehicle-road collaboration dataset released by the Institute of Intelligent Industry of Tsinghua University in China and the interaction dataset released by the University of California, USA. With the help of these open source natural driving datasets, the present invention aims to analyze and refine the core semantic information of vehicle driving. This information will provide key parameter basis for the proposed asymmetric Yukawa traffic potential model and Gaussian lane risk field model.
所述车辆轨迹序列数据包括interaction以及国内的DAIR-V2X车路协同数据集。The vehicle trajectory sequence data includes interaction and the domestic DAIR-V2X vehicle-road collaboration dataset.
步骤二、通过车辆轨迹序列数据建立非对称Yukawa交通势能计算主车行驶过程中某一时刻所面临的风险；通过高精地图数据建立高斯车道线风险计算模块计算动态时空车辆的风险，具体如下：Step 2: Use vehicle trajectory sequence data to establish an asymmetric Yukawa traffic potential to calculate the risk faced by the main vehicle at a certain moment during driving; use high-precision map data to establish a Gaussian lane line risk calculation module to calculate the risk of dynamic space-time vehicles, as follows:
1)通过车辆轨迹序列数据建立非对称Yukawa交通势能计算主车行驶过程中某一时刻所面临的风险的具体方法如下：1) The specific method of establishing asymmetric Yukawa traffic potential energy by using vehicle trajectory sequence data to calculate the risk faced by the main vehicle at a certain moment during its driving process is as follows:
Yukawa势在物理学中用作表征原子核内部的强相互作用力，这种势能随着距离的增加而呈指数型衰减，在交通模型中，车辆会根据其他车辆或障碍物的相对位置来调整自己的行为，如公式(1)所示，Yukawa的数学形式也可以被用来描述车辆之间的相互作用。Yukawa potential is used in physics to characterize the strong interaction force inside the nucleus. This potential energy decays exponentially with increasing distance. In traffic models, vehicles adjust their behavior according to the relative positions of other vehicles or obstacles, as shown in formula (1). Yukawa's mathematical form can also be used to describe the interaction between vehicles.
式中，P(d)为两粒子间的势能，依赖于它们之间的距离d；α为衰减参数，它决定了势能随距离的衰减速度；当α值较大时随着距离的增加，势能快速减小；A为强度常数，这个常数决定了势能的强度；Where P(d) is the potential energy between the two particles, which depends on the distance d between them; α is the decay parameter, which determines the decay rate of the potential energy with distance; when the α value is large, the potential energy decreases rapidly with the increase of distance; A is the intensity constant, which determines the intensity of the potential energy;
基于公式(1)来描述主车周围以及障碍物的交互是不够的，原公式Yukawa势能只有主车与其他车辆距离d的影响因子，但复杂的交通环境，速度以及加速度也是描述车辆行为的重要因素；同时，车辆前方与后方所面临的风险应区别对待。因此，对Yukawa势进行优化改良，具体如下：It is not enough to describe the interaction between the main vehicle and obstacles based on formula (1). The original formula Yukawa potential energy only has the influence factor of the distance d between the main vehicle and other vehicles. However, complex traffic environment, speed and acceleration are also important factors in describing vehicle behavior. At the same time, the risks faced by the front and rear of the vehicle should be treated differently. Therefore, the Yukawa potential is optimized and improved as follows:
提出非对称Yukawa交通势能(Asymmetric Yukawa Traffic Potential，AYTP)，来表征主车与周围物体的相互作用关系，公式如(2)、(3)、(4)所示：Asymmetric Yukawa Traffic Potential (AYTP) is proposed to characterize the interaction between the main vehicle and surrounding objects. The formulas are shown in (2), (3), and (4):
公式(2)，表示主车i与车辆j(前向车辆)与车辆k(后向车辆)相加之后的非对称Yukawa交通总势能；Formula (2) represents the total potential energy of asymmetric Yukawa traffic after the main vehicle i, vehicle j (forward vehicle) and vehicle k (backward vehicle) are added;
公式(3)与公式(4)中，常数A和B为缩放因子，决定风险函数的基线强度。Δv是车辆之间的纵向速度差，Δa是车辆之间的纵向加速度差，参数α，α′，β，β′，γ，γ′是衰减系数，用于调节距离、速度和加速度对势能的影响。指数ζ是一个非线性因子，当车辆距离非常近时，增加势能，强调高速接近的车辆对主车的影响。实际应用参数的具体值需要基于自然驾驶数据集中的车辆轨迹序列数据进行调整，非线性因子ζ根据历史数据预估决定其值。In formula (3) and formula (4), constants A and B are scaling factors that determine the baseline strength of the risk function. Δv is the longitudinal velocity difference between vehicles, Δa is the longitudinal acceleration difference between vehicles, and parameters α, α′, β, β′, γ, γ′ are attenuation coefficients used to adjust the effects of distance, speed, and acceleration on potential energy. The exponent ζ is a nonlinear factor that increases the potential energy when the vehicles are very close, emphasizing the effect of vehicles approaching at high speed on the host vehicle. The specific values of the parameters for actual application need to be adjusted based on the vehicle trajectory sequence data in the natural driving dataset, and the nonlinear factor ζ determines its value based on historical data estimation.
2)通过高精地图数据建立高斯车道线风险计算模块计算动态时空车辆的风险的具体方法如下：2) The specific method of establishing a Gaussian lane line risk calculation module through high-precision map data to calculate the risk of dynamic space-time vehicles is as follows:
本发明提出一种高斯车道线风险模型，如公式(5)所示，该模型利用高斯分布来量化和分析车辆相对于车道中心的横向偏移量(Δdl)、车辆航向和车道方向的角度差(θ)与车辆的尺寸(宽度w和长度l)以及车道线类型(type)，从而得出在特定的车辆参数和静态道路条件下，车辆行驶的风险概率。The present invention proposes a Gaussian lane line risk model, as shown in formula (5). The model uses Gaussian distribution to quantify and analyze the lateral offset of the vehicle relative to the lane center (Δdl ), the angular difference (θ) between the vehicle heading and the lane direction, the vehicle size (width w and length l), and the lane line type (type), so as to obtain the risk probability of vehicle driving under specific vehicle parameters and static road conditions.
公式(5)中引入的权重μ旨在量化不同车道线类型对车辆行为的约束强度。该权重根据车道线的不同类型进一步细分为四个类别：实线(solid)、虚线(dashed)、双实线(double_solid)、车道物理边界(barrier)。每种类型的车道线被赋予不同的权重，从而反映出不同类型车道线对车辆行为的影响程度。权重分配遵循一个明确的顺序：车道物理边界获得最高的权重，其次是双实线，再次是实线，而虚线的权重最低。The weight μ introduced in formula (5) aims to quantify the constraint strength of different lane line types on vehicle behavior. The weight is further subdivided into four categories according to the different types of lane lines: solid, dashed, double solid, and lane physical boundary (barrier). Each type of lane line is assigned a different weight, reflecting the degree of influence of different types of lane lines on vehicle behavior. The weight distribution follows a clear order: the lane physical boundary gets the highest weight, followed by the double solid line, then the solid line, and the dashed line has the lowest weight.
所述车道物理边界通常包括栅栏、树木或花坛等障碍物，与之发生接触往往会导致交通事故。双实线作为分隔不同行驶方向的界线，穿越它们可能会显著增加与对向车辆发生事故的风险。实线通常在交通规则中标示着变道限制，违反这一规则可能会引起与其他车辆的碰撞。相比之下，虚线表示允许变道，因此发生交通事故风险的概率相对较低。The physical boundaries of the lanes usually include obstacles such as fences, trees or flower beds, and contact with them often leads to traffic accidents. Double solid lines are used as boundaries to separate different directions of travel, and crossing them may significantly increase the risk of accidents with oncoming vehicles. Solid lines usually indicate lane change restrictions in traffic regulations, and violating this rule may cause collisions with other vehicles. In contrast, dotted lines indicate that lane changes are allowed, so the probability of traffic accident risks is relatively low.
Δdl是量化车辆横向偏移的指标，表示车辆中心点到最近车道线点的距离。随着Δdl的增加，表明车辆与车道中心线的距离越来越远，反映车辆正在向车道线方向偏移。在车辆横向位置偏差较大的情况下，意味着车辆与车道中心的距离减少，表明车辆更接近于车道的物理边界或可能是对向车流的分隔线等(双实线、实现、虚线)。因此，在风险评估模型中，Δdl的值越大，代表车辆可能因为偏离安全行驶轨迹而产生的风险越高。Δdl is an indicator that quantifies the lateral deviation of the vehicle, indicating the distance from the center point of the vehicle to the nearest lane line point. As Δdl increases, it indicates that the distance between the vehicle and the center line of the lane is getting farther and farther, reflecting that the vehicle is deviating in the direction of the lane line. In the case of a large lateral position deviation of the vehicle, it means that the distance between the vehicle and the center of the lane is reduced, indicating that the vehicle is closer to the physical boundary of the lane or may be the separator of the oncoming traffic, etc. (double solid line, real line, dotted line). Therefore, in the risk assessment model, the larger the value of Δdl , the higher the risk that the vehicle may have due to deviation from the safe driving trajectory.
标准差σ决定高斯分布宽度的参数，它将影响关于车辆横向偏移Δdl的风险概率分布的形状。The standard deviation σ determines the parameter of the Gaussian distribution width, which will affect the shape of the risk probability distribution with respect to the vehicle lateral displacement Δdl .
λ＝arctan2(Δlat，Δlon) (6)λ＝arctan2(Δlat，Δlon) (6)
θ＝φ-λ (7)θ＝φ-λ (7)
h(θ，w，l)＝ecos(θ)·Vsize (8)h(θ,w,l)＝ecos(θ) ·Vsize (8)
在本发明中，定义车辆航向角和其他相关角度均基于笛卡尔坐标系，且高程变化对计算结果的影响不予考虑；如附图2所示，通过公式(6)与(7)计算得到的角度θ(车辆实际行驶方向与车道线方向的角度差)，判定车辆行驶方向与车道线平行还是偏离车辆原本行驶方向θ，详细的计算步骤按照以下方式展开：In the present invention, the vehicle heading angle and other related angles are defined based on the Cartesian coordinate system, and the influence of elevation change on the calculation result is not considered; as shown in FIG2, the angle θ (the angle difference between the actual driving direction of the vehicle and the direction of the lane line) calculated by formulas (6) and (7) is used to determine whether the vehicle driving direction is parallel to the lane line or deviates from the original driving direction of the vehicle θ. The detailed calculation steps are as follows:
在高精地图上，车车道线被定义为一系列经纬度坐标点的连续序列；因此，在任意给定时刻t中，可通过当前车辆所在车道位置距离最近车道线上的两个连续的经纬度坐标P1(lat1，lon1)以及P2(lat2，lon2)，计算两点的精度差(Δlon＝lon2-lon1)和纬度差(Δlat＝lat2-lat1)，使用反正切函数arctan2(Δlat，Δlon)确定P1至P2的方向角λ，车辆的实际航向可以从自然驾驶数据中获取，所需计算的角度θ则定义为车辆航向/>与由两个连续经纬度坐标P1(lat1，lon1)和P2(lat2，lon2)所定义的微小车道线方向λ之间的角度差。On a high-precision map, a lane line is defined as a continuous sequence of longitude and latitude coordinates. Therefore, at any given time t, the two consecutive longitude and latitude coordinatesP1 (lat1 ,lon1 ) andP2 (lat2 ,lon2 ) on the nearest lane line from the current vehicle's lane position can be used to calculate the accuracy difference (Δlon=lon2 -lon1 ) and latitude difference (Δlat=lat2 -lat1 ) between the two points, and the inverse tangent function arctan2(Δlat, Δlon) can be used to determine the direction angle λ fromP1 toP2 , and the actual heading of the vehicle. It can be obtained from natural driving data, and the angle θ to be calculated is defined as the vehicle heading/> The angular difference between the direction λ of the small lane line defined by two consecutive longitude and latitude coordinates P1 (lat1 , lon1 ) and P2 (lat2 , lon2 ).
在本发明中，定义的h(θ，w，l)函数，被称为“车道几何影响因子函数”，如公式(8)所示；该函数综合对其程度以及车辆尺寸对行驶风险的影响；函数中ecos(θ)成分用于量化车辆航向与车道线方向之间的对齐程度；当θ(车辆实际行驶方向与车道线方向的角度差)接近零时，cos(θ)的值接近1，表明车辆正沿车道方向精确行驶，没有变道的趋势；在此情况下，风险权重达到最低，表明车辆处于理想的行驶状态，相应的风险水平也相对较低。In the present invention, the h(θ, w, l) function is defined, which is called the "lane geometry influence factor function", as shown in formula (8); the function comprehensively considers the influence of its degree and vehicle size on driving risk; the ecos(θ) component in the function is used to quantify the degree of alignment between the vehicle heading and the lane line direction; when θ (the angular difference between the actual driving direction of the vehicle and the lane line direction) is close to zero, the value of cos(θ) is close to 1, indicating that the vehicle is driving accurately along the lane direction and has no tendency to change lanes; in this case, the risk weight reaches the lowest, indicating that the vehicle is in an ideal driving state and the corresponding risk level is relatively low.
公式(9)Vsize模块为公式(8)组成的一部分，定义为“尺寸调控因子”，其中w·l是车辆的尺寸，它的乘积给出了车辆的“面积”或者占地面积；尽管车辆尺寸在一定程度上影响行驶安全，但这种影响是有限的，特别是对于大型车辆，例如卡车，由于其体积庞大，在道路上行驶时本质上已经构成了额外的风险因素；因此，风险并不是随着车辆尺寸的增大而无限增加的，为了在模型中体现这一点，特意采用了对数函数作为调节机制；具体来说，模型对车辆尺寸w·l进行了对数转换，即log(w·l)，表明车辆尺寸的每一次倍数增加对Vsize的影响是递减的，防止w·l非常大时Vsize增长过快；同时，车辆是否越界或可能发生的变道行为是比车辆尺寸更关键的因素，因此车辆行驶的方向与车道线方向的一致性以及潜在的变道倾向，应当赋予更高的权重。Formula (9) The Vsize module is part of Formula (8) and is defined as the "size control factor", where w·l is the size of the vehicle, and its product gives the "area" or footprint of the vehicle. Although vehicle size affects driving safety to a certain extent, this effect is limited, especially for large vehicles such as trucks, which inherently constitute an additional risk factor when driving on the road due to their large size. Therefore, the risk does not increase infinitely with the increase of vehicle size. In order to reflect this in the model, a logarithmic function is deliberately used as a control mechanism. Specifically, the model performs a logarithmic transformation on the vehicle size w·l, that is, log(w·l) , indicating that each multiple increase in vehicle size has a decreasing effect on Vsize , preventing Vsize from growing too fast when w·l is very large. At the same time, whether the vehicle crosses the boundary or the possible lane change behavior is a more critical factor than the vehicle size, so the consistency of the vehicle's driving direction with the lane line direction and the potential lane change tendency should be given a higher weight.
c常数，为一个比例因子，直接影响车辆尺寸变量w·l经过对数变换后对Vsize的贡献程度，c因子的选定须考虑到实际车辆尺寸对安全风险的实际影响；若观察到较大的车辆在特定行驶条件下的风险增加幅度不如较小车辆显著，c可以相应调低；相反，如果大型车辆带来的额外风险相对较大，c的值将相应增大。The c constant is a proportionality factor that directly affects the contribution of the vehicle size variable w·l to Vsize after logarithmic transformation. The selection of the c factor must take into account the actual impact of the actual vehicle size on the safety risk. If it is observed that the risk increase of larger vehicles under specific driving conditions is not as significant as that of smaller vehicles, c can be lowered accordingly. On the contrary, if the additional risk posed by large vehicles is relatively large, the value of c will be increased accordingly.
b作为分母中的常数，提供Vsize函数的基础值，它相当于是一个归一化因子，确保即使在w·l的值较小(即车辆尺寸较小)时，Vsize也保持在一个合理的范围内。b, as a constant in the denominator, provides the base value of the Vsize function. It is equivalent to a normalization factor, ensuring that Vsize remains within a reasonable range even when the value of w·l is small (i.e., the vehicle size is small).
步骤三、对自然驾驶数据指定片段中主车行驶过程中某一时刻所面临的非对称Yukawa交通势能模块计算的风险值以及高斯车道线风险模块计算的风险值进行累加，累加的结果为车辆交通态势风险认知总势场，用此值评估当前自然驾驶数据集指定车辆轨迹片段的车辆行车风险。Step 3: Accumulate the risk value calculated by the asymmetric Yukawa traffic potential module and the risk value calculated by the Gaussian lane line risk module at a certain moment in the specified segment of the natural driving data. The accumulated result is the total potential field of vehicle traffic situation risk cognition, which is used to evaluate the vehicle driving risk of the specified vehicle trajectory segment of the current natural driving data set.
进一步的，公式(10)中，通过时刻t车辆的非对称Yukawa总势能与该时刻的高斯车道风险势能，共同构建主导车辆行驶过程中在某一时刻下的综合风险势场模型，从而最终评估处当前时刻的车辆行车风险，进而建立附有风险值标签的车辆风险值数据集。所计算的风险值反映了自然驾驶数据集中，车辆轨迹数据在特定自然驾驶数据集中指定片段的潜在风险水平，提供了一个量化指标来评估车辆在任意给定时段可能遭遇的危险程度。通过遍历选定的自然驾驶数据集，构建带有风险值标签的车辆风险值数据集，以便提供给步骤四的无监督聚类算法进行数据的进一步处理。Furthermore, in formula (10), the asymmetric Yukawa total potential energy of the vehicle at time t and the Gaussian lane risk potential energy at that time are used to jointly construct a comprehensive risk potential field model of the dominant vehicle at a certain moment in the driving process, so as to finally evaluate the driving risk of the vehicle at the current moment, and then establish a vehicle risk value data set with risk value labels. The calculated risk value reflects the potential risk level of the vehicle trajectory data in the natural driving data set in a specified segment in a specific natural driving data set, and provides a quantitative indicator to evaluate the degree of danger that the vehicle may encounter in any given period of time. By traversing the selected natural driving data set, a vehicle risk value data set with risk value labels is constructed to provide it to the unsupervised clustering algorithm in step 4 for further data processing.
步骤四、采用了自组织神经网络SOM的无监督聚类算法，对轨迹片段相关的风险值进行细分。如图3所示，从步骤三得出的带有风险值标签的车辆风险值数据集至作为输入，传递至SOM(自组织神经网络)的竞争层，该层的权值W初始化为二维风险值的随机参考向量。图3中描绘了SOM学习过程的迭代可视化，反映了自组织神经网络通过不断学习以适应风险数据分布的过程。依据PEGASUS项目规范、ISO 34502国际标准及行业专家意见，本方法通过设定阈值明确无监督聚类结果，以有效区分风险类别中的危险场景。鉴于无监督聚类技术的广泛应用和成熟度，本发明将不深入阐述自组织神经网络SOM的技术细节。Step 4: The unsupervised clustering algorithm of the self-organizing neural network SOM is used to subdivide the risk values associated with the trajectory segments. As shown in Figure 3, the vehicle risk value dataset with risk value labels obtained from step 3 to As input, it is passed to the competition layer of SOM (self-organizing neural network), and the weight W of this layer is initialized as a random reference vector of two-dimensional risk values. Figure 3 depicts the iterative visualization of the SOM learning process, reflecting the process of the self-organizing neural network adapting to the distribution of risk data through continuous learning. According to the PEGASUS project specifications, ISO 34502 international standards and industry expert opinions, this method clarifies the unsupervised clustering results by setting thresholds to effectively distinguish dangerous scenarios in risk categories. In view of the wide application and maturity of unsupervised clustering technology, the present invention will not elaborate on the technical details of the self-organizing neural network SOM.
综上，本发明针对智能汽车测试中生成关键危险场景“难”的重要挑战，提出了一种创新性的方法。此方法融合了非对称Yukawa交通势能模型与高斯车道线风险模型，旨在精准评估主车在特定时刻遭遇的交通风险。其次，本专利采用了无监督聚类的机器学习技术，基于计算得出的交通风险指数，有效筛选出对智能汽车安全性测试至关重要的关键危险场景。本方法的创新性在于提出了一种改良的汤川势能(Yukawa)模型-非对称Yukawa交通势能，该模型将物理学概念与智能汽车技术相融合，并由此开发出一种创新的智能汽车关键危险测试场景生成方法。In summary, the present invention proposes an innovative method to address the important challenge of "difficulty" in generating key danger scenarios in smart car testing. This method combines the asymmetric Yukawa traffic potential model with the Gaussian lane line risk model, aiming to accurately assess the traffic risks encountered by the main vehicle at a specific moment. Secondly, this patent uses unsupervised clustering machine learning technology to effectively screen out key danger scenarios that are crucial to the safety testing of smart cars based on the calculated traffic risk index. The innovation of this method lies in the proposal of a modified Yukawa potential model-asymmetric Yukawa traffic potential, which integrates physical concepts with smart car technology, and thereby develops an innovative method for generating key danger test scenarios for smart cars.
尽管已经示出和描述了本发明的实施例，对于本领域的普通技术人员而言，可以理解在不脱离本发明的原理和精神的情况下可以对这些实施例进行多种变化、修改、替换和变型，本发明的范围由所附权利要求及其等同物限定。Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410081136.3ACN117892178A (en) | 2024-01-19 | 2024-01-19 | Scene extraction method based on intelligent automobile traffic situation risk cognition |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410081136.3ACN117892178A (en) | 2024-01-19 | 2024-01-19 | Scene extraction method based on intelligent automobile traffic situation risk cognition |
| Publication Number | Publication Date |
|---|---|
| CN117892178Atrue CN117892178A (en) | 2024-04-16 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410081136.3APendingCN117892178A (en) | 2024-01-19 | 2024-01-19 | Scene extraction method based on intelligent automobile traffic situation risk cognition |
| Country | Link |
|---|---|
| CN (1) | CN117892178A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118658304A (en)* | 2024-08-19 | 2024-09-17 | 青岛理工大学 | Position prediction and risk quantification method for connected autonomous vehicles based on velocity field |
| CN119190019A (en)* | 2024-10-09 | 2024-12-27 | 长沙理工大学 | A vehicle lane-changing decision method and system for mixed passenger and freight traffic scenarios |
| CN119636790A (en)* | 2025-02-20 | 2025-03-18 | 陕西永嘉泰乐工程技术有限公司 | Road traffic behavior anomaly detection method and system based on computer vision |
| CN119714932A (en)* | 2024-12-20 | 2025-03-28 | 江苏汉邦车业有限公司 | A brake detection method for electric tricycle |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118658304A (en)* | 2024-08-19 | 2024-09-17 | 青岛理工大学 | Position prediction and risk quantification method for connected autonomous vehicles based on velocity field |
| CN119190019A (en)* | 2024-10-09 | 2024-12-27 | 长沙理工大学 | A vehicle lane-changing decision method and system for mixed passenger and freight traffic scenarios |
| CN119714932A (en)* | 2024-12-20 | 2025-03-28 | 江苏汉邦车业有限公司 | A brake detection method for electric tricycle |
| CN119636790A (en)* | 2025-02-20 | 2025-03-18 | 陕西永嘉泰乐工程技术有限公司 | Road traffic behavior anomaly detection method and system based on computer vision |
| Publication | Publication Date | Title |
|---|---|---|
| CN111580500B (en) | Evaluation method for safety of automatic driving automobile | |
| CN117892178A (en) | Scene extraction method based on intelligent automobile traffic situation risk cognition | |
| CN110949398B (en) | Method for detecting abnormal driving behavior of first-vehicle drivers in vehicle formation driving | |
| Jin et al. | Gauss mixture hidden Markov model to characterise and model discretionary lane‐change behaviours for autonomous vehicles | |
| CN107492251A (en) | It is a kind of to be identified and driving condition supervision method based on the driver identity of machine learning and deep learning | |
| CN106599827A (en) | Small target rapid detection method based on deep convolution neural network | |
| CN113642108B (en) | A method for generating key test cases for autonomous vehicle intersection traffic scenarios | |
| CN103235933A (en) | Vehicle abnormal behavior detection method based on Hidden Markov Model | |
| CN114743181A (en) | A road vehicle target detection method, system, electronic device and storage medium | |
| CN111950583A (en) | A method for multi-scale identification of traffic signal signs based on GMM clustering | |
| CN112084890A (en) | Multi-scale traffic signal sign identification method based on GMM and CQFL | |
| CN108960074B (en) | Small-size pedestrian target detection method based on deep learning | |
| CN118238847B (en) | Autonomous lane change decision planning method and system adaptive to different driving styles and road surface environments | |
| CN113033363A (en) | Vehicle dense target detection method based on deep learning | |
| Wang et al. | RETRACTED ARTICLE: Intelligent traffic monitoring and traffic diagnosis analysis based on neural network algorithm | |
| Li et al. | Modeling and analysis of mandatory lane-changing behavior considering heterogeneity in means and variances | |
| Tan et al. | RCP‐RF: A comprehensive road‐car‐pedestrian risk management framework based on driving risk potential field | |
| CN116361175A (en) | Method for creating test scenes of automatic driving vehicles in different safety domains | |
| Hamad et al. | How Do drivers behave at roundabouts in a mixed traffic? A case study using machine learning | |
| CN119261915A (en) | A vehicle risk potential field assessment method based on driving style coefficient | |
| Wang et al. | Research on road adhesion condition identification based on an improved ALexNet model | |
| CN118025185A (en) | Driving intention recognition method based on Gaussian mixture-hidden Markov model | |
| CN112328651A (en) | Traffic target recognition method based on statistical features of millimeter wave radar data | |
| CN117237689A (en) | Machine learning-based vehicle lane change intention recognition method and system | |
| CN114387777B (en) | Road data processing method, device, computer equipment and storage medium |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |