ArticlePDF Available

Comparison of Markov Chain Abstraction and Monte Carlo Simulation for the Safety Assessment of Autonomous Cars

Authors:
Matthias Althoff at Technische Universität München
Matthias Althoff
Alexander Mergel
Alexander Mergel
Alexander Mergel
  • This person is not on ResearchGate, or hasn't claimed this research yet.
Download full-text PDFDownload full-text PDF
Read full-text
Download citation

Abstract and Figures

The probabilistic prediction of road traffic scenarios is addressed. One result is a probabilistic occupancy of traffic participants, and the other result is the collision risk for autonomous vehicles when executing a planned maneuver. The probabilistic occupancy of surrounding traffic participants helps to plan the maneuver of an autonomous vehicle, whereas the computed collision risk helps to decide if a planned maneuver should be executed. Two methods for the probabilistic prediction are presented and compared: 1) Markov chain abstraction and 2) Monte Carlo simulation. The performance of both methods is evaluated with respect to the prediction of the probabilistic occupancy and the collision risk. For each comparison test, we use the same models that generate the probabilistic behavior of traffic participants, where the generation of these data is not compared with real-world data. However, the results independently show the behavior generation that Markov chains are preferred for the probabilistic occupancy, whereas Monte Carlo simulation is clearly preferred for determining the collision risk.
Figures - uploaded by Matthias Althoff
Author content
All figure content in this area was uploaded by Matthias Althoff
Content may be subject to copyright.
althoff-2011-ITSStra
ns.pdf
Content uploaded by Matthias Althoff
Author content
All content in this area was uploaded by Matthias Althoff
Content may be subject to copyright.
althoff-2011-ITSStra
ns.pdf
Content uploaded by Matthias Althoff
Author content
All content in this area was uploaded by Matthias Althoff
Content may be subject to copyright.
A preview of the PDF is not available
... Analytical approaches on autonomous driving (AD) system or functional level such as Failure mode and effects analysis (FMEA), Fault Tree Analysis (FTA) and advanced Markov simulation require as input failure rate data, potentially depending on scenario data, time and mission history. However, such data on the failure rate is only available to a very limited extent in the open scientific literature (Richter et al. 2023 (Althoff and Mergel 2011) (Nyberg 2018) (Kaalen, Nyberg, and Bondesson 2019) (Häring et al. 2022) . ...
ACCEPTED PAPER: Failure rates per time for autonomous driving safety assessment from CARLA simulation
Conference Paper
  • Feb 2024
Autonomous Driving simulations like CARLA offer a variety of functionalities including the modeling of subfunctions of autonomous driving such as detection, classification, identification, and tracking of persons, objects and road elements, route planning, and maneuver planning. All of these functionalities operate at high resolution level regarding scenario details such as conditions of weather, daytime, road type, traffic conditions, etc. In contrast, analytical safety assessments like functional Failure Mode and Effects Analysis (FMEA), Fault Tree Analysis (FTA) or Markov modelling and simulation operate at an abstract level in terms of failures per time or conditional failure rates such as failure of object detection given certain weather or daytime conditions. Furthermore, simulation options can be varied over a wide range of parameters and scenario types, including statistical generation of scenarios and active road participants, i.e. spawning of objects. The article presents a quantitative approach to generate failure rates per hour from CARLA simulation for object detection under a range of environmental conditions in terms of precipitation intensity, fog density, and time of the day, single and in combination. Focus is on the derivation of failure rate expressions that are accessible from simulation data. To this end computer vision metrics are used together with additional information available within simulation setup to compute failure rates per hour. Results are presented using sample tables, box plot and violin graphs. The CARLA simulator is used to assess an object detection algorithm that has been fine-tuned with CARLA sample images. It is discussed why the obtained failure rates are consistent but rather high. Further improvement options of the overall approach are provided.
View
... Reliable trajectory prediction is essential for the ego vehicle to comprehend the surrounding traffic environment, and extensive research has been conducted on this regard [8]- [10]. Existing trajectory prediction methods can be roughly categorized into four groups: physics-based [11], Gaussian processes-based [12], model-based [13], and DL-based approaches [14]. The first three categories formulate trajectory prediction as a deterministic regression problem. ...
An Enabling Decision-Making Scheme by Considering Trajectory Prediction and Motion Uncertainty
Article
  • Jan 2024
Decision-making plays a crucial role in enabling automated driving and its performance is highly dependent on accurate motion prediction. Most existing motion prediction methods employ black-box neural network approaches, which may introduce high uncertainty and thus compromise driving safety. This study proposes a decision-making framework that takes motion prediction into account while considering predicted motion uncertainty. First, a multi-modal trajectory prediction method is introduced by combining the motion trend of the ego vehicle to reduce trajectory uncertainty. Then a two-dimensional dynamic risk assessment approach based on the Gaussian distribution is developed to quantify potential collision risk by analyzing occupancy relationship in space-time. Additionally, elementary motion patterns of automated driving in dynamic traffic environments are analyzed, and a decision-making method integrating finite state machines and the Markov decision process is established. Taking into account future motion uncertainty, it can dynamically adjust the ego vehicle's driving states in advance to ensuring driving safety. Finally, the effectiveness of the proposed decision-making framework is verified under different driving scenarios extracted from the NGSIM dataset, demonstrating satisfying results in avoiding potential driving risks.
View
... Study [34] proposed a trajectory prediction model based on constant velocity and acceleration by assuming that the predicted temporal vehicle motion state remains unchanged, but this method does not consider the varying characteristics of vehicle lateral dynamics in the prediction time domain and is not applicable to conditions such as vehicle turning and lane changing. Study [35] studied the variation characteristics of vehicle yaw rate within the prediction time domain based on the above model, which efectively characterizes the vehicle's lateral position changes and has high online calculation efciency. However, the above method assumes that most vehicle states remain unchanged within the prediction time domain, ignoring their uncertainty, and thus only applies to short-term prediction. ...
TCN-SA: A Social Attention Network Based on Temporal Convolutional Network for Vehicle Trajectory Prediction
Article
Full-text available
Vehicle trajectory prediction can provide important support for intelligent transportation systems in areas such as autonomous driving, traffic control, and traffic flow optimization. Predicting vehicle trajectories is an extremely challenging task that not only depends on the vehicle’s historical trajectory but also on the dynamic and complex social-temporal relationships of the surrounding traffic network. The trajectory of the target vehicle is influenced by surrounding vehicles. However, existing methods have shortcomings in considering both time dependency and interactive dependency between vehicles or insufficient consideration of the impact of surrounding vehicles. To address this issue, we propose a hybrid deep learning model based on a temporal convolutional network (TCN) that considers local and global interactions between vehicles. Specifically, we use a social convolutional pooling layer to capture local interaction features between vehicles and a multihead self-attention layer to capture global interaction features between vehicles. Finally, we combine these two features using an encoder-decoder structure to predict vehicle trajectories. Through experiments on the Next-Generation Simulation (NGSIM) public dataset and ablation experiments, we validate the effectiveness of our model.
View
... For example, the authors of [52] use simplifications based on contextual reasoning, while [55] uses Markov chain abstractions to compute the crash probability in a stochastic reachable set framework. It is also common to estimate the probabilities by Monte Carlo simulation, see for instance [56], [57]. Hence, probability-based strategies require good estimates of the predictive distributions, which is contrary to the coarse uncertainty estimates provided by ANNbased uncertainty-aware models, as discussed in the previous section. ...
Prediction-Uncertainty-Aware Threat Detection for ADAS: A Case Study on Lane-Keeping Assistance
Preprint
Full-text available
  • Jan 2023
Advanced driver assistance systems typically support the driver in cases where the driver is likely to fail the driving task. The challenge, from a system perspective, is to accurately detect those cases. Recently, machine learning-based prediction models that are able to estimate the prediction uncertainty in real-time have successfully been introduced for this purpose. However, very little effort has been made on using the prediction uncertainty in the decision-making logic to improve the system's robustness, especially in cases where the input data is affected by noise or anomalies that are not presented in the training data. In this work, four threat-detection methods using uncertainty estimates are proposed and evaluated using a real-world data set. The methods use different strategies for leveraging uncertainty information, where the goal is to ensure that the intervention decision is based on trustworthy predictions. The threat-detection methods' performances are evaluated, using five different learning-based prediction models, in the context of a lane-keeping assistance application.
View
... Existing vehicle trajectory prediction methods can be divided into four categories [5]: physics-based methods, classic machine learning-based methods, deep learning-based methods, and reinforcement learning-based methods. Earlier, scholars mostly predicted the future motion state of vehicles based on physical rules, and such methods mainly include the Constant-Velocity (CV) model [6], Constant-Accelerate (CA) model [7] and Kalman Filtering (KF) based methods [8], Monte Carlo (MC) based methods [9]. However, this category is only applicable to short-term trajectory prediction with a prediction horizon of no more than 1s, and cannot effectively complete the medium-and long-term prediction tasks. ...
Vehicle Trajectory Prediction Based on Multivariate Interaction Modeling
Article
Full-text available
  • Jan 2023
Vehicle trajectory prediction is crucial for the safe driving of the intelligent and connected vehicle, but the existing researches suffer from inadequate interaction modeling, especially lacking modeling for dynamic interaction. To solve this problem, we propose a vehicle trajectory prediction model SDT-ATT based on multivariate interaction modeling. Firstly, a hierarchical extracting module is constructed based on the bi-directional long short-term memory network (Bi-LSTM) to extract the contexts of the target vehicle parameters, neighbor vehicle spatial parameters and neighbor vehicle dynamic parameters, respectively. Secondly, the temporal, spatial and dynamic interactions of vehicles are modeled and learned based on the multi-head attention (MHA) mechanism, then fused to obtain the multivariate interaction features. Next, a direct multi-step prediction module is constructed based on long short-term memory network (LSTM), which directly outputs the future coordinates of the vehicle in the target prediction horizons. Finally, the vehicle trajectory prediction model SDT-ATT proposed in this paper is validated based on the public naturalistic driving datasets I-80 and US-101. The data results show that compared with the CS-LSTM model, the SDT-ATT has 9.3% lower ADE average value, 8.5% lower FDE average value and 8.7% lower RMSE average value, and the inference speed of the SDT-ATT model meets the real-time requirement.
View
... Following Lefèvre et al. (2014), vehicle trajectory prediction methods can be divided into physics-based methods, maneuver-based methods, and interaction-aware methods. Physics-based methods (Treiber et al., 2000;Yu et al., 2018;Sharath and Velaga, 2020;Mo et al., 2021) represent vehicles as entities controlled by physics laws and characterize the dynamics and kinematics of the vehicles by different uncertainty handling methods (Veeraraghavan et al., 2006;Gindele et al., 2010;Althoff and Mergel, 2011). They are restricted to short-term motion prediction and highway scenarios because of the limitation of input features and model parameters (Messaoud et al., 2021). ...
Dynamic-learning spatial-temporal Transformer network for vehicular trajectory prediction at urban intersections
Article
  • Nov 2023
  • TRANSPORT RES C-EMER
Forecasting vehicles' future motion is crucial for real-world applications such as the navigation of autonomous vehicles and feasibility of safety systems based on the Internet of Vehicles (IoV). Vehicular trajectory prediction at urban intersections remains challenging due to the difficulty in modeling temporal dependencies and spatial interactions among traffic agents. This paper proposes a dynamic-learning spatial-temporal Transformer network (DSTTN) with domain adaptation training methods based on two modules for two-dimensional vehicular trajectory prediction at urban intersections. The first module is trajectory maneuver characterization (TMC) which captures latent driving maneuver features and divides trajectory data into different categories with different distributions, which can be regarded as different domains in transfer learning (TL). The second module is trajectory distribution matching (TDM) which adopts a novel spatial-temporal Transformer network with distribution matching loss to dynamically learn domain-invariant maneuvers and achieve accurate trajectory prediction. Experiments and ablation studies collected at two unsignalized urban intersections of the inD dataset first validate the interpretability, transferability, and high prediction accuracy of the proposed DSTTN. The results show that the TMC module has maneuver-representational ability. The proposed STTN achieves good and stable prediction accuracy compared with other baselines, including the state-of-the-art deep learning model, while DSTTN with the TMC and TDM modules further improves accuracy. Additional experiments have been conducted under 70 randomly-selected urban intersection scenarios of the Waymo dataset to validate the good prediction accuracy of the proposed method. The formulation of DSTTN offers new ideas for dividing trajectory data into different domains and general domain adaptation to vehicle trajectory prediction under urban road environments.
View
View
View
Goal-Oriented Pedestrian Motion Prediction
Article
  • Jan 2023
Forecasting the motion of others in shared spaces is a key for intelligent agents to operate safely and smoothly. We present an approach for probabilistic prediction of pedestrian motion incorporating various context cues. Our approach is based on goal-oriented prediction, yielding interpretable results for the predicted pedestrian intention, even without the prior knowledge of goal positions. By using Markov chains, the resulting probability distribution is deterministic—a beneficial property for motion planning or risk assessment in automated and assisted driving. Our approach outperforms a physics-based approach and improves over state-of-the-art approaches by reducing standard deviations of prediction errors and improving robustness against realistic, noisy measurements.
View
Driver Intention & Interaction-Aware Trajectory Forecasting via Modular Multi-Task Learning
Article
Full-text available
  • Oct 2023
Forecasting the trajectories of vehicles in a highway scene is a crucial task in the area of Intelligent Transportation Systems (ITS), as the complexities of a web of driving maneuvers in a highway setting can easily lead to a catastrophe. An effective trajectory prediction thus should take into account first, what the driver wants to do (intention), as well as second, what the surrounding vehicles are going to do (interaction). This, thus, is the goal of this article. We have adopted intention awareness on two data sources and have also implemented the multi-head attention mechanism to achieve interaction awareness in order to achieve accurate future trajectory predictions. The proposed method has been trained and evaluated by following a modular multi-task learning method on two distinct publicly available datasets -NGSIM and Brain4Cars and additionally has been implemented on CARLA for evaluation of results. Experimental outcomes verify that our approach outperforms other state-of-the-art methods.
View
Integrated driving behavior modeling
Article
Full-text available
  • Oct 2005
View
Evaluation of automotive forward collision warning and collision avoidance algorithms
Article
Full-text available
  • Nov 2005
  • VEHICLE SYST DYN
Collision warning/collision avoidance (CW/CA) systems target a major crash type and their devel-opment is a major thrust of the Intelligent Vehicle Initiative. They are a natural extension of adaptive cruise control systems already available on many car models. Many CW/CA algorithms have recently been proposed but the existing literature mainly focuses on algorithm development. Evaluations of these algorithms have been usually based on subjective ratings. The main contribution of this paper is the utilization of a naturalistic driving data set for the evaluation of CW/CA algorithms. We first collect manual driving data from the ICCFOT project, then process the data by Kalman smoothing, and finally identify 'threatening' and 'safe' data sets according to vehicle brake inputs and vehicle range behavior. Five CW/CA algorithms published in the literature are evaluated against the identified data sets. The performance of these algorithms is determined through a performance metric commonly used in signal detection and information retrieval under unbalanced data population.
View
View
Design and evaluation of forward collision warning/collision avoidance algorithms
Article
  • Apr 2005
Collision Warning/Collision Avoidance (CW/CA) systems are actively studied by many automotive companies, with the goal of reducing the frequency and severity of rear end collisions, which are responsible for about one third of all ground vehicle crashes. The majority of existing CW/CA literature focused on the algorithm designs, which are based on widely different philosophies and strategies. Consequently, their performance and characteristics are also quite different. Evaluations of these algorithms were usually subjective and were done based on the feedback of a limited number of drivers. False alarms were found to be a common problem mainly because of the fact the test subjects drive quite differently themselves. In this paper, we utilize a large-scale human driving database for the evaluation, optimization and design of CW/CA algorithms and introduces two important concepts: a "scalable" numerical optimization algorithm for unbalanced data sets, and the concept of the Electronic Brake Light, which was found to significantly improve the performance of CW/CA algorithms. © 2005 Society of Automotive Engineers of Japan, Inc. All rights reserved.
View
View
Free flight collision risk estimation by sequential Monte Carlo simulation
Chapter
  • Jan 2007
Technology allows aircraft to broadcast information about its own-ship position and velocity to surrounding aircraft, and to receive similar information from surrounding aircraft. This development has stimulated the rethinking of the overall concept for future Air Traffic Management (ATM), e.g., to transfer responsibility for conflict prevention from ground to air. As the aircrews thus obtain the freedom to select their trajectory, this conceptual idea is called free flight [57]. It changes ATM in such a fundamental way, that one could speak of a paradigm shift: centralised control becomes distributed, responsibilities transfer from ground to air, fixed air traffic routes are removed, and appropriate new technologies are brought in. Each aircrew has the responsibility to timely detect and solve conflicts, thereby assisted by navigation means, surveillance processing, and conflict resolution systems. Due to the potentially many aircraft involved, the system is highly distributed. This free flight concept idea has motivated the study of multiple operational concepts and implementation choices [33], [37], [41], [44], [54]. One of the key outstanding issues is the safety verification of free flight design, and in particular when air traffic demand is high.
View
A Stochastic Form of a Human Driver Steering Dynamics Model
Article
  • May 2007
This work develops a stochastic form of a human driver model which can be used for simulating vehicle guidance and control. The human motor-control function is complex and can be affected by factors such as driver's training and experience, fatigue, road conditions, and attention. The variations in these effects become more pronounced in hazardous driving conditions such as in snow and ice. One example of such driving conditions is snow removal operation in highway maintenance, where the use of a stochastic driver model seems to be more desirable. This work evaluates and extends existing models of a human driver including stochastic or statistical considerations related to differences in drivers' experiences and their conditions as well as variations in the effect of disturbances such as plowing forces. The aim is to develop a simulation environment that can be used in design and evaluation of driver assistance systems for snow removal operation in an Intelligent Transportation System environment.
View
Representation of Hybrid Systems by Means of Stochastic Automata
Article
  • Dec 2001
In this paper hybrid systems with discrete-event measurements and inputs are considered. Such systems consist of a continuous-variable subsystem and a discrete-event subsystem which are connected in a closed loop. Stochastic automata are used as models of the discrete-event behavior of the hybrid system, which is in general non-deterministic. It is shown that the model can be set up by, first, modeling the discrete-event behavior of the continuous-variable subsystem by means of a stochastic automaton. It is shown, that the transition relation of this automaton can be derived from a continuous-variable description of the subsystem. Second, the resulting model is combined with the model of the discrete-event subsystem, which in general can be represented by a deterministic automaton. As the continuous-variable subsystem and, therefore, also the hybrid system cannot be modeled exactly by an automaton, the model of the hybrid system generates spurious solutions. Therefore, conditions on the model of the continuous-variable and on the model of the hybrid system will be presented which yield the best complete model of the hybrid system, i.e., the model which generates all event sequences which may occur in the hybrid system and the least number of spurious solutions.
View
Lateral Distribution of Mixed Traffic on Two-Lane Roads
Article
  • Jul 2006
Location of the outer wheel path is an important input in the measurement of deflection response of a pavement and its riding quality. In the simulation of mixed traffic flow, it is also required to assign the position of a vehicle within a lane when it enters the test section. The statistical distributions fitted through the placement data recorded on 19 different sections of two-lane highways in India indicate that the placement of vehicles may follow a unimodal or a bimodal Curve depending upon the percentage of slow and fast moving vehicles. Two parameters are introduced in this technical note. These are placement factor (PF) and skewness range (SR). It is observed that the placement data follow a unimodal curve if PF and SR are less than 1.3 and 0.54, respectively. Otherwise it is a case of bimodality.
View
Vehicle Dynamics and Control
Book
  • Jan 2006
Vehicle Dynamics and Control provides a comprehensive coverage of vehicle control systems and the dynamic models used in the development of these control systems. The control system topics covered in the book include cruise control, adaptive cruise control, ABS, automated lane keeping, automated highway systems, yaw stability control, engine control, passive, active and semi-active suspensions, tire models and tire-road friction estimation. In developing the dynamic model for each application, an effort is made to both keep the model simple enough for control system design but at the same time rich enough to capture the essential features of the dynamics.
View