IV 2010 Ph.D. Forum Participants and Abstracts
 
Vehicle Communication
 
Analytical Modeling of Delay-Tolerant Data Dissemination in Vehicular Networks
Ashish AgarwalProf. Thomas D.C. Little
Electrical and Computer Engineering, 2010Boston University
Vehicular networking is an emerging technology to support applications involving communications between vehicles, between vehicles and fixed access points, and between vehicles and the Internet cloud. The goal is to enable vehicles to exchange information for improved safety through situational awareness, enhanced convenience, and achieve increased levels of efficiency in time and energy consumption. Safety messaging, real-time traffic updates, remote diagnostics, Internet access and in-car entertainment are examples of applications targeted by this technology.
    This dissertation considers communication and networking among vehicles that are constrained to navigable roadways. We propose a novel routing technique that incorporates attributed (labeled) messaging; geographic routing; and delay tolerant networking techniques in a solution that operates in a network characterized by rapid mobility and time-varying partitioning (fragmentation). An analytical model is developed to demonstrate the performance of opportunistic data exchange in a delay tolerant network setting.
    Contributions of the work include revelation of phase transition behavior due to vehicle density and transmission range. We are able to identify regimes of density where gains are achieved by exploiting the opportunistic contacts between vehicles traveling in opposing directions. Also significant is the observation that increased mobility of nodes from 0 m/s to 10 m/s yields an order of magnitude increase in the performance of messaging from 0 m/s to 200 m/s. The proposed architecture is compared with existing mobile ad hoc networking schemes and performance gains achieved are detailed. It is demonstrated that in a hybrid environment with intermittently placed access points, large separations are possible when supported by multihop networking. Under delay tolerant networking assumption, minimum delay and maximum propagation rates are achieved for low vehicular traffic densities of 20 vehicles/km, for given parameters. A path based messaging scheme would achieve similar performance at 40 vehicles/km.
 
Evaluating and Improving Broadcast Reliability in Vehicular Ad Hoc Networks
Rex ChenProf. Amelia Regan and Prof. Wen-Long Jin
Computer Science, 2010University of California, Irvine
Traffic congestion and accidents continue to take a toll on our society with congestion impacting billions of dollars in economic cost and millions of traffic accidents annually worldwide. Vehicular Ad hoc Network (VANET) based information systems have considerable promise for improving traffic safety, reducing congestion and increasing environmental efficiency of transportation systems. However, there are numerous technical hurdles for deploying VANET on the road network and its full potential will not be realized until the issues related to communication reliability and security are fully resolved.
    VANET is a specific type of mobile ad hoc network (MANET) with unique characteristics that are different from a general MANET. These attributes include the mobility model (vehicle movements) and the network topology (road layout) imposed by the underlying transportation system. In this dissertation, we study broadcasting schemes for VANET and investigate ways to improve reliability under numerous traffic conditions (free flow and congested flow traffic scenarios). Further, we incorporate an understanding of traffic flow theory to increase the communication reliability and efficiency in dynamic vehicular networks. The contributions in this dissertation will be of significant interest to both the computer networking and transportation research communities.
 
 
Driver Understanding
 
Route Choice: A Behavioral Analysis and Modeling Approach
Aly TawfikProf. Hesham Rakha
Civil and Environmental Engineering, 2011Virginia Tech
Within the context of transportation modeling, driver route choice is typically captured using mathematical programming approaches. These approaches assume that drivers have close-to-perfect knowledge of the transportation network state in attempting to minimize some objective function. Typically, drivers are assumed to either minimize their travel time (user equilibrium) or minimize the total system travel time (system optimum). Given the dynamic and stochastic nature of the transportation system, the assumption of a driver’s close-to-perfect knowledge is questionable. While it is well documented in psychological sciences that humans tend to minimize their cognitive efforts and follow simple heuristics to reach their decisions, especially under uncertainty and time constraints, and that human perceptions, consequently, are often different from actual reality, in the late stages of typical route choice models drivers are assumed to have close to perfect knowledge of their choice set, as well as the travel characteristics associated with each of the choice elements. In addition, only a few of the many route choice models that are described in the literature are based on observed human behavior. With this in mind this research monitors, analyzes and models actual human route choice behavior based on drivers’ actual experiences, perceptions and choices. based on drivers’ actual experiences, perceptions and choices.
 
Investigating Posture and Affect Dynamics at Multiple Levels of Details for Active Safety: A Vision Based Approach
Cuong TranProf. Mohan Trivedi
Computer Science and Engineering, 2012University of California, San Diego
These days, assistance systems for driving safety enhancement have an increasingly important role. To be effective, such systems need to be human centric and have a holistic sensing of environment, vehicle, and driver. This research focuses on looking at driver, using data from cameras (vision-based approach) to infer driver’s activities and cognitive states for active safety. Specifically, the goals are (1) developing novel computer vision techniques for tracking driver posture and extracting semantic information at multiple levels of detail (i.e. whole upper body, hands, head, and facial features), (2) developing a learning framework to study the role of extracted information at each level as well as how to fuse them for better inferring of driver’s activities and states, and (3) Applying this framework for interactive driver assistance. Understanding driver states and activities is crucial to both active safety and effective interactions between advanced assistance systems and drivers. Therefore the outcomes of these studies will benefit policy decisions on future preventative safety measures.
 
Toward Improved Speech Based Emotion Recognition in Driver Assistance System
Ashish TawariProf. Mohan Trivedi
Electrical and Computer Engineering, 2013University of California, San Diego
Automated analysis of human affective behavior has attracted increasing attention in recent years. In context of driving task, particularly, driver's emotion often influences driving performance which can be improved if the car actively responds to the emotional state of the driver. It is important for an intelligent driver support system to accurately monitor the driver's state in an unobtrusive and robust manner. Ever changing environment while driving poses a serious challenge to existing techniques for speech emotion recognition. In our research, we propose to utilize contextual information of the outside environment as well as inside car user to improve the emotion recognition accuracy. In particular, a noise cancellation technique is proposed to suppress the noise adaptively based on the driving context and use gender based context information for developing the classifier.
 
 
On-Road Detection
 
Traffic Sign Detection and Recognition Toward Smart Driver Assistance
Keisuke DomanProf. Hiroshi Murase
Graduate School of Information Science, 2012Nagoya University
The purpose of this research is to design a traffic sign detection and recognition technique toward a smart driver assistance system that can control the amount of information provided to a driver based on the visibility of visual targets. A high-accuracy traffic sign detection and recognition technique has been considered as important for driving safety. However, in recent years the number of driving safety support systems in a car is increasing. As a result, it is becoming important to select appropriate information from them for safe and comfortable driving since too much information may cause driver distraction and may contradictorily increase the risk of a traffic accident. One approach to avoid such a problem is to alert the driver only with information which could easily be missed. Therefore, to realize such a system, this research mainly focuses on estimating the visibility of traffic signs in addition to detecting and recognizing them with an in-vehicle camera. The proposed technique estimates the visibility of traffic signs based on the contrast of multiple image features in a traffic sign and its surrounding region. More details of this technique will be presented in the main session of the symposium.
 
Integrated Vision Systems for Driver Assistance in Intelligent Vehicles
Sayanan SivaramanProf. Mohan Trivedi
Electrical and Computer Engineering, 2012University of California, San Diego
Intelligent vehicles are an active area of research, with significant contributions being made in sensing, control, and information processing. Advances in the field have the potential to improve safety, reduce risk of injury, and save lives. The focus of this research deals with the challenges of utilizing data from cameras looking out of the vehicle, monitoring the driving environment. We propose several relevant research goals that include applied computer vision, information fusion, and systems integration. The ultimate goal is to instrument intelligent vehicles and deploy multi-objective integrated computer vision systems for intelligent driver assistance. The advantages of using multiple cues from integrated systems are substantial, allow for a deeper and more robust understanding of the on-road environment, and promise safer intelligent driver assistance systems.
 
Stereo Vision and Color Image Evaluation for Combined Scene Segmentation and Detection of Traffic Participants
Philip LenzProf. Christoph Stiller
Institute of Measurement and Control Systems, 2013Karlsruhe Institute of Technology
Modern driver assistance systems such as collision avoidance or intersection assistance need reliable information on the current environment. Extracting such information from camera-based systems is a complex and challenging task for inner city traffic scenarios. This paper presents an approach to combined scene segmentation and object detection using stereo and color information. The extracted features must be sufficient for a road geometry estimation in an urban environment. Therefore, a probabilistic formulation of segmentation and object classification is mandatory. This project in particular investigates features that are relevant for inner city road geometry such as house fronts or principle axes of cars.
 
Hybrid Models for vision-based Intersection Understanding
Andreas GeigerProf. Christoph Stiller
Institute of Measurement and Control Systems, 2012Karlsruhe Institute of Technology
Most existing approaches to visual intersection detection and classification heavily rely on lane marking and road texture features. However, in most realistic scenarios these features alone are rendered insufficient due to structured shadows, large-scale occlusions and damaged lane markings. In my work, I exploit additional features such as three dimensional scene flow and object detections obtained using a stereo camera rig. By using an efficient dense stereo matching algorithm the object detection system can be significantly improved and important properties of the scene, such as vanishing points and the road plane are easily extracted. Feature information is fused in a probabilistic manner using a generative graphical modeling of the scene. Common inference techniques such as Markov Chain Monte Carlo sampling can be employed to draw samples from the posterior over road topologies and topographies.
 
 
Vehicle Tracking
 
LIDAR, Camera, and Inertial Sensors Based Navigation and Positioning Techniques for Advanced ITS Applications
Lili HuangProf. Matthew Barth
Electrical Engineering, 2010University of California, Riverside
Sensor fusion techniques have been used for years to combine sensory data from disparate sources. This dissertation focuses on Lidar, camera and inertial sensors based navigation and vehicle positioning techniques. First of all, a unique multi-planar Lidar and computer vision calibration algorithm is proposed. This approach requires the camera and Lidar to observe a planar pattern. Then the geometric constraints of the "views" from the Lidar and camera images are resolved as the coordinate transformation and rotation coefficients. A tightly coupled LIDAR/CV integrated system for vehicle detection and tracking is also proposed in this dissertation. The LIDAR sensor data is transformed into the image coordinates. Different Regions of Interest (ROIs) in the imagery are defined based on the LIDAR object hypotheses. Then a classifier error correction approach is used to choose an optimal position of the detected vehicle. We also present an autonomous positioning solution for urban environment. The positioning solution is derived by combining measurements from both Lidar and inertial sensors. The last chapter of this dissertation is vehicle tracking used in traffic surveillance system. In this dissertation, a real-time multi-vehicle tracking approach is proposed, which combines both local feature tracking and a global color probability model.
 
Vehicle Segmentation and Tracking from Monocular Videos
Anh VuProf. Matthew Barth
Electrical Engineering, 2011University of California, Riverside
We propose a new framework to automate the processing of stationary video data, consisting of three stages. In the first stage, the camera’s internal and external parameters are calibrated. In the second stage, we segment vehicles coming into the view and initialize their poses (position and attitude) expressed in the ground plane frame. This task is carried out by identifying foregrounds from a background subtraction process as candidate regions, where a Genetic Algorithm (GA) is formulated to find vehicle tracking candidates as well as their poses. In the third stage, the vehicles extracted from the second stage are tracked using a particle filter using a predefined propagation model. The features identified on each of the vehicles during the initial pose are stored as templates where they are used as measurements to infer vehicle pose over time. In order to move the framework onto a probe vehicle, we employ an integrated navigation system which combines an Inertial Measurement Unit (IMU), DGPS, vision system(s), LIDAR(s), and ground based radio systems to estimate the state of the probe vehicle as well as the state of the calibrated onboard sensors using an Extended Kalman Filter (EKF). In order to accommodate the vehicle’s motion, stage one is eliminated since we have a system capable of estimating the probe vehicle and its sensors’ positions in a fixed earth frame. In stage two, a vehicle classification process is used to identify vehicle candidates instead of background subtraction. Stage three is also similar except for the fact that we also have to account for the probe vehicle’s motion.
 
Efficient Tracking of Public Transport in Urban Environments
Rohit KumarProf. David Castanon
Electrical Engineering, 2010Boston University
Good public transport is an important requirement for smooth working of an economy. It not only reduces the traffic on the road leading to free flow of goods and services, it also reduces the oil dependence. Advanced traveler information systems (ATIS) is one of the many efforts by the government wherein a rider is provided with accurate real-time information which not only reduces their wait times but it also gives a framework to estimate arrival times. In the first part, we aim to abstract the transportation problem with new models to better estimate the current location and consequently predict the time of arrival more accurately.
     The accuracy of the ATIS depends of the location information of the bus available at the base station. Hence it require frequent communication from the bus to the base station resulting in huge communication cost and prevents transit authorities from implementing such systems. In this second part of the PhD we develop various communication protocols for reducing communication cost. In order to encourage people to use public transport, government has to make it more attractive and user friendly especially in developing countries which is adding more than a million car on road every day.
 
A Unified View on Probabilistic Tracking and Situation Assessment
Robin SchubertProf. Gerd Wanielik
Professorship of Communications Engineering, 2010Chemnitz University of Technology
In-vehicle Advanced Driver Assistance Systems are based on the perception and interpretation of the vehicle’s environment. These tasks are often based on probabilistic tracking algorithms such as the Bayes filter and situation assessment algorithms, e.g., Bayesian networks. However, while these techniques are subject to intensive research, the interface between them has not yet been sufficiently addressed. Thus, the main research objective of this work is to provide a generic, bidirectional, probabilistic interface between tracking and situation assessment in order to allow a unified view on these tasks. For that, it is firstly analyzed how uncertainties from the probabilistic perception of the vehicle’s surrounding can be entered into a Bayesian network in order to directly influence the situation assessment. In addition, it is investigated how uncertain knowledge about the current situation can be used in order to support the tracking performance. For that, an extension of the Interacting Multiple Model (IMM) is proposed which is called the Meta Model Filter. This technique models possible maneuvers of vehicles inside a Bayesian network in order to adaptively adjust the transition probabilities of the IMM according to the current situation. With this approach, a situation-dependent multiple model filtering can be achieved.
 
Direct Trajectory Optimization of Vehicle Evasive Maneuvers
Damoon SoudbakhshProf. Azim Eskandarian
Civil and Environmental Engineering, 2011George Washington University
Study on FARS database showed that 53% of all road fatalities in 2008 were lane/road departure crashes. Many steering control systems have been developed to reduce road departures by performing lane change or evasive maneuvers. In most of these systems a controller was designed to make the vehicle follow a desired trajectory. In this research, finding desired trajectory based on some defined design criteria will be studied by directly optimizing the trajectory. The design criteria are the initial and final conditions and also minimization of some parameters such as lateral acceleration of the vehicle through the maneuver. This will be done by developing a model of the vehicle and using optimization techniques. The results will be compared to the trajectories suggested by other researchers. Based on the results, a new system with preferred trajectory will be developed to assist the driver in performing lane change or evasive maneuver in straight and curved lanes. To test ability of the controller to accommodate disturbances and nonlinearities of the real vehicle, the developed controller will be implemented on a complex model of a car in a commercial software package and the results will be compared to the simulation results.