Master's Thesis
Neighboring Vehicle Behavior Prediction Using A Gated Recurrent Unit Neural Network
Research in vehicular networks has intensified due to their promise of reducing vehicle collisions, providing driver assistance, and increasing fuel economy. A main component of vehicular networks is Vehicle-to-Vehicle (V2V) communication, allowing vehicles to share their status information by frequently broadcasting their state such as current speed or brake pressure. Sharing status information will allow other drivers to become aware of their location and immediate intent, but not necessarily their intended route or future behavior. Knowing the future behavior of neighboring vehicles may provide insight to drivers, allowing them to make more well-informed decisions such whether to perform a lane change or to reduce speed. Motivated by this observation, this paper investigates a new strategy for effectively predicting the intended behavior of all neighboring vehicles in a V2V setting. We believe that we are the first in literature to make use of publicly broadcast V2V messages instead of relying on invasive data to predict driver intent. These messages are collected and used as input for the predictive model based on a Gated Recurrent Unit Neural Network. We demonstrate that our method is faster and more accurate than a Long-Short Term Memory Neural Network, a popular choice among researchers. Experiments are conducted using real world data collected from drivers following a fixed route.
Publications
ePrivateEye: To the Edge and Beyond!. Christopher Streiffer, Animesh Srivastava, Victor Orlikowski, Yesenia Velasco, Vincentius Martin, Nisarg Raval, Ashwin Machanavajjhala, and Landon P. Cox. SEC 2017. San Jose, CA, October, 2017
Edge computing offers resource-constrained devices lowlatency access to high-performance computing infrastructure. In this paper, we present ePrivateEye, an implementation of PrivateEye that offloads computationally expensive computervision processing to an edge server. The original PrivateEye locally processed video frames on a mobile device and delivered approximately 20 fps, whereas ePrivateEye transfers frames to a remote server for processing. We present experimental results that utilize our campus Software-Defined Networking infrastructure to characterize how network-path latency, packet loss, and geographic distance impact offloading to the edge in ePrivateEye. We show that offloading videoframe analysis to an edge server at a metro-scale distance allows ePrivateEye to analyze more frames than PrivateEye’s local processing over the same period to achieve realtime performance of 30 fps, with perfect precision and negligible impact on energy efficiency.
Donghyun Kim, Yesenia Velasco, Wei Wang, R.N. Uma, Rasheed Hussain, Sejin Lee, "A New Comprehensive RSU Installation Strategy for Cost-Efficient VANET Deployment," IEEE Transactions on Vehicular Technology (TVT), vol. 66, issue 5, pp. 4200-4211, May 2017
&
Donghyun Kim, Yesenia Velasco, Zishen Yang, Wei Wang, Rasheed Hussain, and R.N. Uma, "Cost Effective Mobile and Static Road Side Unit Deployment for Vehicular Adhoc Networks," Proceedings of International Workshop on Computing, Networking and Communications (CNC) in conjunction with International Conference on Computing, Networking and Communications (ICNC 2016), February 15-18, 2016, Kauai, Hawaii, USA.
Recently, the studies on vehicular adhoc network (VANET) are booming due to the huge potential. Road side unit (RSU) is a key component of the VANET infrastructure connecting mobile vehicles and the rest of the infrastructure. To maximize the availability of RSUs, RSUs should be densely deployed. Otherwise, blind spots may exist in which vehicles lose the connection to the infrastructure. Unfortunately, the massive deployment of RSUs to seamlessly cover the whole area of interest, which could be a vast metropolitan, can be very expensive. As the effectiveness and the benefits of the VANET are not fully proven yet, such large scale deployment can hardly be a viable option as of today. Motivated by this observation, this paper investigates a new strategy to best deploy RSUs so that their spatio-temporal coverage is maximized under a limited budget. In detail, for the first time in the literature, we consider an innovative RSU deployment framework, which is a well-balanced combination of three different approaches, deploying RSUs on static locations, public mobile transportation, and fully controllable vehicles owned by the local government. We first introduce a new strategy to abstract a map of city area into a grid graph. Then, we formulate the problem as a new optimization problem and show its NP-hardness. To solve this problem, we transform this problem into another optimization problem. Then, we propose a new polynomial running time approximation algorithm for the problem and show that the performance ratio (the ratio between the quality of an output of the proposed algorithm and the quality of the best possible solution) is at least half of the best possible ratio. We also conduct simulations under various setting to study the effectiveness of the proposed approach.
Projects
CAT Vehicle REU Intern, University of Arizona, Tucson, AZ, 2015 summer
The field of autonomous vehicles is still a young one, with much garnered interest from researchers and car manufacturing industries alike. To ensure an improved driving performance of such vehicles, model predictive control (MPC) has gained some attention due to its ability to optimize solutions in real time while in view of constraints and future events, i.e. obstacles or the environment. This is done by repetitively calculating the optimal path represented as a finite set of predicted states, based upon a variety of variables and constraints until a target state is reached. Commonly, MPC is used with a single predictive model which has shown promising results for vehicle maneuvering under certain driving scenarios. The kinematic predictive model, for instance, performs well when driving under low steering angles, but due to its simplified nature, inaccuracy in the model quickly increases. The dynamic predictive model has shown to be more complex and precise than the kinematic and as a result, requires a longer computational time. If the dynamic model fails to return a solution within a reasonable amount of time, the fast, yet simplified kinematic model is preferred. Thus in order to account for a wider range of driving scenarios, multiple predictive models are needed. In this paper a hybrid MPC is used to alternate between several predictive models of an autonomous vehicle with uncontrollable divergence as our switching logic. Uncontrollable divergence is defined as the difference of the actual vehicle state from when the MPC was called to when it returns since the vehicle state is continually shifting while waiting for the MPC to return. The use of uncontrollable divergence as the switching logic leads to the predictive model with the best balance of fast optimization time and low model mismatch to be chosen given the current driving situation. The effectiveness of our solution is demonstrated in both a simulated and physical autonomous vehicle maneuvering successfully through an obstacle course. This method had shown to enhance path planning for autonomous vehicles all while avoiding obstacles.
Rams Intern, Oak Ridge National Laboratory, Oak Ridge, TN, 2014 summer
General Systems Problem Solver (GSPS), an algorithm for building models, looks for order in a finite set of seemingly disordered data. The algorithm’s search, when successful, allows an investigator the possibility of reconstructing the system that produced the data and, in some cases, it allows the investigator to even predict the system’s future behavior. In this study, we attempted to reconstruct a process for the growth of cancer cells, based on empirical data obtained through magnetic resonance imaging and other types of scans. Because GSPS is computationally intensive, it has not been widely used. Further, because there’s not yet much empirical data on cancer-cell behavior, it has been difficult to develop models that can reliably predict cancer-cell behavior. In our research, we decreased the computational needs by binning the data to reduce its variety: the binning allowed us to use a personal computer to perform the necessary GSPS calculations. The data we used came from a series of images of a tumor in different stages of growth. From these series of observations we calculated a mask that best forecasted the next state. Each observation consists of a center volumetric pixel (voxel) v1 and its six neighbors, with each voxel containing seven variables. The mask takes past values of a variable as input from v1 and its neighbors, and then outputs the next value of a different variable at v1. The mask is used to predict future behavior as soon as a pattern has been detected. Future investigations could take into account data on additional tumors, include more variables, and larger numbers of bins. This research could lead to a better understanding of cancer tumors and even a renewed interest in the GSPS.
Intern, Idaho National Laboratory, Idaho Falls, ID, 2016 summer
Worked on simulating the cold cap physics of a nuclear waste melter using Star-CCM+ modeling software, C, and Matlab. My main task was translating the source code for a highly mathamatical model simulation from matlab to C in order to process big data in a much more effeccient way.