Experiments and Theoretical Analysis of Terrain-based Vehicle Localization to obtain GPS-Equivalent Vehicle Location Accuracy

GPS is a fragile sensing system: it is easily blocked by commonplace roadside features, easily confused by multi-path reflections, and easily jammed in wartime. Even so, the potential of using position information to enhance vehicle stability and performance has driven extensive research the past several decades. Recent GPS-based algorithms show great potential to save money, the environment, and human lives through enhanced estimation of driving conditions, improving performance of hybrid vehicle power management controllers, and autonomous vehicle guidance or driver-assist for dangerous situations. This talk will present experimental and theoretical results of an alternate to GPS for roadway positioning that uses the small rocking motions of the driven vehicle as a location-specific “fingerprint” of the road, thus enabling a map-equipped vehicle to calculate its position using only on-vehicle inertial measurements. Experiments using data collected this past year from over 6000 miles of roadway indicate that, by correlating such roll and pitch disturbances, it is relatively straightforward to achieve sub-meter longitudinal localization accuracy in real-time. Further, once can clearly discriminate laterally which lane the vehicle is traveling in, discriminate roadway departure, and identify sensor faults.
The correlation of a disturbance signal to a digital map presents interesting theoretical problems: both the map and disturbance signals are noisy and subject to drift, the probability density function of the uninitialized estimate is inherently multi-modal which makes linear estimators difficult to use. Data-storage limits require that the map be sparsely sampled to a level that makes extended Kalman filters impossible to use. To overcome these issues, we present a hybrid method using both a Particle Filter (PF) to initialize the algorithm, and an Unscented Kalman Filter (UKF) to maintain the estimate at steady-state. Similarities between the two methods are exploited to obtain an explicit solution for the steady-state PF accuracy using the Algebraic Ricattti Equation. Analysis of the sampling step of the PF shows that prediction of the theoretical accuracy requires a solution to the Lambert W function. The predicted accuracy using this hybrid analysis show remarkable agreement to experimentally measured accuracies measured across a wide variety of driving conditions. The talk will conclude with a discussion of ongoing work and future challenges related to advanced vehicle guidance.

Dr. Sean Brennan has been an Assistant Professor of Mechanical Engineering at Penn State University since 2003 and shares a joint faculty appointment with the Thomas D. Larson Pennsylvania Transportation Institute. Since 1998, he has published on topics ranging from vehicle systems dynamics; robust control; and vehicle chassis sensing and control. His current areas of study include modeling and experimental validation of vehicle dynamics, hardware- and human-in-the-loop experimental testing, electric vehicle primary power management and control, and design and control of UGVs. In addition to being the secretary of the International Forum for Road Transport Technology, he is currently the Chair of the ASME Dynamic Systems and Control Technical Committee on Automotive and Transportation Systems, and the Chair of the Education Subcommittee for Transportation Visualization of the Transportation Research Board of the National Academies. In 2006 he was selected for Penn State’s Premier Service Award for his initiation of an annual high-school summer camp focused on advanced vehicle and robotics technologies. In 2007, he won the Engineering College Outstanding Teaching Award, the top honor at Penn State University for engineering faculty, and in 2008 he was selected for the 2008 Spirko Award at Penn State for technology educators, and the 2008 SAE Teetor Award for educators in the area of mobility.