Multi-Vehicle Tracking Using Microscopic Traffic Models
Journal Articles
Overview
Research
Identity
Additional Document Info
View All
Overview
abstract
In this thesis, multiple on-road vehicle tracking problem is explored, with greater consideration of road constraints and interactions between vehicles. A comprehensive method for tracking multiple on-road vehicles is proposed by making use of domain knowledge of on-road vehicle motion.
Starting with raw measurements provided by sensors, bias correction methods for sensors commonly used in vehicle tracking are briefly introduced and a fast but effective bias correction method for airborne video sensor is proposed. In the proposed method, by assuming errors in sensor parameter measurements are close to zero, the bias is separately addressed in converted measurements of target position by a linear term of errors in sensor parameter measurements. Based on this model, the bias is efficiently estimated by addressing it while tracking or using measurements of targets that are observed by multiple airborne video sensors simultaneously. The proposed method is compared with other airborne video bias correction methods through simulations. The numerical results demonstrate the effectiveness of the proposed method for correcting bias as well as its high computational efficiency.
Then, a novel tracking algorithm that utilizes domain knowledge of on-road vehicle motion, i.e., road-map information and interactions among vehicles, by integrating a car-following model into a road coordinate system, is proposed for tracking multiple vehicles on single-lane roads. This algorithm is extended for tracking multiple vehicles on multi-lane roads: The road coordinate system is extended to two-dimension to express lanes on roads and a lane-changing model is integrated for modeling lane-changing behavior of vehicles. Since the longitudinal and lateral motions are mutually dependent, the longitudinal and lateral states of vehicles are estimated sequentially in a recursive manner. Two estimation strategies are proposed: a) The unscented Kalman filter combined with the multiple hypothesis tracking framework to estimate longitudinal and lateral states of vehicles, respectively. b) A unified particle filter framework with a specifically designed computationally-efficient joint sampling method to estimate longitudinal and lateral states of vehicles jointly. Both of two estimation methods can handle unknown parameters in motion models. A posterior Cramer-Rao lower bound is derived for quantifying achievable estimation accuracy in both single-lane and multi-lane cases, respectively. Numerical results show that the proposed algorithms achieve better track accuracy and consistency than conventional multi-vehicle tracking algorithms, which assumes that vehicles move independently of one another.
