Intelligent path prediction for vehicular travel

James Alan Krozel, Purdue University

Abstract

This thesis presents a methodology for intelligent path prediction, predicting the motion of an observed vehicle by reasoning about the actions taken by the operator of the vehicle. The thesis considers only the problem of tracking a vehicle performing a transit mission, a mission that proceeds from a start location to a goal location guided by an intelligent planning criterion. Intelligent planning criteria are modeled by generalized cost criteria, which allow for basis cost criterion, for example, distance, visibility, or safety, to be considered, or convex combinations of such basis cost criteria to be considered. Two main investigations are addressed. The first investigation is to develop a method for identifying a decision-making strategy that seemingly explains the observed vehicle's motion. The start location is assumed to be known, the goal location is assumed to be in a given set of candidate goal locations, and an intelligent planning criterion is assumed to be guiding the vehicle. Given the history of the observed vehicle's path, the objectives are to (1) select the cost criterion that best explains the observed motion, (2) predict the goal location of the vehicle, and (3) predict the future path leading to the goal location. In solving this problem, search pointer information provided by reverse graph searches is exploited. Best fitting a goal location and cost criterion pair to explain the decision-making strategy of the vehicle is accomplished through a path similarity correlation measure comparing the observed path data to search gradient direction information. The second investigation is to develop a method of automatically proposing a candidate goal location for the observed vehicle. Given the cost criteria that best explains the history of the observed vehicle's path, the objectives are to (1) propose a region of plausible goal locations and (2) rank the locations in the set of plausible goal locations based on some heuristic merit. In solving this problem, search pointer information provided by forward graph searches is exploited. Establishing the region of plausible goal locations is accomplished using search tree results. Finally, the region of plausible goal locations is analyzed to rank the locations based on heuristic merit.

Degree

Ph.D.

Advisors

Andrisani, Purdue University.

Subject Area

Aerospace materials|Computer science|Electrical engineering|Artificial intelligence

COinS