Representing and Scoring Track Hypotheses for Multitarget groups.

摘要

This paper presents an economical method for representing track sets for observations of possibly unstable groups of vehicles, and an algorithm to create and update the multi-vehicle track association hypotheses. We represent bundles of tracks to avoid redundant computing, but we remember to multiply our track scores by how many individual target tracks are being represented. This paper is divided into two main parts. The first part describes representing these bundles of tracks efficiently and the related logic of track extension for multi-target observations. To begin, we consider the combinatorial explosion of association hypotheses for tracking groups of target objects that may split or merge. For a population of 10 targets, if five are observed in one spot and then three are observed in another spot, there are 9, 072,000 track combinations represented and covered by a single update. However, as we will show, all of these equivalent possibilities can be represented as a single track or track bundle. We count the combinations represented by one multitrack, then we multiply the single-thread track score times the number of threads represented to get the total probability represented. We will derive a formula for the multitrack score update. The second part of the paper describes an algorithm for extending and initiating multitracks. The algorithm for extending and initiating multitracks is an enhancement of the PSR Tracker Multi- Hypothesis Manager algorithm with pruning. However, where the old algorithm constructed all valid hypotheses for a new observation and then pruned, the new algorithm uses a least cost branch and bound technique to create only those hypotheses likely to survive the pruning process. This results in a significant reduction in the number of hypotheses generated.