Coupling Detection and Data Association for Multiple Object Tracking.pdfVIP

Coupling Detection and Data Association for Multiple Object Tracking Zheng Wu, Ashwin Thangali, Stan Sclaroff and Margrit Betke∗ Departments of Computer Science, Boston University, Boston, MA 02215 {wuzheng,tvashwin,sclaroff,betke}@ Abstract alarms” and “missed detections” otherwise propagate to the data association module and false matches need to be cor- We present a novel framework for multiple object track- rected later. In contrast, we show that error propagation ing in which the problems of object detection and data as- from detection to data association can be avoided if both sociation are expressed by a single objective function. The tasks, detection and data association, are combined into a framework follows the Lagrange dual decomposition strat- single module and solved simultaneously by optimizing a egy, taking advantage of the often complementary nature single objective function. This coupling idea appears attrac- of the two subproblems. Our coupling formulation avoids tive but introduces new challenges as well: 1) What type of the problem of error propagation from which traditional objective function should be used? Many existing detection “detection-tracking approaches” to multiple object track- methods have not even been formalized with an objective ing suffer. We also eschew common heuristics such as “non- function. 2) How can the new objective function be solved? maximum suppression” of hypotheses by modeling the joint Many current data association methods are complicated and image likelihood as opposed to applying independent like-