Estimating Tracking Sources and Sinks

When tracking in a particular environment, objects tend to appear and disappear at certain locations. These locations may correspond to doors, garages, tunnel entrances, or even the edge of a camera view. A tracking system with knowledge of these locations is capable of improved initialization of tracking sequences, reconstitution of broken tracking sequences, and determination of tracking sequence termination. Further, knowledge of these locations is useful for activity-level descriptions of tracking sequences and for understanding relationships between non-overlapping camera views. This paper introduces a method for simultaneously solving these coupled problems: inferring the parameters of a source and sink model for a scene; and fixing broken tracking sequences and other tracking failures. A model selection criterion is also explained which allows determination of the number of sources and sinks in an environment. Results in multiple environments illustrate the effectiveness of this method.

1. Introduction

Given an unoccluded, far-field view of an air field (see Figure 1(a)), an effective tracking system could benefit from the knowledge that a previously untracked object should appear near the edge of the camera's field of view and that the object should be continuously tracked until it again reaches the edge of the camera's field of view. Automatically estimating these locations for more complex scenes can aid in initialization of tracks, termination of tracks, and reestablishing correspondence after tracking failures. Further, the understanding of where objects enter and exit a scene is an important element of activity understanding and in learning relationships between tracked objects across nonoverlapping fields of view.