Object Tracking over Multiple Uncalibrated Cameras Using Visual, Spatial and Temporal Similarities - Advanced Concepts for Intelligent Vision Systems - page 168

The main contributions of this paper are the methods of reusing kernels and
location-based kernel matching.
This paper is organised as follows. In section 2, we propose the Signature-
based Tracking Across Cameras (STAC) multi-camera tracking algorithm. In
section 3, we describe the experimental setup and evaluation methods used. In
section 4, we present results. Finally, in section 5 we make conclusions.
2 The STAC Algorithm
The proposed STAC algorithm runs on each camera in a multi-camera network
as part of a parallel tracking framework. Each camera considers itself to be the
local camera and other cameras as foreign cameras . We assume that each camera
captures video at the same frame rate and that frames have been synchronised 1 .
Time is measured in units of frames . We assume that object detection and single-
camera tracking have already been performed and the results of the single-camera
tracking are the input of the STAC algorithm. This setup is shown in Fig. 1.
Specifically, we assume that for each tracked object in the local camera we know:
the object's centre ( x , y ) in pixels; the height and width of the object's bounding
box in pixels; the object's signature in the most recent frame; and a unique
identifier (track ID) for the track of the object in the local camera. A distance
metric on the signature must be defined. For details of the signature type and
metric used in our implementation see section 3.2.
The process of the STAC algorithm comprises three main steps:
1. finding visual similarities between pairs of objects;
2. finding spatial and temporal similarities between current pairs of objects and
historical pairs of objects; and
3. determining correspondences between pairs of tracks.
Step 1 includes the novel concept of reusing kernels and the method of location-
based kernel matching. Sections 2.1, 2.2 and 2.3 describe the details of steps 1,
2 and 3 above, respectively.
2.1 Finding Visual Similarities between Pairs of Objects
The STAC algorithm first attempts to find relationships between tracked objects'
locations in the local camera's field of view and locations in foreign cameras'
fields of view, based on the information received from the single-camera tracker.
Let kernels represent locations of objects in a field of view, and a linked pair of
kernels represent the visual and temporal relationship between two kernels. For
simplicity, details of this process are described in the following subsections from
the point of view of one camera and for one tracked object.
1 In our experiments, we lift the assumption by employing a framerate compensation
algorithm that examines timestamps of captured frames.