Bridging the Gaps between Cameras

Introduction

In multi-camera surveillance systems, the relationships between the cameras must be determined to enable integration of information from multiple views. Our research deal with two particular challenges:

Automatic Network Camera Calibration: A network of surveillance cameras automatically estimates the relationships between the different views (overlapped, adjacent, distant). Such a capability is very useful especially for large network of cameras, because it allows new cameras to "plug-n-play" in an existing network.
"Blind Tracking": Track objects as they move from one camera to another, even if they move across the "blind" areas of the network, where no camera can see.

Camera views of six cameras

Our Approach

A common solution to this problem is to assume a common "ground plane" which is seen by all the cameras. The position of the cameras are defined on the ground plane and tracking is performed using predictors (such as Kalman filter), extrapolating the motion of the object across the "blind" area. the approach we present here is based on a different model. The relationships between cameras are defined using tempo-probabilistic terms, instead of spatial terms. These links are learnt automatically from observations as demonstrated below. These links define an activity model, expressed by a Dynamic probabilistic Network, that covers the area viewed by the system, including the gaps between the cameras.

Tempo-probabilistic links between cameras

Method

Initially, the entry/exit regions of all the camera views are learnt from observations, using the Expectation-Maximisation Algorithm:

Each detected region becomes a node of the activtiy model.

The following ground plane map shows the positions of the different nodes in the scene

We model the possible link between an exit zone i and an entry zone j as follows:

k: virtual “rest of the world” node
n_i(t), m_j(t): disappearance/appearance signals
p_n, p_m: probabilities of disappearances/appearances
α_ij, α_ik, π_j: transition probabilities

The links of the activity model are estimated by the cross-correlation between the time sequences n_i(t) and m_j(t). If a real link exists between the nodes i and j, the cross-correlation has a clear peak. Additionally, estimates of the the transition times and probabilities can be extracted by the cross-correlation.

Cross-correlation of valid links:

Cross-correlation of invalid links:

The knowledge of the detected links reveals the camera topolofy of the system:

The following figures show the detected links between nodes for a single camera and for three differrent cameras. It is interesting that no tracking information has been used to estimated the links, only cross-correlation of events in different areas.

Tracking example

Publications

About