Video Surveillance Using a Multi-Camera Tracking and Fusion System

Abstract

Usage of intelligent video surveillance (IVS) systems is spreading rapidly. These systems are being utilized in a wide range of applications. In most cases, even in multi-camera installations, the video is processed independently in each feed. This paper describes a system that fuses tracking information from multiple cameras, thus vastly expanding its capabilities. The fusion relies on all cameras being calibrated to a site map, while the individual sensors remain largely unchanged. We present a new method to quickly and efficiently calibrate all the cameras to the site map, making the system viable for large scale commercial deployments. The method uses line feature correspondences, which enable easy feature selection and provide a built-in precision metric to improve calibration accuracy.

The Design and Implementation of a Wireless Video Surveillance System

“…views to cover the overlapping views, we identify the unique objects across cameras within a cluster. To achieve this, Vigil uses the following lightweight approach [33] called re-identification. Figure 7 shows an example where two cameras simultaneously detect the faces of two people in their overlapping views.…”

Grouping Crowd-Sourced Mobile Videos for Cross-Camera Tracking

“…While a number of methods have been developed for cross-camera tracking and automatic calibration of multisensor networks [5,6], very few techniques are applicable to opportunistic, crowd-sourced data sets gathered from a heterogeneous collection of sensors, from different viewpoints, and possibly containing unrelated content, such as would be returned by a YouTube query.…”

Efficient Feature Distribution for Object Matching in Visual-Sensor Networks

“…In this way the network would build a more complex, distributed, multi-view representation of an object, while still preserving the advantages of the proposed, hierarchical, feature-distribution scheme. An alternative approach would be the introduction of supernodes [23] or fusion sensors [24] into the network structure. The task of such nodes would be to aggregate the information from spatially related nodes.…”

“…For example, Kirishima et al [23] propose a framework for multi-view gesture recognition based on distributed image processing. Zhang et al [24] deal with the key challenges for a multi-camera surveillance system, where each camera unit should share its information through the centralized datafusion sensor. Our framework does not explicitly deal with multi-view recognition or distributed calibration; however, those approaches can be used to extend the framework with such a functionality.…”