| This thesis aims to build the groundwork for a distributed network of collaborating, intelligent video surveillance cameras. This work is implemented on a system of low-cost embedded-microprocessor-based camera modules. Each camera develops a person detection classifier using the Winnow algorithm for unsupervised, online learning. Training examples are automatically extracted and labelled from the current frame, and the classifier is then applied to the image to detect and locate person instances.;To improve detection accuracy, multiple cameras with overlapping fields of view collaborate to confirm their results. We present a novel, unsupervised calibration technique that allows each camera module to understand its spatial relationship with the other cameras. During runtime operation, cameras efficiently apply the learned spatial correlations to confirm detections. This method implicitly handles non-overlapping field of view regions that cannot be confirmed by other cameras. Its computational efficiency makes it well-suited to real-time processing on our hardware. |
