Fusion of heterogeneous features via cascaded on-line boosting

In this paper, we employ the recent on-line boosting framework to fuse heterogeneous features for object detection and tracking in a video surveillance application. Detection and tracking are treated as a classification problem by an ensemble of weak classifiers built on heterogeneous feature types and updated on-line. We extend the on-line boosting framework by proposing an algorithm that builds a cascade of classifiers dynamically. The procedure takes into account both the error and the computational requirements of the available features and populates the levels of the cascade accordingly to optimize the detection rate while retaining real-time performance. We show the effectiveness of employing different features on real-world video sequences.