Robust Object Detection by Voting in Multiple Feature Spaces

概要

論文の詳細を見る
In this work, the performance of visual detection methods is improved from an aspect of combining information from different channels. The efforts are two-fold: 1) combining motion information with appearance information, and 2) combining visual and spatial information encoded among the local image features of the same object. Three detection methods are proposed, and the most important component is a voting system in each method. The first detection method is developed for real-time applications. By making time-consuming steps deal with fewer instances, the method combines motion information with appearance information efficiently, and gives promising results in real time. The second method extends the Implicit Shape Model to incorporate motion information, and outperforms the state-of-the-art method on two datasets. The third method does pyramid matching during training and detection for efficiency, makes full use of the visual and spatial information of local image features, and gives robust detection results efficiently.
2014-05-08