Attention-Based Target Localization Using Multiple Instance Learning

Abstract

We propose a novel Multiple Instance Learning (MIL) framework to perform target localization from image sequences. The proposed approach consists of a softmax logistic regression MIL algorithm using log covariance features to automatically learn the model of a target that persists across input frames. The approach makes no assumptions about the target’s motion model and can be used to learn models for multiple targets present in the scene. The learned target models can also be updated in an online manner. We demonstrate the validity and usefulness of the proposed approach to localize targets in various scenes using commercial-grade surveillance cameras. We also demonstrate its applicability to bootstrap conventional tracking systems and show that automatic initialization using our technique helps to achieve superior performance.