Detecting Luggage Related Behaviors Using a New Temporal Boost Algorithm ∗

In this paper we propose an approach to recognize luggage related behaviors in public spaces. We model behaviors in a multiclass learning framework, defining four classes: (i) walking, (ii) not moving, (iii) picking up/leaving bag, and (iv) abandoned bag. We rely on the output of a tracking algorithm to generate targets in each image. Then, we analyze each target separately, by computing three features: (i) optic flow, (ii) motion energy, and (iii) bounding box area. The features are fed into a novel boosting algorithm that adds temporal consistency in a short-term way. This temporal boosting algorithm considers time explicitly in the weak classifiers, leading to an improvement in noise robustness and performance. We show that our approach, with very simple features and a time-based boosting algorithm, is able to generate properly alarms on suspicious behaviors in a sequence of PETS 2007 database.