Multi-Component Spatiotemporal Attention and its Application to Object Detection in Surveillance Videos

Abstract

This paper describes multi-component spatiotemporal attention mechanisms in application to object detection in videos. The detection of objects of interest relies on the analysis of feature-point areas (FPAs), which correspond to the object-relevant focus-of-attention (FoA) points extracted by the proposed spatiotemporal mechanisms of attention focusing. The attention mechanisms give detection priority to object-relevant FPAs with spatial saliency, spatiotemporal coherence, and area temporal change including motion. The preliminary test results of the proposed attention focusing mechanisms for object detection and tracking have confirmed its advantage in terms of robustness over existing visual attention-based detectors with comparable run-times.