Feature Gate Computational Top-Down Model for Target Detection

Abstract

Computer vision is a technique used for processing images and videos which are increasingly becoming ubiquitous day by day. Technologies developed are revolving around human needs and demands high computational power as volume of data increases. The extraction of the necessary information for processing, that is independent of various scene complexity is a challenging task. Computation visual attention methods channelize a way to select the information using psychological studies on the human visual system. This work aims to develop a computational visual attention method to select the target in a scene. Feature Gate Top-Down model is proposed to filter the significant region of a target in the scene. The proposed model is extended as a choice-based system to detect target or salient movement in surveillance videos. Experimental analysis is performed on various scenes for detecting human as a target followed by the analysis on surveillance videos is evaluated. The metrics such as Kullback-Leibler divergence (KL div), Normalized Scanpath Saliency (NSS), Correlation Coefficient (CC) and similarity reveals that the proposed model is more adaptive in identifying the target region by suppressing other dominant objects.