Identification of moving objects in poor quality surveillance data

Abstract

In a world of pervasive visual surveillance and fast computing there is a growing interest in automated surveillance analytics. Object classification can support existing event detection techniques by identifying objects present allowing confident prioritization of the detected events. In this paper we propose an effective object classification algorithm to distinguish between four classes that are important for outdoor surveillance applications: people, vehicles, animals and `other'. A challenging dataset that has been obtained from an industry partner from real deployments of poor quality cameras is used to evaluate the proposed approach. Frame differencing was found to be the most suitable approach to detect moving objects with Histogram of Oriented Gradients (HOG) the preferred choice to represent the objects. An SVM was used for classification. The results show that the proposed approach gives higher accuracy than a similar approach based on SIFT and bag words.