Active Tracking System with Rapid Eye Movement Involving Simultaneous Top-down and Bottom-up Attention Control

Visual tracking of complex objects has been studied extensively in the field of robot vision, visual servoing, and surveillance. Detection of reliable low level visual features has been crucial, in the literature, for the stability of tracking in cluttered scene. However, using only low level features for tracking tends to be illusory for practical vision systems, and the selection of reliable visual features is still an important unsolved issue. In recent studies it has been stressed the importance of high-level features in guiding control for long-duration tracking (Matsugu et al., 2006; Li et al., 2007; Yang et al., 2007), computer vision (Sun & Fisher, 2003), and visual search (Lee et al., 2005). Recent cognitive neuropsychological as well as brain imaging studies also revealed a role of top-down attention in visual search task in human vision system (Patel & Sathian, 2000; Hopfinger et al., 2000; Corbetta & Shulman, 2002; Navalpakkam & Itti, 2006). In this chapter, we present a new object tracking vision system that incorporates both topdown and bottom-up attention processes. Tracking a specific object using both low-level and high-level features is not new and was previously studied (Isard & Blake, 1998) in a stochastic framework. The proposed active vision system in this chapter is task-oriented and tracks a specific object (i.e., person). The tracking process is initiated by the top-down process which is activated by robust object detection module (Matsugu & Cardon, 2004; Matsugu et al., 2004). Subsequent feature selection for tracking involves simultaneous consolidation mechanism between higher level complex features (e.g., face) obtained from the top-down process and low level features from the bottom-up process (detailed description is in Section 3). The active vision system controls eye movement based on prediction of an attended object’s location by the above processes. Main contribution of this chapter is that we propose a stable object tracking algorithm involving selective attention together with FF and FB hybrid control that enable smooth and saccadic pursuit. Specifically, we introduce a coherency measure of tracking features. Tracking using such measure ensures stability and fast recovery from failure (missing the object to be tracked) by way of consolidation among bottom-up and top-down attention cues. For the bottom-up feature-based prediction of tracked object, we use local color histogram, and histogram intersection (Swain & Ballard, 1991; Birchfield & Rangarajan, 2005) is used for feature matching. High-level, top-down feature for tracking is defined as detected face O pe n A cc es s D at ab as e w w w .ite ch on lin e. co m