Real-time edge-enhanced dynamic correlation and predictive open-loop car-following control for robust tracking

Abstract

We present a robust framework for a real-time visual tracking system, based on a BPNN-controlled fast normalized correlation (BCFNC) algorithm and a predictive open-loop car-following control (POL-CFC) strategy. The search for the target is carried out in a dynamically generated resizable search-window. In order to achieve the robustness, we use some edge-enhancement operations before the correlation operation, and introduce an adaptive template-updating scheme. The proposed tracking algorithm is compared with various correlation-based techniques and (in some cases) with the mean-shift and the condensation trackers on real-world scenarios. A significant improvement in efficiency and robustness is reported. The POL-CFC algorithm approximates the current velocity of an open-loop pan-tilt unit, computes the predicted relative-velocity of the object using Kalman filter, and generates the precise control signals to move the camera accurately towards the maneuvering target regardless of its changing velocity. The proposed system works in real-time at the speed of 25–200 frames/ second depending on the template size, and it can persistently track a distant or near object even in the presence of object fading, low-contrast imagery, noise, short-lived background clutter, object-scaling, changing object-velocity, varying illumination, object maneuvering, multiple objects, obscuration, and sudden occlusion.