Fast structured tracker with improved motion model using robust Kalman filter

In this chapter, we discuss the problem of tracking objects without having to learn a model of the object’s appearance in advance (i.e., model-free tracking). Object tracking has many practical applications, which motivated continous developments in the field. A particularly successful tracking technique that has been recently proposed is the structured tracker called Struck. Struck uses support vector machines to learn the object’s appearance as well as to predict its location from one video frame to the next. However, tracking remains a challenging problem due to change of lighting conditions throughout the video, change of scale, rotation of the object of the interest, and occlusions. Here, we present a tracker that is based on Struck. Structured learning is a suitable framework for tracking as it calculates a discriminative function mapping the appearance information from one video frame onto the bounding box containing the tracking object in another video frame. We discuss a number of concepts that help keep limit computational complexity throughout tracking: budget for support vectors and fast computation of the intersection kernel. We then show how the use of the Robust Kalman filter can help smooth tracking results, and consequently make the tracker robust to false-negative detections. We show how the Robust Kalman Filter helps detect short-time occlusions by classifying them as outliers. We discuss the tracking evaluation using different evaluation protocols, and present experimental results that show that the structured tracker with the robust filter improves upon original Struck tracker regardless of the features or kernels used.