Fusing Shape and Color for High-Speed Visual Tracking
MetadataShow full item record
In a computer vision system, one must often assume that the system will encounterobjects it has never seen before. A component that is able to track such objects is therefore an integral part of a complete computer vision solution. Given a region capturing the target object in a video stream, the tracker should compute the region capturing the object for every subsequent frame. This is problem is known as model-free tracking. Recent research has focused on improving tracking performance, disregarding computation time, leading to most modern trackers being unable to operate in a live setting. This thesis explores the task of visual model-free, short-term tracking in real-time. By training and combining two object models, individually shown to perform well in terms of both tracking performance and computation time, a high-speed, high-performance tracker is developed. The first model is of the object s shape, implemented as a kernelized correlation filter on histograms of oriented gradients. The second is a novel model of the object s color, implemented as a Bayes classifier on regularized color histograms of the object and surrounding background. While both models have weaknesses, they naturally complement each other, resulting in a robust short-term tracker. A novel model fusion procedure is developed, which combines the localizations of bothmodels based on the ratio between independent confidence measures. A novel confidence measure is developed for the color model from pixel-wise probabilities, while theshape model s confidence measure is the peak-to-sidelobe ratio. Both confidence measures are efficiently evaluated from integral images. The developed tracker is named the Adaptive Shape and Color Tracker (ASCT). It isevaluated on the Visual Object Tracking Challenge 2016, and compared to other participating state-of-the-art trackers. The evaluation shows that the developed tracker achieves state-of-the-art performance while operating at over a hundred frames per second. Most state-of-the-art trackers are implemented in Matlab, while the ASCT is written in C++ . Its high speed compared to the competition is attributed to a careful implementation in a more efficient language. A complete C++ implementation is included that is easily integrated into existing soft-ware without any overhead. Being written in C++ , the tracker will also run on nearlyany platform.