Object tracking in videos (image sequences) is an important computer vision task with many different applications, including intelligent robots and vehicles, augmented reality, entertainment, scientific and industrial applications, and human-computer interfaces. This paper presents a robust video-based object tracker that can cope with pose and illumination changes, partial occlusion, and background clutter.
The authors propose augmented kernel matrix (AKM) classification, a support vector machine (SVM) learning-based method that can select and add training patterns from a video sequence. It is based on the weighted combination of two complementary extracted features from images: pixel intensity and local binary patterns (LBP). The authors compare the proposed AKM tracker (AKMT) method with several other state-of-the-art video object tracking methods: FragTrack (Frag), multiple instance learning (MIL), visual tracking decomposition (VTD), visual tracker sampler (VTS), StruckH, StruckIL, concatenated feature tracker (CFT), and simpleMKL tracker (SimMKLT). It obtains very good results, always presenting one of the best performances in a set of nine video benchmark image sequences tested.
On the other hand, it would have been interesting if the authors had made the code of their implementation available, and also if they had compared the proposed object tracking method with some well-known methods available in the OpenCV framework, for example, methods for object tracking based on ORB, FREAK, BRISK, SIFT, or SURF feature extraction.