Gestures are natural means of communication between humans, and therefore their application would benefit to many fields where usage of typical input devices, such as keyboards or joysticks is cumbersome or unpractical (e.g., in noisy environment). Recently, together with emergence of new cameras that allow obtaining not only colour images of observed scene, but also offer the software developer rich information on the number of seen humans and, what is most interesting, 3D positions of their body parts, practical applications using body gestures have become more popular. Such information is presented in a form of skeletal data. In this paper, an approach to gesture recognition based on skeletal data using nearest neighbour classifier with dynamic time warping is presented. Since similar approaches are widely used in the literature, a few practical improvements that led to better recognition results are proposed. The approach is extensively evaluated on three publicly available gesture datasets and compared with state-of-the-art classifiers. For some gesture datasets, the proposed approach outperformed its competitors in terms of recognition rate and time of recognition.

Detection of repeatable keypoints is often one of the first steps leading to obtain a solution able to recognise objects on images. Such objects are characterised by content of image patches indicated by keypoints. A given image patch is worth being described and processed in further steps, if the interest point inside of it can be found despite different image transformations or distortions. Therefore, it is important to compare keypoint detection techniques using image datasets that contain transformed or noisy images. Since most of detector evaluations rely on small datasets or are focused on a specific application of compared techniques, in this paper two large datasets which cover typical transformations, as well as challenging distortions that can occur while image processing, are used. The first dataset contains 200,000 transformed images, and it has been prepared for the purpose of this study. The second dataset, TID2013, is widely used for perceptual image quality assessment; it contains 3,000 images with 24 distortions. Finally, interest point detectors are evaluated on four datasets, and repeatability score and time of detection are used as measures of their performance.