AbstractsComputer Science

In recent years, personal computing devices such as laptops, tablets and smartphones have become ubiquitous. Moreover, intelligent sensors are being integrated into many consumer devices such as eyeglasses, wristwatches and smart televisions. With the advent of touchscreen technology, a new human-computer interaction (HCI) paradigm arose that allows users to interface with their device in an intuitive manner. Using simple gestures, such as swipe or pinch movements, a touchscreen can be used to directly interact with a virtual environment. Nevertheless, touchscreens still form a physical barrier between the virtual interface and the real world. An increasingly popular field of research that tries to overcome this limitation, is video based gesture recognition, hand detection and hand tracking. Gesture based interaction allows the user to directly interact with the computer in a natural manner by exploring a virtual reality using nothing but his own body language. In this dissertation, we investigate how robust hand detection and tracking can be accomplished under real-time constraints. In the context of human-computer interaction, real-time is defined as both low latency and low complexity, such that a complete video frame can be processed before the next one becomes available. Furthermore, for practical applications, the algorithms should be robust to illumination changes, camera motion, and cluttered backgrounds in the scene. Finally, the system should be able to initialize automatically, and to detect and recover from tracking failure. We study a wide variety of existing algorithms, and propose significant improvements and novel methods to build a complete detection and tracking system that meets these requirements. Hand detection, hand tracking and hand segmentation are related yet technically different challenges. Whereas detection deals with finding an object in a static image, tracking considers temporal information and is used to track the position of an object over time, throughout a video sequence. Hand segmentation is the task of estimating the hand contour, thereby separating the object from its background. Detection of hands in individual video frames allows us to automatically initialize our tracking algorithm, and to detect and recover from tracking failure. Human hands are highly articulated objects, consisting of finger parts that are connected with joints. As a result, the appearance of a hand can vary greatly, depending on the assumed hand pose. Traditional detection algorithms often assume that the appearance of the object of interest can be described using a rigid model and therefore can not be used to robustly detect human hands. Therefore, we developed an algorithm that detects hands by exploiting their articulated nature. Instead of resorting to a template based approach, we probabilistically model the spatial relations between different hand parts, and the centroid of the hand. Detecting hand parts, such as fingertips, is much easier than detecting a complete hand. Based on our model of the…