Hardware-accelerated Object Tracking

This work investigates hardware acceleration of object tracking by parallelising an algorithm for object classification involving decision trees. Object tracking is the process of recognizing and locating a particular moving object in the spatial as well as in the temporal domain of a video stream. One key application of object tracking is video surveillance, to provide an operator in a control room with novel tools for assessing complex events among hundreds of live videos. Object tracking can be achieved between two consecutive video frames through template-based or feature-based correlation of images. Although this approach is computationally efficient, it can be unreliable or unsuccessful, because the appearance of the object may drastically change or the object may become occluded. As an alternative, one can apply object classifiers in subwindows that vary in scale, size and position [1]. A classifier can identify an object appearances based on a limited feature set. One example of such a feature set are 2-bit Binary Patterns that capture brightness variation in certain rectangular regions of an object’s image. This feature set is used as an input for an ensemble of decision trees known as a random forest [2] that can determine the probability of the object being present in a search window. A single 2-bit Binary Pattern gives a very weak indication that the sought object is present in the current search window, while the mapping of several features on one decision tree, and the combination of several trees, can identify objects with high confidence. A classifier can automatically be trained, starting from a given instance of the object’s appearance [1]. Appearance changes are addressed through P-N learning [3], an machine learning technique that combines a Lucas-Kanade frame-byframe tracker [4] with a random-forest-based classifier. P-N learning identifies positive and negative instances of object appearances in the video stream and uses these instances to update the information captured by the decision trees. Classifier-based object tracking is robust to appearance changes and to total occlusions; however, it is computation-