İnsan hareketini anlama: İnsan aktivitelerinin tanınması ve erişimi

Within the ever-growing video archives is a vast amount of interesting informationregarding human action/activities. In this thesis, we approach the problem of extractingthis information and understanding human motion from a computer vision perspective.We propose solutions for two distinct scenarios, ordered from simple to complex. Inthe first scenario, we deal with the problem of single action recognition in relativelysimple settings. We believe that human pose encapsulates many useful clues for recognizingthe ongoing action, and we can represent this shape information for 2D singleactions in very compact forms, before going into details of complex modeling. Weshow that high-accuracy single human action recognition is possible 1) using spatialoriented histograms of rectangular regions when the silhouette is extractable, 2) usingthe distribution of boundary-fitted lines when the silhouette information is missing.We demonstrate that, inside videos, we can further improve recognition accuracy bymeans of adding local and global motion information. We also show that within a discriminativeframework, shape information is quite useful even in the case of humanaction recognition in still images.Our second scenario involves recognition and retrieval of complex human activitieswithin more complicated settings, like the presence of changing background andviewpoints. We describe a method of representing human activities in 3D that allowsa collection of motions to be queried without examples, using a simple and effectivequery language. Our approach is based on units of activity at segments of the body,that can be composed across time and across the body to produce complex queries.The presence of search units is inferred automatically by tracking the body, lifting thetracks to 3D and comparing to models trained using motion capture data. Our modelsof short time scale limb behaviour are built using labelled motion capture set. Our query language makes use of finite state automata and requires simple text encodingand no visual examples. We show results for a large range of queries applied to acollection of complex motion and activity. We compare with discriminative methodsapplied to tracker data; our method offers significantly improved performance. Weshow experimental evidence that our method is robust to view direction and is unaffectedby some important changes of clothing.