Hand Segmentation from RGB Images in Uncontrolled Indoor Scenarios Using Randomized Decision Forests
Abstract: Hand segmentation is an integral part of many computer vision applications suchas gesture recognition, human activity detection, hand-eye coordination anal-ysis, gaze detection etc. Many of these applications require a solution whichcan segment hands in real-time while working in low-capacity computing en-vironments. Training a classier to classify pixels into hand or background isone of the most popular methods used by researchers in the past. Many stud-ies have pointed out that using a Randomized Decision Forest (RDF) basedclassier brings out the best segmentation performance. However, the test sce-narios used by these studies were restricted in many aspects such as unclutteredimage backgrounds, controlled lighting and/or restricted variation of hand orbackground in images. Additionally, only a limited set of image features wereutilized by these methods to extract properties associated with the pixels.In this thesis, the performance of RDF based pixel classiers in hand seg-mentation when used in uncontrolled indoor environments is evaluated. A pairof new image features were devised to extract information from the neighbor-hoods around pixels and new implementation methods were forged for someexisting image features to make them faster. Seven image features, extractingproperties of pixels such as color and texture, as well as properties of pixelneighborhoods such as dierences, histogram, statistics and probabilities weretested with RDFs for hand segmentation. Additionally, as datasets with imagesof hands in uncontrolled indoor environments were nonexistent, a new datasetnamed ManoHandDB was created through manual annotation. The segmenta-tion performance was optimized by ne tuning the parameters of each imagefeature as well as the parameters of RDF. It was found that using a combinationof the color, texture, neighborhood histogram and neighborhood probability fea-tures outperforms existing methods for segmenting images in restricted as wellas unrestricted indoor environments. The experiments also shed light on theadvantages of each feature over the others as well as on the dependency ofthe feature on the training dataset qualities such as variations in background,illumination, hand scale and visible hand parts.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)