Motion Analysis And Recognition Based On Feature Mapping

In the feld of Pattern Recognition and Machine Learning, generative models anddiscriminative models are two important paradigms. Feature mapping, on the otherhand, provides a method to describe the feature of the samples from generative mod-els, then combined with discriminative models. This kind of hybrid generative anddiscriminative method, could not only possess the merits of generative models in mod-eling data and exploiting hidden information, but also could take the advantage of thegreat power of discriminative models in capturing the boundaries between categories.Thus, in motion analysis and recognition problem of Computer Vision feld, such ashuman action recognition, scene recognition, face recognition, crowds analysis, etc.feature mapping has great potential in research and application.Based on a large amount of articles, the author analyzed the existed feature map-pingmethods, thenextendedthem, andproposedamethodtoobtainmulti-classfeaturemapping. Then in three challenging problems in Computer Vision feld,3D human ac-tion recognition, crowds analysis and one example face recognition, obtained somevaluable experiments results by combining the feature mappings from generative mod-els and discriminative models or clustering method.(1)Inthe3Dhumanactionrecognitionproblem,weproposedarecognitionmethodon3D human joint space based on feature mapping. This method relies on HiddenMarkov Model (HMM), but difers from the previous methods in the way of incor-porating HMM and discriminative classifer, aiming to capture more discriminativeinformation. We break down the human joint sequences into several semantic parts,the model diferent parts separately, then we got the feature mapping through multi-class Posterior Divergence, fnally obtained the entire feature mapping. The derived feature mappings map a variable-length joint sequence to a fxed-dimension featurevector which would be delivered to SVM for classifcation. We evaluate the proposedmethod and related methods on a large number of3D joint sequences. The experimen-tal results show its competitive performance, in comparison with other state-of-the-artmethods.(2)Inthecrowdsanalysisproblem, weproposedamethodofexploitingthehiddeninformation based on feature mapping. This method relies on the tracklets extractedfrom the video, and perceives crowd analysis as a clustering process in probabilisticway, and proved to be robust. Through the use of feature mapping, we could exploitingthe hidden information of long trajectories, which is more efective and discrimina-tive in clustering, comparing with observed information. The experiment results aresatisfactory using this method.(3)In the one example face recognition problem, we proposed a novel parts-baseddata representation for robust face recognition, MSR. This method introduces the con-cept of multiple subcategories into a probabilistic generative model to mimic the pro-cess of generating a face image. Training of the model is totally unsupervised. Oncethe training is completed, a test example from the face class may be recognized assimply a novel combination of learned parts, and the example is mapped on these submodels. The feature mapping is then recognized through similarity metrics. Due to theMSR generative model, the feature mapping extracted encodes hidden discriminativestructured information. This method is robust to varied lighing and illumination andocclusion.