Research On Hierarchical Approaches Based Complex Human Action Recognition

Human action recognition aims at enabling computers to recognize and read body actions involved in images and videos. It requires computers of visual and comprehensive abilities. Due to the limited technical and theoretical level at present, human action recognition still remains at its primary stage and just can achieve some simple tasks resulting in limited development in fields like human-computer interaction, smart home and content retrieval.A recognition system is presented in this essay, which is combined of three layers. The first layer is atomic action recognition using deep learning (stacked denoising auto-encoder). In first layer, recognise body contours in image preprocessing stage, mark it as interested area and operate zooming in or out, deduct data dimensions presenting pyramid network node optional program cutting down the budget. Update the learning rate via the mean square error to improve accuracy. The second layer is sequential action recognition using multiple observation which producted by first layer to build hidden Markov models. The third layer is complicate or interactive activity recognition which functions as describing in context free grammar to recognise. As to observation segmentation algorithm, original backward-looking forward algorithm is improved, continuous similar action judgment rules is applied, HMM Evaluation Demarcation is proposed simplifying segmentation process and increasing fault tolerance. Some logical predicates is added to context free grammar description for applying for much more complicate describe. At last, the whole system is analyzed through an experiment stating its feasibility and expansibility.In this thesis, we use Kinect equipment to gather data, providing deep skeleton image and improving the efficien. Compared with traditional human action recognition, the latest deep learning technique is used to improve atomic action recognition, paving the way for further study. Description approaches of context free grammar makes recognition process more flexible and makes complicate recognition possible.Our method belong to hierarchical approaches. Each algorithms of each layer can be instead by similar algorithms or rules. So the system is flexible and extensible. Under the high accuracy recognized by low layer, the higher layer can achieve good precision by a few training set.