Complex Human Activity Recognition Based On SCFG

Human activity recognition is an important issue of image processing and pattern recognition. This thesis on the study of human activity recognition is based on several problems which are human activity segmentation, estimation of joint point positions of the human body, human action recognition and human activity recognition. Our contribution is concluded in the following aspects.The method of two-steps activity segmentation based on slow feature analysis (SFA) is proposed. For reducing the computation of SFA and increasing the activity segmentation speed, the activity is initially spitted according to the difference of horizontal speed and vertical speed of the silhouette border of the moving object. The further activity segmentation uses SFA to spit activities. The initial window and the stopping formulas of optimal solution calculation are proposed against the high computation and slow segmentation of SFA, which can quickly spit human activity online.The method of automatic estimation of joint points is proposed. The human silhouettes of the object are represented by the vector of the end nodes of the skeletal graph matched by DTW in the data set are used to estimate the ones in the given video. In the following images, the SIFT feature matching is used to estimate the joint points. Kalman filter is used to estimate the lost joint points which are occluded. The joint points estimation method we proposed is not computationally intensive, which is suitable for common situation.The approach of action recognition based on continuous-space relevance model (CRM) is proposed. For solving the problem that the parameter and topic of topic model is not guaranteed, CRM is utilized to recognize human actions characterized by human joint point trajectories. Human action recognition based on CRM uses non-parameters Gaussian kernel density estimator, which avoids parameter setting and parameter estimating and improves the action recognition accuracy.The two-level human activity recognition structure is proposed. Low-level activity analysis extracts the joint point trajectory and uses CRM to recognize human action. High-level activity analysis based on the recognition results of low level activity analysis uses stochastic context-free grammars (SCFGs) to recognize human activity. The proposed algorithm has been trained and tested on the surveillance video. The results showed that it could recognize single-person activity effectively.