Abstract
In this thesis, we propose a human motion analysis system by using two cameras to capture the facade view and the flank view to overcome the occlusion. Using image sequences acquired from 2 views, we recover the 3-D body pose at each time instant without the use of markers. After background subtraction, we partition our system into two phases: feature extraction and recognition phase. To estimate the motion parameters (BAPs) of the human actor, we use the 3-D articulated human model. To track the human actor, we find the best match between the articulated human model and the binary human object to get the joint angles. The vectors of the joint angles are the descriptions of the posture in each frame. We propose a 3-level BAP estimation algorithm which consists of low-level analysis, mid-level estimation and high-level verification and correction. Moreover, to relax the limitations of using the articulated human model, we predefine some auxiliary postures, and develop a non-causal projection profile analysis to recognize the postures of bending and rotating. Finally, we use the estimated BAPs to train HMMs and recognize the human motions. We first train our system using HMM approach and then use the trained HMMs to recognize the input sequence. The test sequence of human motion to be recognized in separately scored against different HMMs. The model with the highest score is selected as the recognized human motion. In our experiment, our system can analyze 15 different human motions