To increase the efficiency of golf training, 3D swing reconstruction is broadly used among golf researchers. Traditional reconstruction methods apply motion capture system (MOCAP) to gain golfers motion data and drive bio-mechanical model directly. The cost of MOCAP system restricts the application area of golf research and the reconstruction quality of swing relies on the accuracy of the motion data. We introduced the dynamical analysis into swing reconstruction and proposed a Dynamic Bayesian Network (DBN) model with Kinect to capture the swing motion. Our model focused on modeling the bio-mechanical and dynamical relationships between key joints of golfer during swing. The positions of key joints were updated by the model and were used as motion data to reconstruct golf swing. Experimental results show that our results are comparable with the ones acquired by optical MOCAP system in accuracy and can reconstruct the golf swing with much lower cost.

Foreground detection is an important part in video surveillance system. The detection results will significantly affect the performance of tracking, abnormal behavior analysis and other following procedures. Many algorithms have been proposed to improve the detection performance. However, these algorithms simply focus on one single frame, ignoring the relationship among the detection results of one target in successive frames. This paper presents a novel foreground enhancement algorithm using Hidden Markov Model (HMM). In a video sequence, one target in successive frames usually has similar shape, size, et al. With this property, the target can be modeled by HMM and enhanced using the result of its prior frame. The observation of HMM is obtained by ViBe. The enhancement result is then estimated by using Maximum A Posteriori (MAP). Experimental results show that compared with the state-of-art algorithm, the proposed method can enhance foreground detection effectively.

The orientation estimation is a critical technique in inertial sensor based motion capture systems. One challenge of the orientation estimation is that it suffers from the acceleration interference due to body segment motion, especially when the acceleration interference is significant. In this paper, we propose a quaternion based orientation estimation algorithm using unscented Kalman filter. In the algorithm, the acceleration interference is taken as an element of the state vector and estimated in the algorithm together with the orientation quaternion, knowing that the acceleration interference can be predicted based on the rotational angular velocity. The experiments were conducted using both computer simulation and in real-world motion scenarios. Both experimental results have shown the effectiveness of the proposed orientation estimation algorithm.