Recognition of Human Motion and Form
thesisposted on 22.02.2015, 00:00 authored by Qingquan Wu
In this report, we first compare the novel RISq method to HMM. Both methods are used for the recognition of general vector sequences. Our comparison shows that RISq performs better than HMM in many aspects. The training of RISq requires only one example from each class and the training is much simpler than training HMM models. Also a few sparse samples from a test sequence are usually sufficient for RISq to achieve robust recognition while HMM needs the entire sequence. This makes RISq essentially much less sensitive to missing vectors. Lastly, our experiments demonstrate that RISq outperforms HMM in term of handling large sets of models or with many vectors dimensions. RISq is also better in noise robustness, computation time, selectivity ratio, etc. For human gesture recognition, we develop a novel system to recognize different hand gestures. We use an Inertial Measurement Unit equipped with accelerometers and gyroscopes to sense the motion of the operator's hand. The IMU is calibrated with the help of Nonlinear Data-Fitting method. Gesture trajectories are reconstructed from inertial sensor measurements using the Inertial Navigation System theory. We develop a novel method named Zero Velocity Linear Compensation (ZVLC) to improve trajectory reconstruction accuracy. Experimental results show that ZVLC provides more accurate reconstruction than the widely used method of Zero Velocity Compensation. At the recognition stage, the novel RISq method is employed to recognize the reconstructed gesture trajectories and achieves a recognition rate of 92%. In the third part of this thesis, we describe a novel method for 3D head reconstruction and view-invariant recognition, which is based on Shape From Shading combined with Hybrid Principal Component Analysis (HPCA). Our novel HPCA algorithm provides good initial estimates of 3D range mapping for the SFS optimization and yields much improved 3D reconstruction. Additional contribution of our chapter is the successful handling of variable and unknown surface albedo in SFS. Experimental results show that our HPCA based SFS method provides accurate 3D head reconstructions and high recognition rates. Our work could have many practical applications such as person recognition from side views when only frontal views are available for modeling.