Kin Hong Wong

Script recognition using hidden Markov models

A handwritten script recognition system is presented which uses Hidden Markov Models (HMM), a technique widely used in speech recognition. The script is encoded as templates in the form of a sequence of quantised inclination angles of short equal length vectors together with some additional features. A HMM is created for each written word from a set of training data. Incoming templates are recognised by calculating which model has the highest probability for producing that template. The task chosen to test the system is that of handwritten word recognition, where the words are digits written by one person. Results are given which show that HMMs provide a versatile pattern matching tool suitable for some image processing tasks as well as speech processing problems.

Recursive Camera-Motion Estimation With the Trifocal Tensor

In this paper, an innovative extended Kalman filter (EKF) algorithm for pose tracking using the trifocal tensor is proposed. In the EKF, a constant-velocity motion model is used as the dynamic system, and the trifocal-tensor constraint is incorporated into the measurement model. The proposed method has the advantages of those structure- and-motion-based approaches in that the pose sequence can be computed with no prior information on the scene structure. It also has the strengths of those model-based algorithms in which no updating of the three-dimensional (3-D) structure is necessary in the computation. This results in a stable, accurate, and efficient algorithm. Experimental results show that the proposed approach outperformed other existing EKFs that tackle the same problem. An extension to the pose-tracking algorithm has been made to demonstrate the application of the trifocal constraint to fast recursive 3-D structure recovery

Pose estimation using four corresponding points

Abstract In this paper, a novel model-based pose estimation system is developed. In this system, correspondences between four points on a rigid solid object and four points in an image are used to estimate the new position and orientation of that rigid object under full perspective projection. Synthetic and real data have been used to verify our method.

Robust 3-D Motion Tracking From Stereo Images: A Model-Less Method

Traditional vision-based 3-D motion-estimation algorithms require given or calculated 3-D models while the motion is being tracked. We propose a high-speed extended-Kalman-filter-based approach that recovers camera position and orientation from stereo image sequences without prior knowledge, as well as the procedure for the reconstruction of 3-D structures. Empowered by the use of a trifocal tensor, the computation step of 3-D models can be eliminated. The algorithm is thus flexible and can be applied to a wide range of domains. The twist motion model is also adopted to parameterize the 3-D motion. It is minimal since it only has six parameters as opposed to seven parameters in quaternion and 12 parameters in matrix representation. The motion representation is robust because it does not suffer from singularities as Euler angles. Due to the fact that the number of parameters to be estimated is reduced, our algorithm is more efficient, stable, and accurate than traditional approaches. The proposed method has been applied to recover the motion from stereo image sequences taken by a robot and a handheld stereo rig. The results are accurate compared to the ground truths. It is shown in the experiment that our algorithm is not susceptible to outlying point features with the application of a validation gate.

An efficient iterative pose estimation algorithm

A novel model-based pose estimation algorithm is presented which estimates the motion of a three-dimensional object from a image sequence. The nonlinear estimation process within iteration is divided into two linear estimation stages, namely the depth approximation and the pose calculation. In the depth approximation stage, the depths of the feature points in three-dimensional space are estimated. In the pose calculation stage, the rotation and translation parameters between the estimated feature points and the model point set arer calculated by a fast singular value decomposition method. The whole process is executed recursively until the result is stable. Since both stages can be solved efficiently, the computational cost is low. As a result, the algorithm is well-suited for real computer vision applications. We demonstrate the capability of this algorithm by applying it to a real time head tracking problem. The results are satisfactory.

Automatic lyrics alignment for Cantonese popular music

From lyrics-display on electronic music players and Karaoke videos to surtitles for live Chinese opera performance, one feature is common to all these everyday functionalities temporal: synchronization of the written text and its corresponding musical phrase. Our goal is to automate the process of lyrics alignment, a procedure which, to date, is still handled manually in the Cantonese popular song (Cantopop) industry. In our system, a vocal signal enhancement algorithm is developed to extract vocal signals from a CD recording in order to detect the onsets of the syllables sung and to determine the corresponding pitches. The proposed system is specifically designed for Cantonese, in which the contour of the musical melody and the tonal contour of the lyrics must match perfectly. With this prerequisite, we use a dynamic time warping algorithm to align the lyrics. The robustness of this approach is supported by experiment results. The system was evaluated with 70 twenty-second music segments and most samples have their lyrics aligned correctly.

Recursive three-dimensional model reconstruction based on Kalman filtering

A recursive two-step method to recover structure and motion from image sequences based on Kalman filtering is described in this paper. The algorithm consists of two major steps. The first step is an extended Kalman filter (EKF) for the estimation of the objects pose. The second step is a set of EKFs, one for each model point, for the refinement of the positions of the model features in the three-dimensional (3-D) space. These two steps alternate from frame to frame. The initial model converges to the final structure as the image sequence is scanned sequentially. The performance of the algorithm is demonstrated with both synthetic data and real-world objects. Analytical and empirical comparisons are made among our approach, the interleaved bundle adjustment method, and the Kalman filtering-based recursive algorithm by Azarbayejani and Pentland. Our approach outperformed the other two algorithms in terms of computation speed without loss in the quality of model reconstruction.

A system for recognising human faces

A human face recognition system, which has a feature extraction subsystem and a recognition subsystem, has been built. An unknown human facial image is captured and digitized by a frame grabber, and features of the digitized image are extracted. The features are stored during the learning phase and are compared with the references in the memory during the recognition phase. A two-test method is used to increase the accuracy of the recognition. The first test eliminates a number of records with large differences in major features. The second test selects the one with minimum differences by calculating the total weighted error. Test results are presented.<<ETX>>

Pose estimation for augmented reality applications using genetic algorithm

This paper describes a genetic algorithm that tackles the pose-estimation problem in computer vision. Our genetic algorithm can find the rotation and translation of an object accurately when the three-dimensional structure of the object is given. In our implementation, each chromosome encodes both the pose and the indexes to the selected point features of the object. Instead of only searching for the pose as in the existing work, our algorithm, at the same time, searches for a set containing the most reliable feature points in the process. This mismatch filtering strategy successfully makes the algorithm more robust under the presence of point mismatches and outliers in the images. Our algorithm has been tested with both synthetic and real data with good results. The accuracy of the recovered pose is compared to the existing algorithms. Our approach outperformed the Lowes method and the other two genetic algorithms under the presence of point mismatches and outliers. In addition, it has been used to estimate the pose of a real object. It is shown that the proposed method is applicable to augmented reality applications.

A projector-based movable hand-held display system

In this paper, we proposed a movable hand-held display system which uses a projector to project display content onto an ordinary cardboard which can move freely within the projection area. Such a system can give users greater freedom of control of the display such as the viewing angle and distance. At the same time, the size of the cardboard can be made to a size that fits ones application. A projector-camera pair is calibrated and used as the tracking and projection system. We present a vision based algorithm to detect an ordinary cardboard and track its subsequent motion. Display content is then pre-warped and projected onto the cardboard at the correct position. Experimental results show that our system can project onto the cardboard in reasonable precision.