Bin Kong

Stereo Vision based Relative Pose and Motion Estimation for Unmanned Helicopter Landing

In this paper, we present a real-time stereo vision based pose and motion estimation system that will be used for landing an unmanned helicopter on a moving target such as a ship deck. The vision algorithm mainly consists of a feature extraction task and a pose and motion estimation task. The 2D planer target with regular features defined can significantly simplify the feature extraction task such as corner detection and feature points matching. To effectively estimate the distance between the camera carrier and target a stereo camera system is applied. By means of sub-pixel corner location the precisions of pose estimation and relative motion detection can be improved. We present results from semi-physical simulation which show that our vision algorithm is accurate and robust. The methodology provides an effective subsystem for the development of autonomous robot helicopter that will land on a given target under the guide of vision.

Calibration of linear structured light system by planar checkerboard

Linear structured light system is an increasing common means for acquiring three-dimensional geometry of objects due to its accuracy and robustness. Accuracy and less complexity of system calibration is a key challenge to future scanner designs. A calibration method by a planar checkerboard to minimize calibration complexity and cost is proposed. This method consists of two steps: the camera calibration with distortion and the projector calibration by plane properties. A simple system was designed and calibrated by this method, and a good 3D reconstruction quality is achieved by the calibrated system.

Eye state detection in facial image based on linear prediction error of wavelet coefficients

Eye state detection in facial image is a significant issue in face recognition, human-computer interface and driver fatigue monitoring system. In this paper, we first located the eye region in the upper area of the face region with AbaBoost algorithm. The linear predictor error distribution of wavelet coefficients was proposed as the statistics model to distinguish the eye states. We collected statistics (mean, variance, skewness, and kurtosis) of the prediction error distribution as eye state features. Build on these eye state features and support machine vector (SVM) with radial basis function (RBF) kernel a non-linear classifier is obtained by training samples of eye images. Experiment results with the classifier demonstrated that our method is an effective eye state detection approach which can satisfy various situations.

A Shooting Training and Instructing System Based on Image Analysis

The main shortcoming of the traditional way to train shooting athletes is that the effect is hard to be evaluated without live firing. We have designed and implemented a computer-aided shooting training and instructing system that can help coaches and athletes do more quality than quantity work. The minimum system consists of 6 components: a gun, a laser aiming device, a target, a camera, an image card and a computer. It works as follows: the camera snaps a target image, then the aiming point which is marked by the laser spot is extracted from the image and the corresponding score is calculated and printed in the screen in real time. The trace of the aiming point in a training session is also drawn in the screen. The system can prompt the best trigger moment. Based on the recorded performance data in the whole training session, the coach can make objective evaluation on the training effect and give pointed instructions to the athlete. The evaluation and instructions can be made by the expert system instead. Up to 16 targets can be processed simultaneously by one system.

Image Recognition Based on Wavelet Invariant Moments and Wavelet Neural Networks

A novel image recognition system is constructed by combining Wavelet invariant moments with Wavelet neural networks in this paper. Firstly, global and local features of the image can be obtained by using Wavelet invariant moments. Secondly, the invariant features are fed into Wavelet neural networks. Finally, supervised invariant pattern recognition can be achieved by utilizing three characters of Wavelet neural networks, which are the automatic ascertaining the number of hidden layer unit, converging rapidly and never running into the partial minimum of networks. The experiment results demonstrate that using Wavelet invariant moments and Wavelet neural networks can achieve higher accuracy of image classification than the algorithm based on normal invariant moments and BP neural networks.

Color Edge Detection Based on YUV Space and Minimal Spanning Tree

A novel color edge detection approach based on YUV space and minimal spanning tree is proposed, which treats the pixels of color images as vectors in YUV space, and uses the idea of minimal spanning tree in graph theory to process them. The experiments carried out show that the strengths of detected edges are clear and precise, reflecting the intensity of actual edges. The final edge results are obtained by automatic fast entropy thresholding algorithm.

Independent component analysis and its application in the fingerprint image preprocessing

Independent component analysis (ICA) is a new method of signal separation developed in recent years. In this paper, the fundamental theory and algorithm of ICA are introduced, and the implementation of ICA method in fingerprint image preprocessing is discussed to separate the fingerprint from background texture. ICA requires the number of observations should be no less than that of independent sources. So it is impossible to apply ICA to a single image directly. The paper presents a technique to generate three input signals from one single image, and then, process it by ICA. The experiment results illustrate that ICA has better performance than traditional methods.

An Improved Model of Producing Saliency Map for Visual Attention System

The iLab Neuromorphic Vision Toolkit (iINVT), steadily kept up to date by the group around Laurent Itti, is one of the currently best known attention systems. Their model of bottom up or saliency-based visual attention as well as their implementation serves as a basis for many research groups. How to combine the feature maps finally into the saliency map is a key point for this kind of visual attention system. We modified the original model of Laurent Itti to make it more corresponding with our perception.

STEREO VISION-BASED ESTIMATION OF POSE AND MOTION FOR AUTONOMOUS LANDING OF AN UNMANNED HELICOPTER

In this paper, a real-time stereo vision based pose and motion estimation system is presented. It is used for landing an unmanned helicopter on a moving target such as a ship deck. The vision algorithm mainly consists of a feature extraction task and a pose and motion estimation task. By the specially designed pattern of the landing target, the feature extraction algorithm can simplify the step of feature points matching of stereo system. In the task of feature extraction, the step of accurate corner detection can get to the precision of sub-pixel, which helps improve the measurement precision in state estimation. We present results from semi-physical simulation which show that our vision algorithm is accurate and robust to allow our vision sensor to be placed in the control loop of unmanned helicopter management system.

A simple calibration method for restricted camera estimation

In real applications, there is often the need of estimating the intrinsic and extrinsic parameters of a camera directly from the images of natural scenes or working scenarios. When there are only several feature points can be determined, the results of current calibration methods are very unstable. In this paper, a geometrical method is presented for estimating a restricted camera. It can estimate f even when only one feature point is valid. Experiments show that it has good robustness and reliability, and it is tolerant to multiple error sources.