To effectively improve the low contrast of human body region in the infrared images, a combing method of several enhancement methods is utilized to enhance the human body region. Firstly, for the infrared images acquired by Kinect, in order to improve the overall contrast of the infrared images, an Optimal Contrast-Tone Mapping (OCTM) method with multi-iterations is applied to balance the contrast of low-luminosity infrared images. Secondly, to enhance the human body region better, a Level Set algorithm is employed to improve the contour edges of human body region. Finally, to further improve the human body region in infrared images, Laplacian Pyramid decomposition is adopted to enhance the contour-improved human body region. Meanwhile, the background area without human body region is processed by bilateral filtering to improve the overall effect. With theoretical analysis and experimental verification, the results show that the proposed method could effectively enhance the human body region of such infrared images.
In a binocular visual system, to recover the three-dimensional information of the object, the most important step is to acquire matching points. Structure tensor is the vector representation of each point in its local neighborhood. Therefore, structure tensor performs well in region detection of local structure, and it is very suitable for detecting specific graphics such as pedestrians, cars and road signs in the image. In this paper, the structure tensor is combined with the luminance information to form the extended structure tensor. The directional derivatives of luminance in x and y directions are calculated, so that the local structure of the image is more prominent. Meanwhile, the Euclidean distance between the eigenvectors of key points is used as the similarity determination metric of key points in the two images. By matching, the coordinates of the matching points in the detected target are precisely acquired. In this paper, experiments were performed on the captured left and right images. After the binocular calibration, image matching was done to acquire the matching points, and then the target depth was calculated according to these matching points. By comparison, it is proved that the structure tensor can accurately acquire the matching points in binocular stereo matching.
This paper presents a new depth measuring method for the dual-view stereo camera based on the converted relative
extrinsic parameters. The relative extrinsic parameters between left and right cameras, which obtained by the stereo
camera calibration, can indicate the geometric relationships among the left principle point, right principle point and
convergent point. Furthermore, the geometry which consists of the corresponding points and the object can be obtained
by making conversion between the corresponding points and principle points. Therefore, the depth of the object can be
calculated based on the obtained geometry. The correctness of the proposed method has been proved in 3ds Max, and the
validity of the method has been verified on the binocular stereo system of flea2 cameras. We compared our experimental
results with the popular RGB-D camera (e.g. Kinect). The comparison results show that our method is reliable and
efficient, without epipolar rectification.
The spatial position of convergent point of dual-view stereo camera is a key parameter. To solve the problem that lack of simple and effective convergent point positioning method at present, we present two methods of convergent point positioning. The first method for convergent point positioning is by observing the difference between the corresponding points of principal points in left and right images. The second method is by computing the relative extrinsic parameters between right and left cameras. The experimental results show that the first method is convenient for the stereo camera which consists of adjustable left and right cameras; the second method is convenient for the stereo camera which consists of stable left and right cameras. Both of the methods are available for convergent point positioning.