The video imagery gives important information to the dynamic analysis of human motion in the field of sports training or rehabilitation. For understanding the dynamics of human motion from TV or video image sequences, however, there are two complicated subjects. One is an image processing, for example, automated recognition of human's feature points such as head, elbow, knee etc.. The second is, how to estimate camera orientation parameters in the photogrammetry. The authors analyzed the sports dynamics, e. g. Carl Lewis and Boat Rowing using TV images. In these cases, the camera orientation were achieved by utilizing special information such as goal line, course lines and buoys which marks course line. However, the camera orientation for images without any information became important subjects. If CCD camera is mounted on the theodolite, ω andκcan be acquird as a vertical and horizontal angle, and φis regarded as 0 degree because the theodolite is leveled. Furthermore, these rotation parameters can be recorded as a image data by means of superimpose. With this in mind, the authors developed the video theodolite system which consist of CCD camera, theodolite and video recorder to estimate camera rotation parameters with real time and to record image data. This paper presents the application of video theodolite system for the dynamic analysis of human motion with sequential images.
A new contour/grid conversion method is proposed which is based on smoothing filter. By applying the smoothing filter to the contour image repeatedly, the image is gradually changed to target grid data. The implementation of this algorithm is extremely simple, because the algorithm only requires 3×3 masking operation. Also, the proposed method can be applied to contour image with cut lines. This feature is important in reducing the work burden because generating a complete contour image automatically is almost impossible and generally requires much efforts for experienced human operator. To compare the performance of the proposed method with previously reported methods, several algorithms which satisfy the simple implementation condition is actually implemented. In the experiment, the proposed algorithm showed best performance in maximum error. Also, it was most stable for contours with cut lines. Even though the result for average error was not the best, the error was 71cm for contours with lOm elevation interval, which is thought to be sufficient for practical use in many applications.
In order to repair pavements systematically, development of a method for inspecting pavement distress automatically is being hoped for. It is difficult, however, to classify crack types directly have two or more types of connected cracks. This necessitates segmentation of regions containing overlapping and connected cracks so that crack types can be classified properly. This paper proposes a method of segmenting crack region using crack elements extracted from images of pavement surface. The proposed method consists of three steps: (1) encording the orientation of connection using low-resolution images of pavement surface, (2) reencording by using pavement images of hierarchical resolutions, and (3) segmenting the reigions by tracing and merging the pixels encoded as linear regions. By applying the method to images of actual pavement surface, the author succeeded in identifying transverse and longitudinaly crack regions, two-dimensional crack regions, and isolated crack regions. The proposed region segmentation method has made it possible to apply conventional crack types classification techniques to actual pavement surface.
The paper briefly reviews the state of the art of researches on artificial intelligence and discusses their roll and effects on photogrammetry and remote sensing with a specific stress on the vision theories. The research trends can be classified into three : hardware-oriented researches including simulative and theoretical analyses of neural networks, realization of intellectual functions and knowledge-based systems. The purpose of an artificial vision is 3D representation of space from one or two (stereo) images. The paper reviews the history of pattern recognition, computational theory of vision and current topics of new image theories, and further discusses their applications to land cover classification by satellite images and solution of the stereo matching problem as combinatorial optimization.