法工学鑑定のための画像の幾何解析プログラムの開発

抄録

 The importance of image analysis in the forensic science field has been increasing rapidly alongside the increased usage of security cameras. Specifically, the analysis of geometric information, such as position, length, and shape (i.e., vehicle speed measurement, height measurement, and gait analysis, respectively), is in high demand. In this paper, we aim to understand the mathematical principles of geometric image analysis and to expand its range of applications by using data-driven analysis. We created a set of image geometric analysis programs for legal engineering appraisals related to the following two topics: 1) lens distortion correction, and 2) analysis of the perspective projection image (s).

 Regarding lens distortion correction, an existing method of analyzing a single-parameter model from a single straight line has been extended to a data-driven framework that analyzes multi-parameter models from multiple lines. As a result, high-precision distortion correction has become possible by expanding the model even when lens radial distortion is strong and the distortion center does not coincide with the image center. Furthermore, even when the region of interest (ROI) is located near the periphery of the image, the information loss due to resizing and/or multiple interpolations can be avoided by performing one interpolation with translation and enlargement.

 Regarding the analysis of the perspective projection image (s), we revisit classical analyses that use direct linear transform (DLT)-based projection matrix estimation and Gauss-Newton (GN)-based external parameter estimation. In the case of DLT, it can be understood that there is a unified approach for both 2D-2D correspondence analysis (planar projective transformation) and 2D-3D correspondence analysis (3D camera calibration). The analysis based on line correspondence, which has not yet been utilized in the Japanese forensic field, is introduced. By experimentally comparing the three methods based on point correspondence, line correspondence, and both point and line correspondences, respectively, it was revealed that the combined approach tended to be more precise. In the case of three-dimensional-position estimation, where it was pointed out that analysis can be possible even when the number of correspondences is small, and that the field-work procedure is not necessary in the case of relative position estimation. It was shown that multi-view-based 3D human surface reconstruction is possible even in the case of low surveillance camera resolution and non-synchronized settings.