Yohei Ishii

Face and head detection for a real-time surveillance system

This paper describes a face and head detection method for a real-time surveillance system. Since there is no guarantee that surveillance cameras can capture frontal face or full-body of human, face and head detection has an advantage for the practical use. Proposed method employs four directional features and linear discriminant analysis. It can detect face and head simultaneously, and reduce computation cost. In the experiments, comparison of two classifiers and evaluation of proposed human detection method were performed using still images and video scenes. The results showed that the performances of two classifiers were almost equivalent. Thus, the classifier labeled face samples to one class was better in terms of computation cost. In the human detection experiment, the results were 87.2% (48/55) for human detection rate, and 83.6% (832/995) for reliability of detection. The proposed detection method was implemented on a PC and run at approximately over 10 fps for VGA input with motion detection.

Real-time face and head detection using four directional features

Real-time human detection is an important part in a surveillance system using computer vision. In this paper, a real-time face and head detection method is proposed for such human detection. The method has an advantage of detecting peoples who are not facing a camera, by detecting their heads. It employs four directional features (FDF) and linear discriminant analysis in order to save computation cost for scanning and classification. Since FDF represents edge directional information in low resolution, it is resistive to changes in lighting conditions and scales. The proposed method was evaluated through an experiment using 26 video scenes. The results of experiment were 83.6% (46/55) for human detection rate, and 84.6% (1048/1239) for reliability of detection. The human detection system implemented with the proposed method runs on a PC at approximately over 10 fps for VGA input with motion detection.

Precise Top View Image Generation without Global Metric Information

We describe a practical and precise calibration method for generating a top view image that is transformed so that a planar object such as the road can be observed from a direction perpendicular to its surface. The geometric relation between the input and output images is described by a 3×3 homography matrix. Conventional methods use large planar calibration patterns to achieve precise transformations. The proposed method uses much smaller element patterns that are placed in arbitrary positions within the view of the camera. One of the patterns is used to obtain an initial homography. Then, the information from all of the patterns is used by a non-linear optimization scheme to reach a global optimum homography. The experiment done to evaluate the method showed that the precision of the proposed method is comparable to that of the conventional method where a large calibration pattern is used, making it more practical for automotive applications.

A Practical Calibration Method for Top View Image Generation

This paper describes a practical calibration method to generate a top view image from an image of a rear view camera for automobile. The geometric relation between the input and output images is described by a 3 x 3 homography matrix. Conventional methods have to use a large planar pattern to achieve precise calibration. The proposed method uses much smaller patterns, and enables to place those patterns on arbitrary positions within the view of the camera. The experimental results show that the method has equivalent calibration precision with much easier setting than the conventional methods.

Comparison of different methods for gender estimation from face image of various poses

Recently, gender estimation from face images has been studied for frontal facial images. However, it is difficult to obtain such facial images constantly in the case of application systems for security, surveillance and marketing research. In order to build such systems, a method is required to estimate gender from the image of various facial poses. In this paper, three different classifiers are compared in appearance-based gender estimation, which use four directional features (FDF). The classifiers are linear discriminant analysis (LDA), Support Vector Machines (SVMs) and Sparse Network of Winnows (SNoW). Face images used for experiments were obtained from 35 viewpoints. The direction of viewpoints varied ±45 degrees horizontally, ±30 degrees vertically at 15 degree intervals respectively. Although LDA showed the best performance for frontal facial images, SVM with Gaussian kernel was found the best performance (86.0%) for the facial images of 35 viewpoints. It is considered that SVM with Gaussian kernel is robust to changes in viewpoint when estimating gender from these results. Furthermore, the estimation rate was quite close to the average estimation rate at 35 viewpoints respectively. It is supposed that the methods are reasonable to estimate gender within the range of experimented viewpoints by learning face images from multiple directions by one class.