Research on recognition mechanisms of low-quality images for multi-viewpoint video surveillance

随着信息和通信技术的发展，众多的视觉传感器，如配备摄像头的移动的手机，车载摄像头，监控摄像头，现在无处不在的现实世界的环境中。 因此，需要通过利用从这些传感器获得的数据来实现安全和愉快的社会的监视技术。然而，用小型相机设备从真实世界获得的图像具有诸如分辨率不足以及由于照明条件和不同视点而导致的对象外观的变化的问题。该项目旨在开发一种方法，通过整合从不同视点和/或在不同时间获得的多模态信息，准确地识别即使是人眼也无法从单个图像中识别的“极低质量图像”。以下是该项目的实际成果（1）提出了低质量字符识别的方法，该方法通过鲁棒统计分析从连续帧中获得联合收割机相似性，并合成退化字符用于训练（2）提出了根据待识别对象的类别确定多个摄像机的最佳布置的方法。提出了一种在训练样本较少的情况下识别效果较好的方法。(3)提出了车载摄像机的方法，例如基于在不同时间获得的街景视频数据检测街景变化，从劣化视频数据识别交通标志，以及基于行驶汽车前护罩上的雨滴检测的天气识别。

As the Information and Communication Technologies develop, numerous visual sensors such as camera-equipped mobile phones, car-mounted cameras, surveillance cameras now exist ubiquitously in the real-world environment. Accordingly, surveillance technologies that realize a safe and pleasant society by making use of data obtained from these sensors are needed. Images obtained from the real-world with small camera devices, however, have problems such as insufficient resolutions, and the variance of the appearance of objects due to lighting conditions and different viewpoints. This project aimed at developing a method that accurately recognizes from "Very-low quality image" that even human eye& cannot recognize from a single image, by integrating multimodal information obtained from different viewpoints and/or at different timings. The following are the actual fruits of the project (1) Proposed methods for low-quality character recognition, that combine similarities obtained from consecutive frames by robust statistical analysis, and that synthesizes degraded characters for training (2) Proposed a method for deciding optimal arrangements of multiple cameras according to the class of objects to be recognized. Proposed also a method that recognizes well with a small number of training samples. (3) Proposed methods for car-mounted cameras such as detection of streetscape changes based on streetscape video data obtained at different timings, recognition of traffic signs from degraded video data, and weather recognition based on raindrop detection on the front shield of a running car.