FUSE Takashi The University of Tokyo, Graduate School of Engineering, Research Associate, 大学院・工学系研究科, 助手 (80361525)
SHIMIZU Tetsuo The University of Tokyo, Graduate School of Engineering, Associate Professor, 大学院・工学系研究科, 助教授 (40272679)
TSUTSUMI Morito University of Tsukuba, Institute of Policy & Planning, Sciences, Associate Professor, 社会工学系, 助教授 (70292886)
|Budget Amount *help
¥2,800,000 (Direct Cost: ¥2,800,000)
Fiscal Year 2003: ¥900,000 (Direct Cost: ¥900,000)
Fiscal Year 2002: ¥1,900,000 (Direct Cost: ¥1,900,000)
Sequential images from high altitude platforms are suitable for traffic flow surveillance, which requires fixed-point continuous observation of vehicle movements over a wide area. High altitude platforms are expected to result in high spatial and time resolution images at specific areas for continuous observation. These high resolution and continuous images certainly make observation of vehicle movement easier. As a result, the high altitude platforms have a great potential with wider scope of utilization to derive useful traffic information on vehicle trajectories. In this study, we explore the possibility of vehicle maneuvers recognition with high resolution and time-serial aerial images, which are on the assumption of the use of helicopters and stratospheric platforms.
Specifically, we develop a new technique for vehicle maneuvers recognition. In accordance with human perception, we develop spatio-temporal clustering method, which is composed of geometric correction, background subtraction, shadow detection, optical flow extraction, clustering and labeling. Employing the two features, that are background subtraction value and optical flow, the pixels that are adjacent and have similar features are grouped. Consequently, vehicle clusters are formed in the spatio-temporal image.
The method, however, required adjustments to some parameters, depending on the images. In addition, the results were given in image coordinates (pixels) instead of real world coordinates. This research describes a method to adjust the parameters and a framework to evaluate the vehicle positioning accuracy using real distances.
The proposed framework is applied to real sequential images obtained from a helicopter, and the vehicle positioning accuracy is estimated. The sequential images have 10cm, 30 cm spatial resolution and 1/30 s time resolution. The framework provides exact and efficient vehicle recognition. The positioning accuracy of the recognized vehicles is better than 1.92m.