:: ICCAS 2015 ::

WB04		Object/Human Visual Recognition
Time : 13:30~15:00 Room : 104		Chair : Hiroshi Harada (Kumamoto University, )



13:30~13:45	WB04-1

Sliding Window based Landmark Extraction for Indoor 2D SLAM Using Forward-Looking Monocular Camera
Heewon Chae, Chansoo Park, Jae-Bok Song(Korea University, Korea, Republic of )

In this study, we propose a method to extract landmarks for pose-graph SLAM using only the forward-looking monocular camera mounted in a small mobile robot. By using the sliding window approach, the region of interest is determined to triangulate the landmarks. The landmark extraction process uses multiple sets of bearing information for triangulation. To test its accuracy, the laser rangefinder is used as a ground truth. Also the various sliding window sizes are tried for a landmark to investigate the effect of the window size on the accuracy.



13:45~14:00	WB04-2

Parallel optical flow estimation by dividing image in section
Yusuke Miyajima, Teruo Yamaguchi, Hiroshi Harada(Kumamoto University, Japan)

In this study, we evaluated consuming time of the optical flow estimation when calculating in parallel. The method of parallelization includes dividing input image into several region and assigning each region into each multi-core. Because spatio-temporal differentiation method is composed of a local calculation, it is compatible with the parallelization method. It was found feasible to accelerate the optical flow estimation by the method. Experimental result shows that the most suitable acceleration can be achieved when dividing calculation region into parts of the same number of the cores.



14:00~14:15	WB04-3

Implementation of Optical Flow Measurement System with an Embedded Processor
Yukihiro Sugiki, Teruo Yamaguchi, Hiroshi Harada(Kumamoto University, Japan)

In this research, process time of optical flow calculation with Raspberry Pi was evaluated. Calclation time of the program based on spatio-temporal differentiation method was compared with those using optical flow function in OpenCV. As a result, the program was about the same speed as HS method, and took about six times as long as LK method. The program had more right direction than LK and HS method. When an object moves at high speed, output had wrong velocity vectors. Process time will be improved by GPU computing. Estimation accuracy need to be improved with compensation method.



14:15~14:30	WB04-4

Moving object analysis with optical flow using multi-resolution images and its application in traffic environment
Yuta Eto, Teruo Yamaguchi, Hiroshi Harada(Kumamoto University, Japan), Takeshi Tsusue(Mitsubishi Electric Corporation, Japan)

Conventional spatiotemporal differentiation scheme enables us to calculate velocity distribution rapidly, but error in the approximation of derivative coefficients increases as the displacement of the moving pattern between the successive frames becomes large. So we introduced the compensation method and realized how to determine the compensation velocity using multiple resolution images. We applied this method to images captured by an in-vehicle camera in order to calculate velocity distribution in situation where objects of various velocities exist and to locate traffic obstacles.



14:30~14:45	WB04-5

Detection of 3-Dimensional Textured Object without Depth Information
Sangil Lee, H. Jin Kim(Seoul Nat’l University, Korea, Republic of )

We propose a feature-based approach for detection of a 3-dimensional textured object in a single image. Each object has a training set which contains various poses in the template of small size. And then multiple objects are placed in the background, some of which are mixed with untrained objects. Background of some images is complicated and textured, which makes it difficult to find target object. In this experimental environment, the proposed algorithm finds the textured object comparing with the training template set we construct.