Showing papers by "Thanarat H. Chalidabhongse published in 2012"

The region-based distance of oriented gradient and motion direction for human action classification

Abstract: Human action in the image sequence can be seen as the relation of the movement of body parts. Since, human has an articulated body, each body part cannot move freely. In each action, the specific directions of body parts arrangement cause a change in posture and movement over time. In this paper, the image oriented gradient and histogram of motion direction are used to describe the posture and movement information in a bounding box. We use distances between regions in a bounding box which contains the whole body to find the relations of the oriented gradient and motion direction over time. The cosine distance is used to measure the similarity of direction histograms. These features are combined and concatenated with the previous frames to construct a feature vector. Then, K-nearest neighbor is used to classify actions in frame by frame. We test the system performance with a human action dataset and a dataset from a surveillance camera. The system can classify action in human action dataset for each frame and can detect the anomaly from the surveillance dataset.

Head pose estimation using motion subspace matching on GPU

Abstract: The head pose estimation is a process of recovering 3D head position in term of yaw, pitch and roll from 2D images. However, the reduction of information from 3D to 2D leads to an ill-posed problem. In this paper, we propose a novel algorithm of head pose estimation that includes facial features tracking for Thai sign language recognition. In order to estimate head pose correctly, feature points tracking requires high precision. Nevertheless it is difficult for low cost cameras where input image quality may be generally poor. To overcome this problem, we introduce an automatic camera signal calibration such that the features can be tracked correctly despite the quality of the input image sequences. Finally, as our approach bases on the state space searching, the local minima problem is common. Hence, we divide the search space into sub spaces and perform parallel computation on GPU.