Hyongseop Kim

Moment-based human motion recognition from the representation of DMHI templates

This paper presents a noble appearance-based recognition approach of human motion and gestures of several peoplespsila several actions from uncalibrated camera by employing motion history-based representation. It employs the basic motion history image-based and the directional motion history image-based representation and then exploits these motion templates to recognize various motions having more than one motion direction or complex motion. Traditionally, the basic MHI approach used seven Hu moments for feature vector calculation. This paper analyzed the implementation of a better feature vector calculation for our directional approach. We tried with two different feature vector sets for Hu moment. Moreover, due to its better performance, we computed another feature vector set based on the top twelve orders of Zernike moments. Due to computational cost, we finally ignored to employ Zernike moments for the DMHI template for recognition. This new feature vector calculation approach can reduce the calculation and shows good recognition rate. Finally, this paper raised some future concerns of this method.

SURF-based spatio-temporal history image method for action representation

Researches on action understanding and analysis are very crucial for various applications in computer vision. However, these face numerous challenges to represent and recognize different complex actions. This paper presents a noble spatio-temporal 3D (XYT) method for recognizing various complex activities, with a blend of local and global feature-based approach for motion representation. We incorporate SURF (Speeded-Up Robust Features), which is a scale- and rotation-invariant interest point detector and descriptor. Based on the interest points, optical flow-based directional motion history and energy images are developed. In this approach, the flow-based motion vectors are split into four different channels. From these channels, the corresponding four directional templates are computed. 56-D feature vector is calculated according to the Hu invariants for each action. k-nearest neighbor classification scheme is employed for recognition. We employ leave-one-out cross-validation method for partitioning scheme. We apply our method to outdoor dataset and we achieve satisfactory recognition results. We compare our method with some of other approaches and show that our method outperforms them.

Solving boundary problem of the motion database for improved human motion recognition

Development of a high accuracy and robust human motion recognition system is the fundamental requirement to apply in the real-life applications. Therefore, the recognition systems are being investigated to improve the systems performance and to enhance with more functionality. As a result of such an investigation, the formerly adopted structured motion database, capable of suitable retrieval of known motion patterns, is modified for obtaining more improved performance over the earlier systems [11]. The earlier systems has the boundary problems for the motion points those lie on the edge of the searching boundary within the query space at the time of retrieving similar motions from within the motion database. This problem might lead to the misrecognition of human motions in high degree. To overcome this problem, we make use of two sets of query spaces, namely original and shifted query space set which is to be switched based on the input motion. Motion History Image (MHI) and Exclusive-OR (OR) image are used as the motion representations, and the directional eigenspace strategy is used for constructing vector spaces. The results guarantee the significant improvement of the proposed system.