Human Motion Tracking And Recognition

Visual analysis of human motion in image sequence is an important research area in the field of computer vision. Human motion tracking and analysis includes the following steps: moving people detection in backgrounds; human motion tracking in video; motion feature extraction; and motion recognition. In this paper, we make research in human motion tracking, feature extraction and motion analysis. Our research work focus on the following aspects:(1)A tracking algorithm combining Mean Shift algorithm and multi-feature particle filter is proposed in this paper. After each sample is measured by dynamic model, mean shift analysis is applied to each sample based on observation density. After mean shift iterations, samples are "herded" to the local modes of the observation. So the degeneracy problem is efficiently overcome and it decreases the computational cost. At the same time, considering the limitation of single observation feature, the color and texture is adopted as the observation features to improve the algorithm's performance.(2)A method of feature extraction, which is based on motion history image, feature image, speed and the ratio of human silhouette's height and width, is proposed in the paper. Speed is computed by optical flow estimation based on gradient. The ratio of height and width can distinguish different motions very well. And motion history images can express the area of motion in space and how the motion takes place in temporal very well. Feature image created by recursive filtering indicates the speed and direction of the motion implicitly.Motion history images and feature images are extracted from video streams in action duration; then the principal component analysis is used to lower the dimensionality, and eigenspaces corresponding to several motions is then formed. A human motion image and a feature image are then projected to a point in the individual eigenspace. The complex motion feature is the vector composed by the feature points, speed and the ratio of human silhouette's height and width. (3)A method for motion recognition, which is based on improved fuzzy c-means clustering and learning vector quantization, is adopted in the paper. We present an adaptive algorithm to initialize the cluster centers. Choosing the first cluster center at random and the remainder can be found adaptively by the algorithm. After clustering the dataset, using the result as the labeled samples to training the LVQ network, and through the LVQ network, we can get the accurate cluster centers and the well classified recognition results.The experimental results show that the tracking algorithm based on particle filter and mean shift in this paper is robust and fast. And using the complex character based on MHI, speed and the ratio of height and width, the recognition method based on improved FCM and LVQ is active. It can recognize six actions: walk, run, skip, pjump, side and jump.