Moving Object Detection In Video Sequences Classification

The motion analysis based on video have a wide range of applications in intelligent transportation system, safety monitoring and military areas. The moving targets detection and classification are two very important steps in the motion analysis based on video, they are the foundation of the targets tracking and targets behavior understanding and description, so they are of high research value.This paper presents a thorough study of the moving targets detection and classification in video sequences. On the research of the targets detection, it studys the conditions in the static background and in the moving background respectively. Firstly, it introduces the adjacent frame difference, background subtraction and optical flow method, through the experimental analysis, their merits and demerits are presented respectively. Then it uses an algorithm developed from the adjacent frame difference and background subtraction. The experimental results show that the developed algorithm get better detection results. This paper choose a global motion estimation algorithm based on RANSAC+LS, and use the bilinear interpolation method to do the background compensation. The experiments show that the targets detection under moving background are finished effectively.On the research of the targets classification, it proposes to use the ratio of target height to the width in the half height and the ratio of target height to the width in the quater height besides the conventional features such as height-width ratio and duty ratio, and use the SVM to classify the moving targets. The targets are first classified to three categories:automobile, people riding on a bike and pedestrian, then it proposes to select the ratio of height in an eighth to height, the ratio of height in seven eighths to height and the ratio of width in three quarters height to width to further classify the automobiles, the automobiles are classified to car and bus. We detect the targets in the video sequence using the moving targets detection algorithm and extract the selected features of them, then transform the features to the LIBSVM format. We can get the SVM model after training the features. At last, we test the data in the test set, and compute the accuracy rate. The experiments show that the classification results improved obviously through the new features.