Motion Classification Based On Motion Vector Statistics

Abstract

Motion analysis and classification is one of the most important and widely used applications of computer vision systems. These systems are used for content based video retrieval, anomaly detection and surveillance. A system is proposed that can work both on-line and offline for the classification of different types of motions for different applications. Kanade–Lucas–Tomasi (KLT) feature tracker was used for the construction of feature vectors. Statistical features of these motion vectors were then extracted (Average velocity, Standard Deviation of Velocity , Number of Motion Vectors, Mean Vector Length, Standard deviation of vector length, Mean Magnitude, Standard Deviation of Magnitude, Average Angle, Standard deviation of angle, Range of motion along Horizontal axes and Range of motion along Vertical axes) and used for classification purpose. The proposed system was tested on two applications i.e. traffic congestion classification where an accuracy of 96.3 % was achieved and human action recognition with an accuracy of 90.1%. The accuracy achieved is compatible with that achieved by different techniques.