SVM-Based Interactive Retrieval for Intelligent Visual Surveillance System

This paper proposes an interactive retrieval framework for intelligent visual surveillance system which introduces a SVM-based relevance feedback mechanism to perform semantic retrieval. In each round of retrieval, several objects are returned to user for labeling. The concept of user is learned by training a SVM classifier from the feedbacks. A trajectory feature extraction algorithm is also proposed to give an effective description of the trajectory. The trajectory features are extracted by mapping a Hausdorff distance based metric space to a vector space through a distance preserving transformation. Experimental results on real scenes demonstrate the effectiveness of the proposed algorithm.