A Multiple Instance Learning Framework for Incident Retrieval in Transportation Surveillance Video Databases

Traffic incidents are frequent query targets in a transportation surveillance video database. Therefore, understanding and retrieving transportation videos based on their semantic contents becomes an urgent task. For this purpose, this paper proposes an interactive multiple instance learning (MIL) framework for semantic video retrieval. It incorporates techniques in multimedia processing, data mining, and information retrieval. By tracking vehicles\´ trajectories in a video and modeling semantic events, the framework initiates a progressive learning process guided by the user\´s relevance feedback (RF). The choice of RF is for reducing the "semantic gap" between the machine-readable features and the high level human concepts, which is a popular technique in the area of Content-based Image Retrieval (CBIR). With the information provided by RF. a mapping between semantic video retrieval and MIL is established. Due to its robustness to high-dimensional data. One-class SVM is selected to be the core learning algorithm for MIL in this framework. Although the proposed work is intended for transportation surveillance videos, it is designed as a general framework and can be tailored to other applications as well. The effectiveness of the algorithm is demonstrated by our experiments on real-life transportation surveillance videos.