Fast scene analysis for surveillance & video databases

In professional/consumer domains, video databases are broadly applied, facilitating quick searching by fast region analysis, to provide an indication of the video contents. For realtime and cost-efficient implementations, it is important to develop algorithms with high accuracy and low computational complexity. In this paper, we analyze the accuracy and computational complexity of newly developed approaches for semantic region labeling and salient region detection, which aim at extracting spatial contextual information from a video. Both algorithms are analyzed by their native DSP computations and memory usage to prove their practical feasibility. In the analyzed semantic region labeling approach, color and texture features are combined with their related vertical image position to label the key regions. In the salient region detection approach, a discrete cosine transform (DCT) is employed, since it provides a compact representation of the signal energy and the computation can be implemented at low cost. The techniques are applied to two complex surveillance use cases, moving ships in a harbor region and moving cars in traffic surveillance videos, to improve scene understanding in surveillance videos. Results show that our spatial contextual information methods quantitatively and qualitatively outperform other approaches with up to 22% gain in accuracy, while operating at several times lower complexity.