Data-driven fusion of multi-camera video sequences: Application to abandoned object detection

Due to the potential for object occlusion in crowded areas, the use of multiple cameras for video surveillance has prevailed over the use of a single camera. This has motivated the development of a number of techniques to analyze such multi-camera video sequences. However, most of these techniques require a camera calibration step, which is cumbersome and must be done for every new configuration. Additionally, these techniques fail to exploit the complementary information across these multiple datasets. We propose a data-driven solution to the problem by making use of the inherent similarity of temporal signatures of objects across video sequences. We introduce an effective solution for the detection of abandoned objects using this inherent diversity based on the transposed independent vector analysis (tIVA) model. By taking advantage of the similarity across multiple cameras, the new technique does not require any calibration and thus can be readily applied to any camera configuration. We demonstrate the superior performance of our technique over the single camera-based method using the PETS 2006 dataset.