Adaptive background modeling in multicamera system for real-time object detection

We present an adaptive and efficient background modeling strategy for real-time object detection in multicamera systems. The proposed approach is an innovative multiparameter adaptation strategy of the mixture of Gaussian (MoG) background modeling algorithm. This approach is able to efficiently adjust the computational requirements of the tasks to the available processing power and to the activity of the scene. The innovative approach allows one to adapt the MoG without a significant loss in the detection accuracy while contemporarily adhering to the real-time constraints. The adaptation strategy works at the local level by modifying, independently, the MoG parameters of each task, and then, whenever the results of the local strategy are not satisfactory, a global adaptation strategy starts that aims at balancing the workload among the tasks. Our approach has been tested on three different data sets, including several image sizes, heterogeneous environments (indoor and outdoor scenarios), and different real-time constraints. The results show that the proposed adaptive system is well suited for multicamera applications thanks to this efficiency and adaptability; it guarantees real-time highly accurate detections.