Multi-target and Multi-camera Object Detection with Monte-Carlo Sampling

In this paper, we propose a general-purpose methodology for detecting multiple objects with known visual models from multiple views. The proposed method is based Monte-Carlo sampling and weighted mean-shift clustering, and can make use of any model-based likelihood (color, edges, etc.), with an arbitrary camera setup. In particular, we propose an algorithm for automatic computation of the feasible state-space volume, where the particle set is uniformly initialized. We demonstrate the effectiveness of the method through simulated and real-world application examples.