Association of Moving Objects Across Visual Sensor Networks

We present a novel inter-camera trajectory association algorithm for partially overlapping visual sensor networks. The approach consists of three steps, namely Extraction, Representation and Association. Firstly, we extract trajectory segments in each camera view independently. These local trajectory segments are then projected on a common-plane. Next, we learn dynamic motion models of the projected trajectory segments using Modified Consistent Akaike’s Information Criterion (MCAIC). These models help in removing noisy observations from a segment and hence perform smoothing efficiently. Then, each smoothed trajectory is represented by its curvature. Finally, we use normalized cross correlation, as a proximity measure, to establish correspondence among trajectories that are observed in multiple views. We evaluated the performance of the proposed approach on a simulated and real scenarios with simultaneous moving objects observed by multiple cameras and compared it with state-of-the-art algorithms. Convincing results are observed in favor of the proposed approach.