A switching fusion filter for dim point target tracking in infra-red video sequences

Motion and depth perceptions allow, to the human visual system, of interpreting the object movements by surrounding environmental information processing. The cognitive science applied to computer vision field can be considered an important innovation in order to increase the detection and tracking performances. These tasks play a fundamental role for detecting and tracking of dim moving point targets in Infra-Red (IR) images, which are characterized by low levels of SNR. In such cases, by means of the paradigm of Track-Before-Detect (TBD) based detection algorithm, it is possible to distinguish the target from image background. This paper presents an innovative TBD based approach relies on interacting multiple target models, which is called Fusion Filters (FFs), for far objects in IR sequences. Specifically, through two different Kalman filters it is possible to estimate separately position and dimension of the target. By means of switching probabilistic models, the proposed framework infers on the different target motion percepts. Such a process permits to obtain the global state of the object by merging position with size estimates. The experimental results on real and simulated sequences demonstrate the effectiveness of the proposed approach.