A real-time, semi-automatic method for discriminant target initialization in thermal imagery

Target initialization is critical in both appearance and feature based trackers since it determines what to track for the tracker. Therefore, any insignificant or spurious information, i.e. parts of objects similar to common background or patches from background, may result in redundant features or deceptive appearance that can cause premature track losses. In this sense, we propose a semi-automatic target initialization algorithm for IR images which takes an image coordinate as input, then generates target bounding box for the most salient segment in 5 ms on a 270MHz DSP. Experiments demonstrate that proposed algorithm is able to compensate input errors, i.e. exactly same target initialization can be achieved in the case of perturbations in input coordinate; and also boosts performance of several tracking algorithms in terms of number of frames in which successful tracking is achieved since discriminative targets are obtained.