Regularity flow inspired target tracking in FLIR imagery

In this paper, a novel approach for target tracking in FLIR (Forward Looking Infra-red) imagery is presented. Generally IR (infra-red) signatures of targets are more prominent than background and clutter, and this contrast is commonly used as a clue for detection of targets and initialization of tracking algorithms. But in the case of small targets with poor SNR, detection based on this feature alone becomes challenging. The present approach also relies on this contrast based clue, but in place of detecting small targets, limited to a couple of pixels in X - Y plane the trajectory followed by the targets is explored in a sequence of X - T frames. A small group of frames of the video sequence is taken to form a 3D data cuboid with X, Y, T axes. This cuboid is re-represented as a stack of contiguous X - T slices over Y. As only a few of these X - T slices contain information related to the trajectory of horizontally moving objects, we can use Hough transform to detect lines of arbitrary inclination which would represent the trajectory in these selected X - T slices. Subsequently, the detected trajectory in X - T can be reprojected to the X - Y frames to label different moving objects. This method is commonly called 'track before detect'. A dataset of infra-red sequences having targets like tanks, AFVs and other targets of military importance in presence of substantial amount of clutter and variable atmospheric and thermodynamic conditions is used to validate our method. The results obtained demonstrate effectiveness and robustness of our approach. Multiple object tracking capability, occlusion handling capability and tracking without initialization are inherent advantages of our approach.