Real-time user selected dynamic template tracking for UAV

Template matching algorithms describe the way a tracker is able to follow an object by comparing two templates. Representing information about the object to be tracked and the other about the frame that is being analysed. Previous implementations prove to be sensible to partial occlusion, updating the model template even in cases where the line of sight between the camera and the object is being partially blocked, producing a misrepresentation of the object's features. This paper proposes a dynamic template update, using normalised cross correlation as a similarity metric. Using the response given by the similarity, it is possible to determine ranges in which to update the model of the object. With this, the algorithm is able to keep relevant information about the object when it is partially or completely occluded. The main assumption made during the development of most template tracking algorithms is the prior knowledge of the object's location and dimensions in the initial frame. This paper proposes an interactive implementation where the prior information that is needed can be obtained from a single point in the object using segmentation. The implementation of the algorithm described produced reliable and real-time results (30 frames per second) on the NVIDIA Jetson TX1 platform.