Evaluation of a Proposal for Resetting Trackers in Video Object Tracking

Object tracking is an essential task in various applications within the area of ​​computer vision, which include the development of video surveillance systems, robotic systems, autonomous navigation systems, airspace monitoring systems, interfaces for human interaction -computer, computer-assisted surgery systems, sports performance analysis systems, augmented reality systems, among others. Due to its potential use in this variety of applications, the scientific community has proposed different techniques for tracking objects, called trackers; however, there is no tracker robust enough to perform this task with high efficiency under real conditions. Current trackers rely heavily on an appearance model to facilitate their operation. However, in many cases the performance limitations of the trackers are due to the fact that they operate with appearance models contaminated by the incorporation of information that is not very related to the object of interest. To date, there are very few studies aimed at preventing or stopping the contamination of these models; There are also no studies carried out on the evaluation of the impact that these alternatives could have on the performance of the trackers. It is for this reason that in this work we propose to evaluate the impact generated by the reinitialization of the trackers (as a method of preventing contamination of the appearance models) on their operating performance.

Object tracking is a task within computer vision that deals with estimating the states of an object of interest over the frames of any video. In general, the state describes properties associated with the object, such as the position and the area of ​​extension that it occupies in a given frame (Forsyth and Ponce, 2011).
At present, a great variety of computational algorithms have been proposed to carry out this task, among which those that use online learning stand out, a characteristic that allows them to adapt to changes in the object over time, and to which we will refer in forward only as trackers.
To perform state estimation, a tracker typically creates an appearance model of the object at the beginning of tracking and updates it during its operation with the intention of representing the possible appearances that the object might assume over time (Yang et al., 2011); thus originating a strong dependency between the performance of the tracker and the quality of its appearance model that concentrates information on the object.
Under real conditions, the appearance model can be contaminated with information dissociated from the object due to factors intrinsic and extrinsic to the tracker, which occur during tracking and negatively impact its performance (Smeulders et al., 2014; Wu et al., 2015). ). However, most trackers are designed to operate despite contamination levels in their skin models.
There are very few studies that focus on reestablishing these models once certain levels of contamination are exceeded, which can be evaluated through the reliability of their estimates; process that will be the subject of our study and that we propose to call "tracker reset".