Analysis of Visual Distinctive Features for the Global Tracking of Multiple Objects

Tracking has many possible applications as in surveillance systems, robot vision, automotive safety and so on. Therefore tracking of multiple targets in image sequences has become a classical task in computer vision. An example for an application is the use in driver assistance systems to avoid collisions. In this work a multi-target tracking framework which formulates multi-target tracking as minimization of a continuous energy function and solves minimization by conjugate gradient method with periodic trans-dimensional jumps is extended by an appearance-based energy term. Other than a number of recent approaches not only an appearance model consisting of a descriptor and a metric is selected. Instead many other possibilities such as weighting, prefilter and so on are considered. A couple of visual distinctive features with their parameters are optimized by random search and analyzed to find a good appearance model for the integration in the framework. The extension of the multi-target tracking framework is evaluated in experiments on two image sequences with variation in the influence of the new energy term. The addition of the appearance-based energy term leads to an increase in the performance of the tracking results