Evolving Parameters of Surveillance Video Systems for Non-overfitted Learning

Abstract

This paper presents an automated method based on Evolution Strategies (ES) for optimizing the parameters regulating video-based tracking systems. It does not make assumptions about the type of tracking system used. The paper proposes an evaluation metric to assess system performance. The illustration of the method is carried out using three very different video sequences in which the evaluation function assesses trajectories of airplanes, cars or baggage-trucks in an airport surveillance application. Firstly, the optimization is carried out by adjusting to individual trajectories. Secondly, the generalization problem (the search for appropriate solutions to general situations avoiding overfitting) is approached considering combinations of trajectories to take into account in the ES optimization. In both cases, the trained system is tested with the rest of trajectories. Our experiments show how, besides an automatic and reliable adjustment of parameters, the optimization strategy of combining trajectories improves the generalization capability of the training system.