A video-based tracking system for airport surveillance, composed by modules performing vision tasks at different levels, is adapted for operational conditions by means of Evolution Strategies (ES). An optimization procedure has been carried out considering different scenes composed of representative trajectories, supported by a global evaluation metric proposed to quantify the system performance. The generalization problem (the search of appropriate solutions for general situations, avoiding over-adaptation to particular conditions) is approached considering evaluation of ES-individuals over combinations of trajectories to build the fitness function. In this way, the optimization procedure covers sets of trajectories representing different types of problems. Besides, alternative operators for aggregating partial evaluations have been analysed. Results show how the optimization strategy provides a sensitive tuning of performance related to input parameters at different levels, and how the combination of different situations improves the generalization capability of the trained system. The global performance final system after optimization is also compared with representative algorithms in the state of the art of visual tracking.
|Number of pages||16|
|Journal||International Journal of Imaging Systems and Technology|
|Publication status||Published - 2007|
- Surveillance systems