Resumen de Contributions to robust people detection in video-surveillance
Computer vision is a field with multiple lines of research and different application domains, being video surveillance one of the most developed during the last years. During the past years, automatic video surveillance systems have experienced a great development driven by the growing need of security. These automatic systems include several image and video processing techniques for monitoring purposes. Among the different video surveillance tasks, the main objective of this thesis has been the exploration of the state of the art in people detection, analyze the most representative approaches, identify their weaknesses and propose contributions to improve current people detection state of the art.
The people detection task consists mostly of, firstly, the design and training of a person model based on characteristic parameters (motion, dimensions, silhouette, etc) and, secondly, the adjustment of this model to the candidate objects in the scene. Thus, the critical tasks in any people detection algorithm are the generation or extraction of the initial object hypotheses to be people from the scene and the person model used to classify those initial object hypotheses. Firstly, in order to analyze the people detection problems in surveillance scenarios the critical tasks in any people detection algorithm have been identified and a consequently framework for their evaluation have been designed. Secondly, three different people detection algorithms have been proposed and compared with the state of the art, covering all the people detection issues previously identified. Finally, two different people detection post-processing subtasks focused on improving the final detection results have been also proposed.
The performance of the proposed people detection algorithms and post-processing subtasks has been thoroughly evaluated on the proposed evaluation dataset. The experiments conducted demonstrated the advantages and disadvantages of every proposed people detection approach in typical surveillance scenarios. Finally, the inclusion of the proposed post-processing subtasks provides robustness and improves the final detection results.
