Computer Vision Models in Surveillance Robotics
University of Trieste, Italy
http://hdl.handle.net/10077/4468
@article{moro2011computer,
title={Computer vision models in surveillance robotics},
author={Moro, A.},
year={2011},
publisher={Universit{\`a} degli studi di Trieste}
}
In this Thesis, we developed algorithms that use visual informations to automatically perform, in real time, detection, recognition and categorisation of moving objects independently on the environmental conditions and with the best accuracy. To this end, we developed upon several concepts of computer vision, namely the identification of the objects of interest in the whole visual scene (monocular or stereo), and their classification. In the course of development, several approaches have been tested, including the detection of possible candidate by image segmentation with weak classifiers and centroids, image segmentation algorithms enhanced by stereo information and reduction of noise, combination of popular features scale invariant (SIFT) combined with distance information. We developed two main categories of solutions associated with the type
of system used. With mobile cameras, we favourite the detection of known objects by scanning window; with fixed camera we use also foreground detection algorithms. In the case of foreground detection, detection rate and classification rate increases if the quality of the objects extracted is high. We proposed methods to reduce the effects of shadow, illuminations, and repetitive moving objects. An inportant aspect we studied is the possibility to use a foreground detection by moveable camera. Efficient solutions are getting complex, but also the devices to compute the algorithms are more powerful, and in the recent years, GPU architecture offer a big potential. We proposed a GPU implementation of an improved background management in order to increase the detection performance. In this Thesis we studied the detection and tracking of humans for applications such as the prevention of situation of risk (crossing street), and counting for analysis of traffic. We studied these problems and explored the various aspects of human detection, group detection, and detection in crowded scenarios. However, in a generic environment, it is impossible to predict the configuration of the objects that are captured by the camera. In these cases, we require to “abstract the concept” of an object. With this requirement in mind, we explored the property of stochastic methods and
show that good classification rates can be obtained provided that the training set is big enough. A flexible framework have to be able to detect moving regions and recognize the objects of interest. We developed a framework to manage the detection and classification problem. Compared to other methods, the proposed systems offer a flexible framework for objects detection and classification, and can be used efficiently in different indoor and outdoors environments.
of system used. With mobile cameras, we favourite the detection of known objects by scanning window; with fixed camera we use also foreground detection algorithms. In the case of foreground detection, detection rate and classification rate increases if the quality of the objects extracted is high. We proposed methods to reduce the effects of shadow, illuminations, and repetitive moving objects. An inportant aspect we studied is the possibility to use a foreground detection by moveable camera. Efficient solutions are getting complex, but also the devices to compute the algorithms are more powerful, and in the recent years, GPU architecture offer a big potential. We proposed a GPU implementation of an improved background management in order to increase the detection performance. In this Thesis we studied the detection and tracking of humans for applications such as the prevention of situation of risk (crossing street), and counting for analysis of traffic. We studied these problems and explored the various aspects of human detection, group detection, and detection in crowded scenarios. However, in a generic environment, it is impossible to predict the configuration of the objects that are captured by the camera. In these cases, we require to “abstract the concept” of an object. With this requirement in mind, we explored the property of stochastic methods and
show that good classification rates can be obtained provided that the training set is big enough. A flexible framework have to be able to detect moving regions and recognize the objects of interest. We developed a framework to manage the detection and classification problem. Compared to other methods, the proposed systems offer a flexible framework for objects detection and classification, and can be used efficiently in different indoor and outdoors environments.
October 9, 2011 by hgpu
