| Summary: | In the focus of this thesis are the new and modified algorithms for object detection, recognition and tracking within the context of video analytics. The manual video surveillance has been proven to have low effectiveness and, at the same time, high expense because of the need in manual labour of operators, which are additionally prone to erroneous decisions. Along with increase of the number of surveillance cameras, there is a strong need to push for automatisation of the video analytics. The benefits of this approach can be found both in military and civilian applications. For military applications, it can help in localisation and tracking of objects of interest. For civilian applications, the similar object localisation procedures can make the criminal investigations more effective, extracting the meaningful data from the massive video footage. Recently, the wide accessibility of consumer unmanned aerial vehicles has become a new threat as even the simplest and cheapest airborne vessels can carry some cargo that means they can be upgraded to a serious weapon. Additionally they can be used for spying that imposes a threat to a private life. The autonomous car driving systems are now impossible without applying machine vision methods. The industrial applications require automatic quality control, including non-destructive methods and particularly methods based on the video analysis. All these applications give a strong evidence in a practical need in machine vision algorithms for object detection, tracking and classification and gave a reason for writing this thesis. The contributions to knowledge of the thesis consist of two main parts: video tracking and object detection and recognition, unified by the common idea of its applicability to video analytics problems. The novel algorithms for object detection and tracking, described in this thesis, are unsupervised and have only a small number of parameters. The approach is based on rigid motion segmentation by Bayesian filtering. The Bayesian filter, which was proposed specially for this method and contributes to its novelty, is formulated as a generic approach, and then applied to the video analytics problems. The method is augmented with optional object coordinate estimation using plain two-dimensional terrain assumption which gives a basis for the algorithm usage inside larger sensor data fusion models. The proposed approach for object detection and classification is based on the evolving systems concept and the new Typicality-Eccentricity Data Analytics (TEDA) framework. The methods are capable of solving classical problems of data mining: clustering, classification, and regression. The methods are proposed in a domain-independent way and are capable of addressing shift and drift of the data streams. Examples are given for the clustering and classification of the imagery data. For all the developed algorithms, the experiments have shown sustainable results on the testing data. The practical applications of the proposed algorithms are carefully examined and tested.
|
|---|
