Feature Extraction and Object Classification in Video Sequences for Military Surveillance

Abstract

Object detection and recognition requires a learning system that can automatically identify a group of objects independently of the input data. To be able to perform this kind of identification, this system needs to previously analyze a large group of data, so it can memorize special features of different objects. This procedure it's called training and it's the first step in all the detection and recognition processes of machine learning. Although there are already many models that perform detection and recognition for a large group of objects, one of the goals of this project is to specify this identification into a small and special group of objects. This will be achieved by using transfer learning, that is a process that uses the knowledge gained by one of these models while solving one problem and applies it to a different one. Basically, it takes advantage of the feature extraction procedure outputs and use them to learn how to identify other kind of objects. Feature extraction is a group of processes with the goal of simplifying big groups of data by creating small sets of non-redundant information. These small groups are more manageable and can fully describe the original data set and, by using them, the resources necessary to analyse a large set of input data are decreased. In this context, the data to be analyzed will be captured by a camera implemented at a stationary point or in a vehicle. When dealing with the capture of visual information, it's normal that a large number of data is generated. So, it's important to analyze it efficiently and achieve relevant information identification. This dissertation focuses in military uses, therefore these operations are going to be used to automatically identify objects in the military field, that is, tanks, guns, people and vehicles (cars and trucks), achieving territorial surveillance.