The developments introduced in this thesis belong to the area of computer vision and are aimed to provide with ideas to solve the problem of automatically segmenting objects in images acquired in environments with activity, i. e., those in which objects are moving and the illumination is not controlled. In order to develop the ideas and evaluate their performance, two different problems, from the point of view of requirements and environment conditions, are given a solution. Both of the problems are considered as challenges by the research community in the area of computer vision.

 The first problem to be considered consists on segmenting, and later identifying, the container's codes of trucks. For this challenge, images taken at the entrance of a commercial port were used.  In this case, segmentation techniques that allow the extraction of concrete objects fom indidivual images,  characters in this case, were implemented. Natural light is not the only challenge in this case, but also the conditions of the containers themselves.  In this context, we study different techiques from the literature as LAT, Watershed, Otsu's algorithm, local variation and thresholding algorithm, using them to segment gray-tone images. Based on this study, a solution is proposed combining several different techniques as an approach to successfully extract characters regardless of the environmental conditions.


Joining several segmentation techniques into a single method produces noisy segmentations. The knowledge of the features of the seeked objects helps designing filters which can discriminate the valid objecs from noise, avoiding this way to relay only on a classifier. The proposed system does not need any parameter tuning in order to adapt to light variations and achieves a high level of segmentation and identification of characters, although the system performance depends greatly on the classifier's capacity.  Afterwards, experiments with sequences of images are performed in order to improve system's response.  Each sequence contains several images of the same truck in consecutive moments of time. 

 
The second problem considered in this thesis,  is extracting all the objects which do not belong to an scene, using streams of sequences and by designing background models which can adapt to changes in the scene, specially, light changes.

Based on thecniques proposed in the literature for background subtraction and bearing in mind the memory and computational time constraints imposed by some intelligent systems, in this thesis several techniques are proposed to obtain adaptive background models with requirements specially suited to automatic surveillance systems. The proposed techniques, called  BAC (Background Adaptative with Confidence) and FBS (Fuzzy Background Subtracion), use a measure of similarity and a computation of probability based on experimental studies and are able to model the background adapting it to changes in the environment providing at the same time with a confidence measure of the built model. BAC and FSB subtract a frame from the background by assigning each pixel with a possibility of belonging either to foreground or background.

At this point, our developments are evaluted with several video sequences obtained indoors where  problems shadows or  sudden light changes may appear, and also the Wallflower benchmark, accepted in the literature as a good means to test background modelling technqiues. Results obtained by BAC and FSB are promising when compared to those obtained by other algorithms accepted by the community as representative.